← HomeHOTAI.NEWS / AI WIKI

Understand AI,
choose the right models and tools.

Find models and tools, understand the concepts, and follow practical steps to get work done.

Search the wiki

Filter by topic

410 articles

Browse by content type

Article type	Your question	What the article covers
Core concepts	What is it, and how does it work?	Definitions, examples, and limitations
Models and Families	What can this model do, and which tasks suit it?	Model purpose, task selection and version limits
Tools and Products	How does this tool fit into your work?	Product purpose, usage and permission limits
Comparisons and choices	How do you choose the right approach?	Comparable conditions and selection criteria
Solve a problem	How do I do it, and what should I do when it fails?	Troubleshooting steps and outcome checks

Core concepts

What is it, and how does it work?

Start with a definition and concrete examples, understand the limitations, and explore related concepts.

Core conceptsWhat is accuracy? The fraction of all predictions that are correctExplain classification accuracy, distinguish precision and recall and assess imbalance and critical errors. Core conceptsWhat is an agent harness? Execution loops and state management around a modelHow harnesses connect models, tool execution and task state, with cross-session progress, recovery and approvals distinct from models and multi-agent orchestration. Core conceptsWhat is agent memory? Conversation state, persistent data, and model parametersUnderstand agent memory through a project assistant’s preferences, distinguishing short-term state, persistent storage, retrieval, provenance, updates, and deletion. Core conceptsWhat is agent orchestration? Delegation, state handoffs, and verificationUnderstand orchestration through organizing product documents, including fixed paths, dynamic delegation, dependencies, parallelism, retries, and termination. Core conceptsWhat are Agent Skills? Skill files, scripts, and on-demand loadingUnderstand reusable skill packages through a report-checking example, distinguishing prompts, tools, training, progressive loading, dependencies, and permissions. Core conceptsWhat is A2A? Agent discovery and cross-system task collaborationUnderstand Agent2Agent through delegated research, including discovery, task state, artifacts, differences from MCP, and limits of trust and authorization. Core conceptsWhat is AI code review? Turning comments into reproducible findingsUnderstand review suggestions, real defects, context, line references, and reproduction through a null-handling change, including why no comments do not prove correctness. Core conceptsWhat is AI Mode? Generative exploration and follow-ups in Google SearchExplain AI Mode’s complex queries and links, distinguish AI Overviews and indexing and verify answers and eligibility. Core conceptsWhat are AI Overviews? Google Search summaries with supporting linksExplain AI Overviews, distinguish generated summaries from source pages and check triggering, evidence and eligibility. Core conceptsWhat is an API key? A credential, not ordinary frontend configurationExplain API key authentication, distinguish endpoints and OAuth tokens and validate integration without exposing credentials. Core conceptsWhat is an API? How applications request functions and dataUse model text processing to explain API inputs and responses, distinguish SDKs and UIs and check real results. Core conceptsWhat is an approval workflow? Separate decisions from execution resultsExplain requests, approval states and subsequent execution, checking timeouts, rejection and duplicate notices. Core conceptsWhat is artificial general intelligence? Understanding AGI definitions and disputesDistinguish AGI from specialized capability, compare definitional scope, and examine why demonstrations and rankings cannot establish generality or autonomy. Core conceptsWhat is an aspect ratio? Proportions, pixel dimensions, cropping, and stretchingUnderstand aspect ratios versus resolution through landscape-to-portrait conversion, including crops, bars, expansion, safe areas, and display checks. Core conceptsWhat is automatic speech recognition? Transcription and common error boundariesUnderstand ASR through interview recordings, distinguishing transcription, translation, and summaries, with language, noise, names, numbers, and negation checks. Core conceptsWhat is batch inference? Grouped tasks still need individual reconciliationExplain batch inference versus concurrency and reconcile identifiers, completion states and partial failures. Core conceptsWhat is a benchmark? Scores apply to defined test conditionsExplain benchmark tasks, settings and scoring and distinguish public rankings from task acceptance. Core conceptsWhat is BM25? Term frequency and document length in relevance rankingUnderstand BM25 through model codes and error strings, including rare terms, frequency saturation, length normalization, and differences from factual accuracy. Core conceptsWhat is a browser agent? Completing web tasks through observations and actionsBrowser agents using page structure, screenshots and structured tools, distinguished from fixed scripts, broader computer use and WebMCP. Core conceptsWhat is a canonical URL? A preferred version for duplicate contentExplain canonicals versus redirects and indexing controls and align content, links and sitemaps. Core conceptsWhat is chain-of-thought? Uses and limits of step-by-step reasoning textUnderstand chain-of-thought prompting through a quantity calculation, distinguishing generated steps, internal computation, and evidence, including checks for misleading explanations. Core conceptsWhat is chunk overlap? Why adjacent passages repeat and what it costsUnderstand overlap through cross-boundary qualifications, including target versus actual overlap, continuity, duplicate retrieval, and context usage. Core conceptsWhat is document chunking? Preserving conditions and provenance in RAGUnderstand character, token, and structure-based chunking through a returns manual, including headings, tables, exceptions, and retrieval checks. Core conceptsWhat is citation provenance? Tracing AI claims to original evidenceUnderstand claim-to-passage links through fee conditions, distinguishing citation presence, evidential support, and source reliability. Core conceptsWhat is classifier-free guidance? Understanding strength and image changesUnderstand conditional and unconditional predictions, prompt alignment, quality tradeoffs, negative prompts, structural controls, and model-specific scales. Core conceptsWhat is code completion? Symbol suggestions, AI continuations, and applied editsUnderstand language-service and generative completion through a list-processing function, including context, acceptance, version matching, and runtime checks. Core conceptsWhat is code explanation? Checking AI descriptions against actual behaviorUnderstand code reading through missing and zero values, including inputs, outputs, callers, side effects, and differences between comments, intent, and implementation. Core conceptsWhat is code refactoring? Changing structure while preserving behaviorUnderstand renaming and extraction, distinguish refactoring from bug fixes, and check references, exceptions, public interfaces, and diffs. Core conceptsWhat is a coding agent? Reading, editing, testing, and delivering codeUnderstand coding agents through an export-field fix, including repository rules, runtime permissions, checks, and the distinction between a PR and deployment. Core conceptsWhat is a command-line interface? Commands, arguments, and exit statusUnderstand terminals, shells, programs, arguments, working directories, and status through a version query, including checks for AI-suggested commands. Core conceptsWhat is computer use? How models interact with software interfacesScreenshot observation, actions and feedback, distinguishing computer use from browser agents and direct APIs, with execution sessions, coordinates and authorization. Core conceptsWhat is a connector? Integrating external functions still needs account checksDistinguish connectors, configured connections and underlying services, checking actions, fields and recovery. Core conceptsWhat is context compaction? Shortening history while preserving task stateUnderstand compaction through a long coding conversation, distinguishing summaries, deletion, and caching, with checks for constraints, provenance, and provider-specific forms. Core conceptsWhat is context engineering? Selecting relevant, reliable information at every model stepContext engineering for support and coding tasks: selecting instructions, retrieved material, tool results and history, distinct from prompting, RAG and training. Core conceptsWhat is a context window? Per-request capacity and the limits of recallUnderstand context windows through multi-contract questions, distinguishing input length, output limits, and persistent memory, with budgeting and long-document checks. Core conceptsWhat is ControlNet? Guiding generation with edges, depth, and posesUnderstand control maps versus source images through pose control, distinguishing structure from identity, with model compatibility, preprocessing, and strength checks. Core conceptsWhat is data drift? Changed inputs do not imply self-changing model weightsExplain input-distribution changes versus quality degradation and use monitoring, labels and revalidation. Core conceptsWhat is deep learning? How multilayer neural networks learn complex patternsUnderstand multilayer models through image recognition, distinguish training from use, and examine data bias and overfitting before choosing a more complex model. Core conceptsWhat is Deep Research? Multi-step investigation and evidence synthesis across productsDeep Research planning, retrieval and reports, distinguishing research functions and specialized models from search summaries and explaining source and access limits. Core conceptsWhat is a diffusion model? From noise-based training to image generationUnderstand diffusion training and generation, text conditioning, pixel versus latent spaces, and why step counts and realism do not establish truth. Core conceptsWhat is direct preference optimization? Learning from preferred response pairsUnderstand DPO through support responses, chosen and rejected examples for one prompt, the reference model, and why preference labels do not establish factual accuracy. Core conceptsWhat is an embedding? Vector representations for similarity and retrievalUnderstand embedding vectors through differently worded support questions, including model matching, similarity, retrieval, and limits of truth and access. Core conceptsWhat is an evaluation dataset? Held-out examples for checking task performanceExplain evaluation inputs, references and records and check coverage, leakage and versioning. Core conceptsWhat is few-shot prompting? Demonstrating a task without training the modelUnderstand few-shot prompting through message classification, including differences from fine-tuning, example coverage, consistent labels, and evaluation separation. Core conceptsWhat is fine-tuning? Adapt an existing model with training examplesUnderstand fine-tuning through support classification, separate parameter updates from prompting and retrieval, and examine data quality, held-out evaluation, and platform availability. Core conceptsWhat is first-last-frame generation? Endpoint images and intermediate motionUnderstand endpoint constraints, references, interpolation, compatibility, transitions, and model support through open-close images. Core conceptsWhat is forced alignment? Matching supplied text to audio timingUnderstand alignment versus recognition through corrected subtitles, including language, dictionaries, missing words, boundaries, offsets, and caption pacing. Core conceptsWhat is a foundation model? Adapt pretrained capabilities to different tasksDistinguish foundation models from task-specific models through summarization and classification, and examine adaptation, downstream validation, and usage terms. Core conceptsWhat is FP16? Half-precision capacity and numerical limitsExplain sixteen-bit floating point, distinguish BF16 and mixed precision and verify labels against actual runtime. Core conceptsWhat is frame rate? FPS, playback speed, and new motion detailUnderstand FPS versus resolution and bitrate, repeated frames, interpolation, timestamps, audiovisual sync, and analysis sampling. Core conceptsWhat is full-text search? Indexing, term matching, and rankingUnderstand full-text search through product-model queries, distinguishing indexed terms, literal lookup, and semantic retrieval, with analyzers, fields, and missed-result checks. Core conceptsWhat is function calling? Model-generated arguments versus program executionUnderstand declarations, argument validation, call identifiers, and result returns through an order-status lookup, distinguishing requested actions from confirmed execution. Core conceptsWhat is generative AI? How generation differs from retrieving informationUnderstand generative AI through email drafts and illustrations, distinguish it from classification and search, and learn what to verify before using the result. Core conceptsWhat is GEO? Improving answer-source usability without citation guaranteesExplain GEO as research and industry practice, distinguish Google requirements from experiments and check evidence, indexing and observed effects. Core conceptsWhat is a GPU? Why graphics processors also run AI computationExplain GPU parallel processing, distinguish compute and memory from model speed and define task-validation conditions. Core conceptsWhat is grounding? Source-based answers still need verified citationsExplain grounding and retrieval and verify claim support, versions and missing evidence. Core conceptsWhat are AI hallucinations? Plausible answers can lack factual supportExplain incorrect or unsupported generation, distinguish retrieval gaps and unfaithful answers and verify sources and unknowns. Core conceptsWhat is human-in-the-loop? Integrating human review into AI executionUnderstand human involvement through an outgoing quotation draft, including observation, edits, approval, pause-resume behavior, and preventing repeated execution. Core conceptsWhat is hybrid search? Combining lexical and vector retrievalUnderstand merged candidates, rank fusion, filters, and reranking through a product-code question, distinguishing hybrid methods from simple concatenation. Core conceptsWhat is idempotency? Preventing duplicate effects during retriesUnderstand idempotency through a timed-out ticket creation, including intent, keys, matching parameters, lifetimes, and limits beyond a single operation. Core conceptsWhat is image-to-image? Combining input images, strength, and textUnderstand input conditioning and strength through sketch-to-illustration, including structural preservation, local edits, and comparison checks. Core conceptsWhat is image-to-video? Turning an image into motion-conditioned videoUnderstand first frames, appearance references, movement, cameras, preservation, and intermediate-frame checks through a product-display shot. Core conceptsWhat is an inference endpoint? A service entry point for model requestsExplain endpoints versus weights using classification requests, readiness and result checks across managed and self-run services. Core conceptsWhat is model inference? Turn new inputs into outputs after trainingDistinguish inference from training and reasoning tasks through classification and text generation, and learn how to compare latency, output length, and runtime settings. Core conceptsWhat is information extraction? Locate source fields rather than invent valuesExplain entity and date extraction, separating structure from evidence and checking unknowns, positions and relationships. Core conceptsWhat is inpainting? Masks, boundary blending, and preservation checksUnderstand mask semantics, context, and boundaries through replacing a background object, including checks for changes outside the selected area. Core conceptsWhat is INT4? Four-bit storage does not define all arithmeticExplain INT4 quantization, distinguish other four-bit formats and check scales, metadata and backend compatibility. Core conceptsWhat is an IDE? Integrating editing, debugging, and project toolsUnderstand IDEs and editors through navigation and debugging, including indexes, extensions, environment dependencies, AI features, and runtime evidence. Core conceptsWhat is intelligent document processing? Text recognition is only one stageExplain recognition-to-extraction pipelines and separate OCR, layout and business validation. Core conceptsWhat is IP-Adapter? Combining image prompts with text conditionsUnderstand reference-image features, encoders, and adapters through style guidance, distinguishing pose, identity, and pixel preservation. Core conceptsWhat is an AI knowledge base? Organization, updates, and answer evidenceUnderstand knowledge bases through public company manuals, distinguishing upload, indexing, and training, with updates, permissions, citations, and error diagnosis. Core conceptsWhat is knowledge distillation? How a smaller model learns from a teacherLearn about teacher and student models, soft targets, and generative distillation through support classification, including differences from quantization and checks for inherited errors. Core conceptsWhat is a knowledge graph? Connecting entities, relations, and evidenceUnderstand knowledge graphs through products, vendors, and versions, including graph databases, embeddings, entity resolution, time qualifiers, and evidence. Core conceptsWhat is the KV cache? Reusing attention history to reduce repeated workExplain key-value caching during generation, distinguish weights, answer caches and summaries and account for context memory. Core conceptsWhat is a large language model? From token prediction to summaries and answersLearn how large language models use tokens and context, distinguish models from chat applications, and examine knowledge limits and verification through document Q&A. Core conceptsWhat is latency? Time first output separately from completionExplain end-to-end latency, first output and queueing under matched inputs and quality criteria. Core conceptsWhat is least privilege? Limiting AI-tool access to the taskUnderstand identity, data scope, actions, read-only tools, temporary access, and revocation through an inventory assistant, without relying on prompts as access control. Core conceptsWhat is lip sync? Audio-driven mouth motion versus authentic speechUnderstand dubbing, conversion, mouth generation, face detection, profiles, durations, artifacts, and authorized use through a permitted character narration. Core conceptsWhat is a local model? Execution location differs from open-source statusExplain device-local inference versus local interfaces to remote models, including dependencies, offline conditions and output checks. Core conceptsWhat is LoRA? Trainable parameters, base models, and adaptation limitsUnderstand low-rank updates, adapters, full weights, quantization, rank, compatibility, and evaluation, without treating small files as standalone models or guaranteed gains. Core conceptsWhat is machine learning? Understand learning from examples through email classificationUnderstand training and prediction through email classification, the role of labels and independent tests, and why data shifts and overfitting still matter. Core conceptsWhat are maximum output tokens? Output limits and truncation checksDistinguish output caps, context windows, and actual response length through a long-summary example, including parameter settings, finish reasons, and completeness checks. Core conceptsWhat is an MCP client? Its role between a host and a serverUnderstand MCP clients through an editor connecting to two data services, including host roles, discovery, version compatibility, permissions, and end-to-end checks. Core conceptsWhat is an MCP resource? Readable context versus executable toolsUnderstand resource URIs, types, and templates through a project file, distinguishing listing, reading, context inclusion, and authority. Core conceptsWhat is an MCP server? Exposing tools and contextual capabilitiesUnderstand MCP servers through project data, including local and remote deployment, capability declarations, access to underlying data, and result handling. Core conceptsWhat is metadata filtering? Restricting versions, categories, and access in retrievalUnderstand metadata quality, pre- and post-filtering, relevance, and scope through manual versions and visibility, including denied-result and missed-result checks. Core conceptsWhat is a mixture-of-experts model? Expert routing and parameter countsUnderstand expert routing, total versus active parameters, and the limits of a support-ticket analogy, with checks for memory, load balancing, and deployment. Core conceptsWhat is model calibration? Probabilities should match observed frequenciesExplain probability calibration versus accuracy, reliability groups, held-out data and verbal confidence. Core conceptsWhat is a model checkpoint? Saving, resuming, and releasing training statesUnderstand model weights and optimizer states through interrupted training, including the files needed for inference versus resuming a run. Core conceptsWhat is MCP? Connecting tools and data through a context protocolUnderstand MCP hosts, clients, servers, tools, resources, and prompts through project-data lookup, distinguishing compatibility from authorization and capability. Core conceptsWhat are model parameters? Understand weights, biases, and parameter countsUnderstand weights and biases through a simple prediction formula, distinguish parameters from hyperparameters and context, and interpret parameter counts and numeric precision. Core conceptsWhat is a multi-agent system? Collaboration benefits and handoff problemsUnderstand roles, models, concurrency, independent work, shared state, and conflict resolution through a product-comparison task. Core conceptsWhat is a multi-turn conversation? History, current responses, and stateUnderstand conversation history, server-managed state, and persistent memory through an evolving travel plan, including corrections, references, and context limits. Core conceptsWhat is a multimodal model? Checking image, audio, and text capabilities separatelyUnderstand multimodality through image questions and speech inputs, distinguishing input and output support, combined understanding, contradictions, and file processing. Core conceptsWhat is a negative prompt? Steering away from content and its limitsUnderstand negative conditioning, ordinary negation, safety filters, and model support through an uncluttered poster background, including checks for unwanted removals. Core conceptsWhat is a neural network? Weights, hidden layers, and nonlinear computationUnderstand neural networks through handwritten-digit classification, including inputs, layers, outputs, learned weights, and checks for reliable predictions. Core conceptsWhat is open source AI? Why available weights alone are not enoughUse the OSI definition to examine materials needed for use and modification, distinguish code from weights, and learn how to inspect licenses, training code, and data information. Core conceptsWhat is an open-weight model? What to check after downloading weightsUnderstand what open weights provide, distinguish downloading and execution from open-source permission, and examine tokenizers, quantized versions, and model cards. Core conceptsWhat is OCR? Image text extraction still needs layout checksExplain OCR outputs and limits, distinguish table parsing and vision-language models and verify small text and reading order. Core conceptsWhat is outpainting? Canvas extension, composition, and preservationUnderstand canvas extension, stretching, super-resolution, masks, and boundaries through converting a portrait image into a banner. Core conceptsWhat is precision? How many selected results are truly positiveExplain precision using email classification and distinguish accuracy, recall, thresholds, false positives and zero denominators. Core conceptsWhat is pretraining? Learning broad capabilities before later adaptationUnderstand pretraining through continuation learning, including its relationship to chat behavior, continued pretraining, fine-tuning, and data freshness. Core conceptsWhat is prompt caching? Reusing computation for repeated contextUnderstand prompt caching through repeated manual questions, including prefix matching, lifetimes, usage checks, and differences from answer caching or permanent memory. Core conceptsWhat is prompt engineering? Iterating task design with representative examplesUnderstand organizing instructions, context, and examples through invoice extraction, including evaluation, version records, and distinctions from training and tools. Core conceptsWhat is prompt injection? Commands in source material are not user authorizationExplain instructions embedded in external content, separate user tasks from source data and validate tool boundaries with harmless samples. Core conceptsWhat is a prompt? Writing task instructions, references, and input boundariesUnderstand prompt components through rewriting an announcement, including instructions, source text, examples, output requirements, and missing facts. Core conceptsWhat is a pull request? Diffs, reviews, checks, and merge stateUnderstand commits, branches, PRs, deployment, base branches, checks, and review comments through an added export field, including AI-created proposals. Core conceptsWhat is model quantization? Fewer bits still require validationExplain weight quantization and distinguish file size, runtime usage and task quality under matched conditions. Core conceptsWhat is a random seed? Starting states, repeatability, and environment differencesUnderstand seeds through image comparisons, including random sequences, consumed state, cross-platform limits, and records needed for reproducibility. Core conceptsWhat are rate limits? Time-window quotas differ from account balancesExplain request and token limits, distinguish concurrency and spending caps and use service responses for retries. Core conceptsWhat is a reasoning model? Distinguish problem solving from model inferenceUnderstand reasoning models, thinking budgets, and final answers through a scheduling example, including why longer deliberation can still fail. Core conceptsWhat is recall? How many actual targets the system findsExplain recall, missed positives and reference scope in classification and retrieval. Core conceptsWhat is red teaming? Deliberately probing failure boundariesExplain adversarial model and application checks with authorized scope, failure evidence and normal-task regression. Core conceptsWhat is a reference image? Appearance, structure, and first-frame rolesUnderstand appearance prompts, structural maps, endpoints, multi-image conflicts, and preservation through character and composition requirements. Core conceptsWhat is RLHF? Turning human preferences into training feedbackUnderstand demonstrations, preference rankings, and reward models in a typical RLHF process, including why a thumbs-up does not establish immediate learning. Core conceptsWhat is reinforcement learning? Actions, feedback, and long-term rewardsUnderstand reinforcement learning through a simulated warehouse route, including reward signals, exploration, delayed returns, and the difference from a chat rating. Core conceptsWhat is repository indexing? How AI locates code and what it can missUnderstand repository indexing through cross-file calls, distinguishing indexing, reading, and training, with updates, exclusions, permissions, and revision checks. Core conceptsWhat is reranking? Reassessing relevance among retrieved candidatesUnderstand reranking through feature documentation, including candidate retrieval, cross-encoders, semantic rankers, missing evidence, latency, and citations. Core conceptsWhat is retrieval evaluation? Checking coverage and ranking separately from generationUnderstand retrieval evaluation through manual questions, including queries, corpora, relevance labels, Recall@k, ranking metrics, and missed-evidence diagnosis. Core conceptsWhat is RPA? Rule-based software actions still need business verificationExplain RPA versus model interpretation, checking UI targeting, input rules and duplicate submissions. Core conceptsWhat is a robots meta tag? Controlling indexing and previewsExplain noindex and preview controls, distinguish page tags, headers and robots.txt and verify crawler access. Core conceptsWhat is robots.txt? Crawl rules are not confidentiality controlsExplain robots exclusion versus indexing and inspect host scope, path matching and responses. Core conceptsWhat is sample rate? Samples per second, resampling, and file metadataDistinguish sample rate, bit depth, bitrate, and channels, with correct resampling for speech inputs and checks for duration, pitch, and formats. Core conceptsWhat is sandboxed execution? Constraining files, permissions, and networksUnderstand process isolation, containers, read-only mounts, shared kernels, and network boundaries through unknown-script execution, with allowed and denied checks. Core conceptsWhat is SEO? Help search engines discover and understand pagesExplain technical and content SEO, distinguish crawling, indexing and serving and validate links and titles. Core conceptsWhat is search indexing? Candidate inclusion differs from actual displayExplain discovery, crawling and indexing and inspect noindex, canonicals and content processing. Core conceptsWhat is self-hosting? Operating and maintaining your own model serviceExplain self-hosting control and responsibilities through weights, dependencies and recovery, distinguishing it from local-only execution. Core conceptsWhat is semantic search? Finding relevant material beyond exact wordingUnderstand semantic retrieval through return-related paraphrases, distinguishing embedding search from semantic reranking, with checks for codes, negation, and freshness. Core conceptsWhat is an SDK? Client tooling does not replace service validationExplain SDKs in model integration, distinguish APIs and models and verify types, retries and version changes. Core conceptsWhat is speaker diarization? Who spoke when versus real identityUnderstand diarization through a two-person interview, including anonymous labels, channels, overlap, transcription, name mapping, and cross-segment consistency. Core conceptsWhat is SSML? Pauses, pronunciation, and supported markupUnderstand speech-synthesis markup through a standard XML pause example, distinguishing specifications, vendor extensions, voice support, and listening checks. Core conceptsWhat is speech-to-speech? Dialogue, translation, and voice conversionUnderstand speech input-output paths through a voice assistant, distinguishing transcription pipelines, native audio, replies, conversion, interruption, latency, and transcript consistency. Core conceptsWhat are stopping conditions? Completion, pauses, and failure in agent loopsUnderstand stopping conditions through repeated searches, including result evidence, step caps, human interruption, and the difference between termination and success. Core conceptsWhat is streaming output? Incremental content still needs completion checksExplain incremental model events, interruptions, structured data and final states so partial answers are not accepted as complete. Core conceptsWhat is streaming transcription? Interim live text versus finalized segmentsUnderstand audio chunks, interim revisions, finalization, latency, disconnects, reconnection, and key-term checks through live captions. Core conceptsWhat is structured data? Standard page descriptions still need truthful contentExplain Schema.org and search markup and distinguish syntax, visible-content consistency and actual display. Core conceptsWhat is structured output? Schema conformity does not establish factual truthExplain field constraints and distinguish valid JSON, schema conformity and factual correctness, including unknowns and failures. Core conceptsWhat is super-resolution? Enlarging images without proving recovered detailUnderstand enlargement, denoising, and synthesized detail through a low-resolution product image, including text and texture checks against originals. Core conceptsWhat is supervised fine-tuning? Demonstrations, formats, and evaluationUnderstand SFT through ticket extraction, including input-output examples, chat templates, data leakage, and its difference from supplying a document in context. Core conceptsWhat is supervised learning? Labels for classification and predictionUse ticket categories and parcel times to understand classification and regression, label consistency, dataset splits, and evaluation leakage. Core conceptsWhat is a system prompt? Global behavior instructions versus ordinary messagesUnderstand system prompts through support scope restrictions, including role boundaries, provider naming differences, and why prompts do not enforce permissions. Core conceptsWhat is temperature? Separating sampling randomness from factual accuracyUnderstand how temperature affects token probabilities through headline variation, including support limits and why low temperature does not guarantee truth. Core conceptsWhat is text classification? Define labels before assigning themExplain single-label, multilabel and scored classification and validate references, unknowns and thresholds. Core conceptsWhat is text summarization? Condense sources without inventing their claimsDistinguish summaries from rewriting and fact checks, assessing omissions, attribution and negation. Core conceptsWhat is text-to-image? Text conditions, composition, and image acceptanceUnderstand text-to-image through an event-poster concept, including description versus precise layout, composition, references, text, product details, and repeated outputs. Core conceptsWhat is text-to-sound-effects? Generating ambience and action soundsUnderstand sound effects, narration, music, event ordering, duration, looping, and video synchronization through a door-and-footsteps example. Core conceptsWhat is text-to-speech? Pronunciation, prosody, and audio checksUnderstand text, voice, expression, numbers, ambiguous pronunciation, chunk joins, sample rates, and delivery checks through product narration. Core conceptsWhat is text-to-video? Shot descriptions, temporal generation, and acceptanceUnderstand actions, camera movement, editing, continuity, audio, text, model support, and file checks through a short-shot concept. Core conceptsWhat is throughput? Work completed per unit time differs from latencyExplain request and token throughput and compare capacity using concurrency, lengths and failures. Core conceptsWhat is a token? Its relationship to text length, usage, and context budgetsUnderstand token units, counts for Chinese and code, input-output-reasoning usage, and how to account for a complete request. Core conceptsWhat is tokenization? Mapping text into model inputsUnderstand segmentation and token IDs, subwords versus linguistic words, and checks for matching vocabularies and special markers when changing models. Core conceptsWhat is a tool result? Returned content, execution state, and evidenceUnderstand links, retrieved content, write confirmation, call IDs, errors, and structured data through file-search results, without treating returned text as trusted instructions. Core conceptsWhat is tool use? Querying and acting through external programsUnderstand tool requests, execution environments, evidence, argument permissions, errors, and untrusted returned content through weather and file examples. Core conceptsWhat is top-k retrieval? Result counts and evidence coverageUnderstand returned counts, internal candidates, reranking cutoffs, filters, and approximate search through multi-passage questions. Core conceptsWhat is top-p sampling? Selecting candidates by cumulative probabilityUnderstand nucleus sampling, top-k, and temperature through probability examples, including why thresholds are not accuracy scores and why support must be checked. Core conceptsWhat is a Transformer? Attention mechanisms and model architecturesUnderstand attention through pronoun relationships, distinguish encoders, decoders, and trained weights, and learn why architecture names do not establish capability. Core conceptsWhat is a trigger? It starts a workflow without proving completionExplain event and schedule triggers and distinguish inputs, execution states and business outcomes, including duplicates and recovery. Core conceptsWhat is unit testing? Checking behavior with fixed inputs and assertionsUnderstand unit tests with locally run Python assertions that preserve zeros, including pass status, coverage, delivery limits, and shared assumptions in AI-written code and tests. Core conceptsWhat is unsupervised learning? Finding structure without predefined labelsUnderstand unsupervised learning through clustering customer messages, including features, distances, business categories, and why cluster IDs are not ground-truth labels. Core conceptsWhat is a vector database? Similarity search, metadata, and source textUnderstand vectors, identifiers, metadata, exact and approximate neighbors, model migrations, and access updates through manual retrieval. Core conceptsWhat is video extension? New continuation versus stretching timeUnderstand continuation versus repeated frames and slow motion, including inputs, character drift, joins, and whether output includes the original clip. Core conceptsWhat is video memory? Weights are only part of runtime usageExplain GPU memory versus system memory and how loading, context and concurrency affect capacity checks. Core conceptsWhat is video understanding? Frames, audio, and temporal localizationUnderstand generation versus analysis, summaries, event localization, sampling, audio sources, and timestamps through an instructional demonstration. Core conceptsWhat is a vision-language model? Image questions, reading, and spatial limitsUnderstand VLM image questions, text reading, spatial relations, and multi-image comparison through a table screenshot, including scaling and unseen details. Core conceptsWhat is voice activity detection? Separating speech regions from silenceUnderstand VAD, ASR, and diarization through a long meeting, including thresholds, pauses, clipping, quiet speech, and global timing. Core conceptsWhat is voice cloning? Reference voices, synthesis, and authorizationUnderstand cloning versus TTS and conversion through authorized self-narration, including reference quality, language, similarity, and identity limits. Core conceptsWhat is a web crawler? Fetching a page does not mean indexing itExplain crawler discovery and fetching and distinguish indexing and training, checking logs, identity and access rules. Core conceptsWhat is WebMCP? Exposing website tools to browser agentsWebMCP tool registration and form interactions, distinct from MCP servers, browser automation and account permissions, with community-specification and preview limits. Core conceptsWhat is AI? Understand artificial intelligence through email classification and generated contentUnderstand artificial intelligence through spam detection and document summaries, distinguish machine learning from large language models, and learn how training differs from using a model and how to check AI outputs. Core conceptsWhat is an AI agent? Compare it with a chatbot through a document taskUse a meeting-document task to understand an LLM agent’s tool calls and feedback loop, distinguish agents from fixed workflows, and learn which results establish that a task is complete. Core conceptsWhat is RAG? How an AI knowledge base retrieves information before answeringUnderstand retrieval-augmented generation through an expense-policy example, including chunking and common retrieval methods, the difference between RAG and a vector database, and why a cited answer can still be wrong. Core conceptsWhat are word-level timestamps? Word locations, caption grouping, and precisionUnderstand word boundaries from recognition and alignment through long-recording captions, including word-character granularity, offsets, pauses, and playback. Core conceptsWhat is a workflow? Traceable tasks built from steps and conditionsExplain predefined control flow and model decisions through inputs, branches, recovery and final states. Core conceptsWhat is a world model? Environment prediction, interactive generation and physical realityWorld models learning environment states and transitions, illustrated by classic World Models research and Genie, distinct from ordinary video and reliable physics simulation. Core conceptsWhat is an XML sitemap? Discovery hints rather than an index guaranteeExplain sitemap URLs and modification dates and distinguish submission, crawling and indexing during maintenance. Core conceptsWhat is zero-shot prompting? Defining tasks without demonstrationsUnderstand zero-shot prompting through email classification, distinguishing no demonstrations from no training, with labels, uncertainty, evaluation, and when to add examples.

Models and Families

What can this model do, and which tasks suit it?

Check model families, inputs, outputs and access options, with version and license differences distinguished from observed results.

Models and FamiliesWhat Is Claude? The Model Family and Its Coding ToolsClaude is an Anthropic model family and is also the name of its chat product. This article distinguishes the models, the Claude app and Claude Code, including tool execution and model ID boundaries. Models and FamiliesWhat Is DeepSeek? Hosted Models and Open WeightsDeepSeek includes hosted services and several model series with published weights. This article explains API aliases, thinking mode and self-hosted versions, so that old tutorial names are not mistaken for the current underlying model. Models and FamiliesWhat Is Doubao? The Chat App and the Seed Model FamilyDoubao commonly refers to a chat app and also appears in Volcengine Ark model-service names. This article distinguishes the app, general Seed models, Seedream and Seedance, and explains how to choose a specific version. Models and FamiliesWhat is FLUX? Distinguishing the model family, editing tools and open weightsFLUX is a Black Forest Labs model family. Generation, reference editing and inpainting depend on the version. Learn how hosted APIs differ from open weights and why code and model licenses are separate. Models and FamiliesWhat is Gemini Robotics? Action models and embodied reasoningGemini Robotics is Google DeepMind’s robotics family, separating vision-language-action models from ER embodied reasoning. Learn about Robotics 2 inputs and outputs and why a preview API is not a complete robot controller. Models and FamiliesWhat Is Gemini? Multimodal Models and App FeaturesGemini is a Google model family and the name of its consumer app. Using Gemini 4 Argon as an example, this article explains input and output modalities, context limits and the distinction from Gemma. Models and FamiliesWhat Is Gemma? Deploying Google’s Open-Weight ModelsGemma is Google’s open-weight model family, distinct from the Gemini app. This article explains Gemma 4 specifications, effective parameters and memory, and why licensing must be checked by generation. Models and FamiliesWhat is Genie? Interactive world models versus generated videoGenie is Google DeepMind’s world-model research family. Learn how Genie 3 and Project Genie generate environments that respond to actions, where consistency fails and how a research prototype differs from a general API. Models and FamiliesWhat Is GLM? Z.ai Models and Changes to Thinking ModeGLM is Z.ai’s model family, including hosted APIs and open-weight releases. Using GLM-5.3 and earlier 4.5 and 4.7 releases, this article explains text input, thinking controls and migration differences. Models and FamiliesWhat is GPT Image? Image generation, regional edits and API choicesGPT Image is OpenAI’s image generation and editing model family. Learn how it differs from ChatGPT and what to check for masked edits, transparent output and version selection. Models and FamiliesWhat Is GPT? The Model Family, ChatGPT and the APIGPT is an OpenAI model family. This article distinguishes it from the ChatGPT app and uses the official catalog to explain model selection for text, image understanding and tool use. Models and FamiliesWhat is HunyuanVideo? Model branches and local generation requirementsHunyuanVideo is Tencent Hunyuan’s video generation family, with distinct text-to-video, image-to-video and avatar branches. Learn about public version 1.5, super-resolution, device requirements and community-license conditions. Models and FamiliesWhat Is Kimi? Models for Long Tasks and Legacy APIsKimi is Moonshot AI’s model and application brand. This article uses the current model catalog and K2.5 open weights to explain long-context uses, legacy API retirement and the limits of the modified MIT license. Models and FamiliesWhat Is Llama? Meta’s Open Weights and LicensingLlama is Meta’s model family, distinct from a chat app or llama.cpp. Using Scout and earlier releases, this article explains mixture-of-experts parameters, language support and specific community license limits. Models and FamiliesWhat is Lyria? Choosing songs, short clips and real-time musicLyria is Google DeepMind’s music generation family. Learn how text or images guide music and how Lyria 3.5, short clips and real-time generation differ in lyrics, audio output and editing limits. Models and FamiliesWhat is MiniMax H3? Multimodal video generation and its relationship to HailuoMiniMax H3 combines audiovisual references and text instructions to generate video with native stereo audio. Learn how it differs from older Hailuo models, how base weights work with services for 2K video and which community license applies. Models and FamiliesWhat is MiniMax Music? Full songs, section control and open-weight conditionsMiniMax Music creates music from descriptions and lyrics. Learn how Music 3.0 handles full-song structure, audio formats and local deployment, including community-license conditions and non-streaming limitations. Models and FamiliesWhat is MiniMax Speech? Text-to-speech, voice cloning and real-time repliesMiniMax Speech is a text-to-speech model family. Learn how Speech 2.8 and streaming interfaces handle voices, pronunciation and cloning, and how a synthesis model differs from a complete voice assistant. Models and FamiliesWhat Is MiniMax? Language, Speech and Video Model LinesMiniMax is a company brand spanning several model lines. This article focuses on general language and coding models, distinguishing the M3 series, earlier M2.5 weights, and the roles of Speech, Music and H3. Models and FamiliesWhat Is Mistral? Models, Hosted Catalogs and Open WeightsMistral is an AI company and model brand offering hosted interfaces and open weights. This article distinguishes general models from OCR services and explains third-party catalog entries and model-specific licenses. Models and FamiliesWhat is Nano Banana? Distinguishing Gemini image models from the chat appNano Banana names Gemini’s native image generation and editing models. Learn how text and reference images guide revisions, how versions differ, and how the API relates to the Gemini app. Models and FamiliesWhat is Qwen-Image? Chinese text rendering versus Qwen-VL understandingQwen-Image is Qwen’s image generation and editing family, with Chinese text rendering a stated focus. Learn about versions, reference editing and its distinction from Qwen-VL image analysis. Models and FamiliesWhat is Qwen-Omni? Understanding audiovisual perception and spoken outputQwen-Omni combines audiovisual inputs with text and can produce speech in supported variants. Learn how Qwen3-Omni’s Thinker and Talker differ, what each outputs and what real-time applications require. Models and FamiliesWhat is Qwen-VL? Image and video understanding versus image generationQwen-VL is Qwen’s vision-language family for reading images, understanding video and answering in text. Learn how Qwen3-VL handles documents, temporal evidence and GUI tasks, distinct from Qwen-Image generation. Models and FamiliesWhat Is Qwen? Choosing Hosted APIs or Local ModelsQwen is a model family from Alibaba Cloud’s Qwen team, with both hosted services and open-weight releases. Using Qwen3, this article explains model naming, thinking controls and local deployment choices. Models and FamiliesWhat is Seedance? Joint audiovisual generation, references and JimengSeedance is ByteDance Seed’s video generation family, with later versions supporting joint audiovisual output and multimodal references. Learn how Seedance 2.5 handles longer clips and timed edits and how models differ from apps. Models and FamiliesWhat is Seedream? ByteDance image generation, references and editingSeedream is ByteDance Seed’s image creation family for generation and editing. Learn how to choose tasks for infographics and product edits and distinguish model capabilities from app features and API availability. Models and FamiliesWhat is Veo? Native audiovisual generation and retiring preview versionsVeo is Google DeepMind’s video generation family. Learn how text and reference visuals guide clips with audio, what to check in the result and which Veo 3.1 Gemini API previews are retiring. Models and FamiliesWhat is Wan? Choosing tasks across the video model family and cloud servicesWan is Alibaba’s visual generation model family with multiple versions and task branches. Learn how text-to-video and image-to-video relate to open weights, cloud APIs and audio-capable outputs.

Tools and Products

How does this tool fit into your work?

Understand how the tool runs and uses models, with practical examples, permissions and access conditions.

Tools and ProductsWhat is Claude Code? How Claude models participate in project developmentHow Anthropic Claude Code reads projects, edits files and executes commands, separating Claude models from the agent product and explaining interfaces and permissions. Tools and ProductsWhat is Codex? Distinguishing the coding agent, clients and underlying modelsHow OpenAI Codex reads projects, edits code and uses tools, with distinctions between GPT models, completion, local execution, permissions and verification. Tools and ProductsWhat is Cursor? How its editor, agents and model choices work togetherHow Cursor organizes code editing and agent tasks, separating tools, rules and models and distinguishing completion, cross-file edits and review. Tools and ProductsWhat is Google Antigravity? Distinguishing the development platform from GeminiGoogle Antigravity’s desktop, terminal, editor and programmatic forms, with the roles of its shared agent harness, Gemini models, artifacts and permissions. Tools and ProductsWhat is OpenClaw? How self-hosted assistants, Gateways and chat channels work togetherOpenClaw’s self-hosted assistant and persistent Gateway, with the roles of sessions, models, channels and trust boundaries rather than treating it as a model or IDE. Tools and ProductsWhat is OpenCode? How an open-source coding agent connects different model providersOpenCode project initialization, provider connections, model selection and permissions, distinguishing the client from local inference and a complete development environment. Tools and ProductsWhat is Qoder? Distinguishing its standalone desktop app, IDE and CLICurrent Qoder product forms, Coding and General modes, Knowledge Center and terminal use, separating task software from models and avoiding an IDE-only description. Tools and ProductsWhat is Qwen Code? Distinguishing the coding client from Qwen modelsHow Qwen Code uses project tools, authentication and multi-protocol providers, separating model weights, the open-source client and account-service permissions. Tools and ProductsWhat is WorkBuddy? How Tencent’s workspace combines files and reusable skillsTencent WorkBuddy tasks, experts, skills, connectors and knowledge bases, including local-directory authorization and deliverables rather than treating the workspace as a model.

Comparisons and choices

How do you choose the right approach?

Compare approaches for the same task, with selection criteria and the trade-offs involved.

Comparisons and choicesAd hoc document chat versus shared knowledge bases: separate one-off files from maintained versionsCompare one-off file input with maintained retrieval, checking updates, access scope and citation traceability. Comparisons and choicesSkills versus MCP: combine task methods with external connectionsExplain how Agent Skills package procedures and MCP connects tools and material, then check their combination against a single research task. Comparisons and choicesAI agents vs. workflows: what is the difference, and which should you use?Compare AI agents and predefined workflows by who decides the next step, suitable tasks, operating costs and maintenance. Learn how to combine them and evaluate both on the same tasks. Comparisons and choicesAI app or model API: choose around the destinationCompare ready-made AI applications and model APIs on the same document task, including interface use, integration and maintenance. Comparisons and choicesAI product scenes versus photo retouching: product details must matchCompare generated scenes with adjustments to real photos, checking color and structure against the same product rather than equating polish with fidelity. Comparisons and choicesAI upscaling versus interpolation: sharper details may be inventedCompare generative super-resolution with conventional resampling on one low-resolution original, checking text and texture rather than treating more pixels as recovered facts. Comparisons and choicesAI and human code review: comment counts do not establish qualitySeparate automated feedback from maintainer acceptance using the same diff, and evaluate real defects, false positives and suggested fixes alongside review effort. Comparisons and choicesAPI tools versus computer use: compare the actions that can actually be completedCompare structured interface calls with screen-based computer interaction, and assess field stability, layout changes and result verification for the same application. Comparisons and choicesCER versus WER: align segmentation and normalization firstCompare character and word error rates, including Chinese segmentation, number formatting and insertions in transcription evaluation. Comparisons and choicesASR versus audio understanding: distinguish words from acoustic cuesCompare transcription with direct audio analysis and check spoken words, nonverbal sounds and timing separately. Comparisons and choicesCode completion versus coding agents: suggestions and task execution differCompare editor completions with agents that modify and validate code, using one repair task to assess context, scope and acceptance effort. Comparisons and choicesBackground removal versus replacement: validate transparency and the new scene separatelyDistinguish transparent subject extraction from generating a scene, and check alpha, edges and subject fidelity on the same product image. Comparisons and choicesBreaking news versus evergreen guides: separate timely records from lasting helpCompare event-led publishing with maintained problem guides, checking dates, versions and page consolidation. Comparisons and choicesCloud flows versus desktop flows: check connectors and interfaces separatelyCompare service-connected execution with workstation automation, checking app coverage, environments and final states on the same task. Comparisons and choicesDense models versus MoE: distinguish total parameters from active parametersUsing the original Mixtral and Switch Transformer papers, examine how expert routing changes computation and how to compare memory needs and task results when self-hosting. Comparisons and choicesPatches versus whole-file rewrites: preserve unrelated contentCompare local diffs with full replacements and assess base versions, patch checks and content loss in AI-assisted code changes. Comparisons and choicesEvent triggers versus schedules: separate timely processing from reconciliationCompare change-triggered execution with periodic polling, checking duplicates, omissions and time windows. Comparisons and choicesExact match versus rubric grading: distinguish fixed answers from open expressionCompare exact text matches with explicit content rubrics, checking normalization, omissions and grader disagreement. Comparisons and choicesFAQ versus topic guides: organize independent answers and connected explanations by intentCompare question lists with complete guides, checking duplicate answers, canonical URLs and structured data without promising search features. Comparisons and choicesFirst-frame versus endpoint control: inspect motion between matching endpointsCompare starting-image guidance with start-and-end constraints, checking transition goals, incompatible images and model support. Comparisons and choicesFixed versus semantic chunking: keep answer prerequisites togetherCompare length-based and content-boundary chunking using cross-paragraph questions, and account for reindexing and duplicate passages. Comparisons and choicesFrontend direct calls versus backend proxies: keep persistent keys out of browsersCompare direct browser requests with backend forwarding, distinguishing persistent keys, temporary credentials and result validation. Comparisons and choicesFull conversation history versus summaries: preserve details that need traceabilityCompare resending full conversations with compacting history, identify information that summaries may lose and use earlier task details to check whether context remains usable. Comparisons and choicesFunction calling versus MCP: separate model tool selection from system connectionExplain function calling within model requests and MCP in external connections, using the same task to check tool discovery, arguments and execution results. Comparisons and choicesGenerative video versus template animation: separate scene creation from layout controlCompare model-generated scenes with timeline rendering of known assets, assessing text, brand layout and batch updates. Comparisons and choicesGraphRAG versus basic RAG: separate relationship questions from passage lookupUse the GraphRAG paper and project documentation to compare graph structures and community summaries with passage-based retrieval for global and local questions. Comparisons and choicesHybrid versus vector search: check whether exact identifiers are missedExplain combining lexical and semantic candidates, and compare recall, ranking and maintenance using identifiers and natural descriptions from the same collection. Comparisons and choicesIDE versus CLI coding agents: the interface does not determine execution locationCompare navigation, terminal commands and diff review, check execution environments separately and assess effort on the same repository task. Comparisons and choicesInteractive versus background agents: separate interim decisions from completion noticesCompare conversational execution with background tasks, design clarification, status queries and result retrieval, and verify completion after client timeouts. Comparisons and choicesKeyword versus semantic search: test exact wording and natural descriptions separatelyCompare lexical matching with semantic similarity, explaining how paraphrases, proper names and index maintenance influence document retrieval. Comparisons and choicesLLM drafting versus template merge: separate flexible prose from fixed fieldsCompare new prose from material with filling fields in fixed text, checking missing values, wording variation and batch revisions. Comparisons and choicesLLM judges versus human evaluation: calibrate the scoring standard firstCompare model grading with human review, checking judge bias, references and disagreement resolution. Comparisons and choicesLocal versus cloud ASR: compare complete workflows on the same audioCompare self-run speech models with hosted transcription, including hardware, uploads and correction effort. Comparisons and choicesLocal versus cloud coding agents: reproduce the dependency environment firstCompare execution in an existing workspace with a remote isolated environment, checking dependencies, uncommitted files and result retrieval for one repair. Comparisons and choicesLong context versus RAG: validate full input and retrieved selectionsCompare supplying complete material with retrieving excerpts, and assess omissions, citations and updates in knowledge-based tasks. Comparisons and choicesLow-code versus AI-generated repositories: plan for ongoing maintenanceCompare platform-built applications with generated projects using one business prototype, assessing connections, modification and migration. Comparisons and choicesManaged APIs versus self-hosting: account for control and maintenanceCompare hosted inference with operating model services, checking versions, concurrency and recovery across the full delivery cost. Comparisons and choicesManager tools versus handoffs: establish who owns the final replyCompare calling a sub-agent as a tool with handing over a conversation, and define responsibility for context transfer, final replies and failure recovery. Comparisons and choicesNative generated audio versus post dubbing: balance synchronization and revisionCompare audio generated with video against later scripted speech, checking exact lines, sound layers and local rework. Comparisons and choicesNews summaries versus explainers: separate events from lasting questionsCompare timely reporting with lasting answers, checking sources, explanations and update ownership instead of expanding pages by paraphrase. Comparisons and choicesOCR pipelines versus vision models: separate text location from visual reasoningCompare extracted text pipelines with direct visual understanding, checking numbers, coordinates and cross-region relationships in one document image. Comparisons and choicesOffline evaluation and online A/B: separate fixed samples from real behaviorCompare fixed-dataset quality checks with randomized product experiments, assessing coverage, assignment and stopping rules. Comparisons and choicesOn-demand versus batch APIs: collect and reconcile every delayed resultCompare interactive requests with asynchronous batches, checking deadlines, identifiers and partial failures. Comparisons and choicesOpen weights or proprietary service: check your rights firstDistinguish open weights, open-source AI and proprietary services, and compare rights, deployment control, operations and task quality. Comparisons and choicesOriginal tests and aggregation: separate execution evidence from source reviewCompare claims supported by first-hand testing and source synthesis, checking versions, inputs and failures for credible reviews. Comparisons and choicesPrompt caching versus answer caching: reuse computation or reuse a resultDistinguish provider-side prompt caching from saved application answers, and explain how prefix matching, changing material and access conditions affect reuse and validation. Comparisons and choicesPrompt templates versus Skills: choose packaging around task materialsCompare a prompt template with an Agent Skill containing instructions and resources, explaining why repeated writing and script-dependent tasks benefit from different storage methods. Comparisons and choicesPublic benchmarks versus task evaluations: high scores do not transfer automaticallyCompare standardized results with real task samples, checking model versions, prompts and scoring before applying rankings. Comparisons and choicesQuantized model vs higher precision model: first verify task degradationExplain how quantization changes the numerical representation of model weights by checking memory usage, runtime conditions, and output quality with the same base model and task samples, not treating bit depth as a general speed guarantee. Comparisons and choicesRAG vs. fine-tuning: what is the difference, and which suits knowledge-base answers?Distinguish RAG’s role in supplying context from fine-tuning’s role in adapting model behavior. Compare both on matched questions, data freshness, maintenance and error diagnosis before combining them. Comparisons and choicesRaster versus vector logos: an SVG extension does not ensure editabilityCompare pixel and path graphics, checking embedded raster images, fonts and small-size exports when delivering brand assets. Comparisons and choicesReasoning or general model: decide from task failuresCompare reasoning and general models on conditional analysis and rewriting, including thinking budgets, latency and evaluation examples. Comparisons and choicesReference images versus character tuning: recurring subjects still need consistency checksCompare inference-time references with trained adaptation using one character in different poses, and include training and version maintenance. Comparisons and choicesRule parsers versus LLM extraction: distinguish stable syntax from semantic variationCompare regex or field rules with LLM extraction, checking format changes, unknown values and source locations. Comparisons and choicesSDK versus direct HTTP: wrappers do not replace response understandingCompare official client libraries with handcrafted requests, checking types, retries and errors rather than treating installation as connectivity. Comparisons and choicesSEO and GEO: Google AI visibility still depends on search fundamentalsDistinguish general search optimization from content usability in generated answers, using Google guidance on eligibility, evidence and measurement limits. Comparisons and choicesServerless versus persistent inference: measure cold starts and idle resourcesCompare demand-based scaling with retained instances, checking loading, concurrency queues and recovery for inference services. Comparisons and choicesSingle versus multiple agents: locate the failing step firstCompare a single assistant executing sequentially with several sharing work, and validate tool problems, task decomposition and handoff costs separately. Comparisons and choicesLong takes versus edited clips: balance action continuity and local repairCompare continuous generated action with edited shots, checking continuity, cut points and rework scope. Comparisons and choicesSingle models versus model routing: correct dispatch still needs quality checksCompare one-model serving with rule-based dispatch, checking selection, fallback and evaluation samples for maintainable services. Comparisons and choicesSmall or large model: establish a classification baselineCompare small and large models on matched classification tasks, considering labels, error groups and actual usage rather than parameter counts alone. Comparisons and choicesSource separation versus denoising: define the sound to preserveDistinguish separating mixed sources from suppressing noise, checking target loss and residuals on the same recording. Comparisons and choicesDiarization versus speaker identification: Speaker A is not a nameDistinguish time-based speaker grouping from matching known identities, and check meeting labels, reference audio and overlapping speech. Comparisons and choicesSpeech-to-speech versus cascaded voice: interaction and inspectability differCompare direct voice sessions with transcription-reasoning-synthesis pipelines, testing interruptions, intermediate text and complete responses. Comparisons and choicesSpreadsheet formulas versus LLMs: separate deterministic calculation from text judgmentCompare rule-based calculation with open-ended text processing, checking formula preservation, changed inputs and recalculation. Comparisons and choicesStandard TTS versus custom voices: distinct voices require ongoing workCompare preset voices with custom voices, assessing recording material, access conditions and pronunciation checks for narration. Comparisons and choicesStreaming versus non-streaming: visible text does not mean completionCompare incremental model events with complete responses, checking first output, structure parsing and disconnections separately. Comparisons and choicesStructured outputs versus JSON mode: valid syntax still needs field checksDistinguish valid JSON, schema-conforming output and factual field accuracy, and explain how programs should choose constraints and handle exceptional model responses. Comparisons and choicesSubtitles versus AI dubbing: validate reading and listening separatelyCompare original audio with subtitles and synthesized target speech, checking translation, timing and spoken content separately. Comparisons and choicesSystem instructions versus user prompts: separate stable rules from current materialExplain the roles of system-level requirements and individual task input, organize multi-turn history, retrieved material and template variables, and test rules with conflicting examples. Comparisons and choicesText-to-image versus editing: evaluate new composition and source fidelity separatelyCompare generation from text with editing existing images, and explain why references, strength settings and masks do not replace fidelity checks. Comparisons and choicesText-to-video versus image-to-video: separate composition freedom from fidelityCompare text-created motion with image-guided video, checking subject fidelity, camera instructions and the complete clip. Comparisons and choicesText models versus multimodal models: screenshot information cannot rely solely on transcriptionWhen processing files with screenshots, compare the differences between extracting text first and directly inputting images, explaining how to separately verify table relationships, visual cues, and character-by-character accuracy. Comparisons and choicesAI-generated unit versus end-to-end tests: identify the failure boundaryDistinguish function rules from full user workflows and assess AI-generated tests using independent assertions, isolated data and deliberate defects. Comparisons and choicesVideo extension versus regeneration: preserve usable footage or rebuild the actionCompare continuing existing footage with rebuilding a shot, checking boundary continuity, accumulated changes and model support. Comparisons and choicesVideo understanding versus transcript analysis: locate the evidenceCompare visual video analysis with speech transcripts, assessing demonstrations, on-screen text and evidence timestamps. Comparisons and choicesWeb search versus AI search: verify original discovery and synthesized answersCompare result links with generated answers, checking source coverage, complex questions and time-sensitive facts. Comparisons and choicesZero-shot versus few-shot prompts: examples should cover real ambiguitiesCompare instructions alone with input-output examples, showing how example selection, input distribution and added context affect results and maintenance.

Solve a problem

How do I do it, and what should I do when it fails?

Start with the symptom, work through actionable troubleshooting steps, and check whether the issue is resolved.

Executed tool testImage enlargement at 2× and 4×: downloaded dimensions, text edges and loading timeA synthetic 160 × 120 image tested in HotAI Image enlargement, with both PNG downloads, WASM phase timings and a comparison image. Executed tool testImage OCR test: mixed Chinese and English, decimal points and I/l errorsThree public synthetic samples show HotAI OCR output, degradation errors and separate initialization and recognition timings. Solve a problemAn agent hits its step or time budget: how do you distinguish partial work from success?Reaching a step or time limit means an agent was stopped, not that its task succeeded. Record the stop reason, confirmed results and remaining work separately, using the target artifact to judge success. Solve a problemAre fewer agent calls always better: how do you verify outcomes and necessary actions together?Fewer calls are not necessarily better. Verify actual completion, required checks and absence of duplicate side effects before comparing resource use and time. Solve a problemAn agent keeps calling tools: how to diagnose and bound the loopDiagnose agent loops through tool calls and results, apply step limits and duplicate signatures, and distinguish stopping from completion with a locally tested simulator. Solve a problemAn AI answer includes source links: how to verify its claimsBreak an AI answer into checkable claims, verify source scope, versions and numbers, and distinguish related links from evidence that supports a conclusion. Solve a problemHow do you record AI citations without treating a few manual searches as total exposure?For AI-citation observation, define product, questions and qualifying links and preserve times and destinations. A few manual searches describe only those tested conditions, not total exposure or stable ranking. Solve a problemAI keeps changing code without fixing the bug: how do you establish a minimal failing path?When repeated AI changes do not resolve a bug, fix a reproducible minimal input, startup command and observable failure before editing a specific path. Without reproduction, changing implementations repeatedly can add unrelated behavior. Solve a problemAI changed unrelated files: how do you narrow the patch without losing existing user edits?When AI edits unrelated files, compare the before-and-after working tree and individual changes to distinguish its edits from existing user work. Do not reset the repository or discard all uncommitted files. Solve a problemAI impressions rise but visits do not: how do you distinguish exposure, clicks and business outcomes?Distinguish displayed links, clicks and on-site business results when AI impressions rise without visits. Google’s dedicated report shows Search AI impressions; dividing ordinary Search clicks by that number does not produce AI-specific CTR. Solve a problemHow do you test an AI-generated data migration on a copy and verify recovery?Run an AI-generated migration on an independent copy and verify an actual backup restore before production. Valid syntax does not establish safe data semantics, and a rollback function does not establish recovery of lost information. Solve a problemAI refactoring subtly changes output: how do boundary samples reveal regressions?When refactoring subtly changes results, define required compatibility and compare old and new behavior on identical inputs. Shorter code and existing passing tests do not establish equivalence. Solve a problemHow do you verify backup and rollback before deploying an AI service?AI service recovery requires matching code and model versions plus configuration and data compatibility, not just an old image tag. Verify remote backups and a known-good release before deployment and execute an explicit fallback if production checks fail. Solve a problemAI-generated mobile pages overlap text: how do you test breakpoints and real content?For overlapping mobile text, reproduce with real content length and the intended viewport, then inspect fixed heights, absolute positioning and wrapping restrictions. A normal desktop screenshot does not establish responsive correctness. Solve a problemA task lacks a time range or subject: how do you make AI ask when needed instead of guessing?If missing conditions change the answer, AI should identify the gap and ask a specific question; details affecting only style can be handled with a stated assumption. Decide whether the missing information changes the outcome, then ask the minimum necessary questions. Solve a problemAn API returns 401: how do you verify authentication without exposing credentials?For API 401 errors, check destination, authentication format and credentials actually read by the process instead of retrying as though it were a network timeout. A key in a file does not establish correct runtime loading or suitability for the service. Solve a problemAn API returns 429: how do you design bounded backoff and queueing?For 429 responses, identify request-rate, token-rate or other quota constraints and read supplied retry information. Immediate looping increases pressure, and layered SDK and application retries can multiply calls. Solve a problemA model API intermittently returns 5xx: how do you bound retries and resource use?For occasional 5xx responses, inspect typed errors and apply bounded retries according to service guidance rather than looping on every failure. SDK retries count toward the total budget. Solve a problemAn API rejects an overlong input: how do you calculate the actual context and remove irrelevant content?For a context-related 400, read the specific error and establish whether input, output or total window is exceeded. Character counts are not token counts, and tools, history and attachments can contribute. Solve a problemThe user sees a failed request but the backend succeeded: how do request IDs connect both outcomes?When the user sees failure but records exist, associate a business operation identifier with every network attempt’s request ID and inspect client, gateway and service outcomes separately. UI timeout does not prove nonexecution, and backend success does not prove delivery to the user. Solve a problemAn API times out: how do you distinguish connection, response and streaming failures?For timeouts, first locate failed connection establishment or sending, then distinguish response waiting from mid-stream interruption instead of raising one global timeout. Stages have different causes, and SDK retries and proxies alter elapsed time. Solve a problemArticle citations fail or change content: how do you update evidence without changing the original meaning?For broken or changed citations, identify the supported claim before finding primary evidence for the same fact. Do not substitute an accessible homepage or silently alter meaning through a new source. Solve a problemChanging an image to 3:4 or 16:9 crops the subject: how do you plan a safe composition?When conversion to 3:4 or 16:9 crops important content, mark required regions before choosing cropping, padding or canvas extension. A changed display ratio does not automatically recompose the original, and filling a container commonly crops edges. Solve a problemTranscription gets names and amounts wrong: how do you review them against the original audio?For incorrect names and amounts, replay the corresponding audio instead of merely asking a language model to make sentences fluent. Preserve original transcripts and timestamps, distinguish unclear speech from vocabulary errors and provide verified terms where supported. Solve a problemAudio plays at the wrong speed or pitch: how do you verify sample rates and headers?For abnormal speed or pitch, inspect actual sample data, headers and processing steps. Changing a sample-rate label differs from actual resampling and can directly change playback speed and pitch. Solve a problemBrowser model requests are blocked cross-origin: how do you distinguish CORS from authentication?For cross-origin blocking, inspect browser CORS rules separately from authentication instead of calling every console warning a key error. Check preflight and actual requests for permitted origins, methods and headers. Solve a problemThe same character looks different across images: how do you compare references and settings?For character drift across images, fix references and required features before comparing composition, prompts and generation settings separately. Repeating a character name does not guarantee identity, and a fixed seed does not lock a person. Solve a problemOld conversation content affects a new task: should you start a new chat or summarize the history?If old conversation content affects a new task, compare the same new instruction in a minimal fresh conversation to determine whether history is responsible. Inspect the actual messages, retain valid constraints and remove unrelated tasks or explicitly expired requirements. Solve a problemRAG loses a rule’s conditions or exceptions: how should chunks be repaired?When a RAG answer omits a condition or exception, inspect whether the actual context preserves that sentence before switching to a larger model. Chunking may separate a rule from its exception, or context assembly may drop neighboring text after retrieval. Solve a problemLocal builds pass but CI fails: how do you compare environments, lockfiles and commands?When local builds pass and CI fails, fix the same commit and compare actual commands, working directories and dependency installation before investigating environment differences. Do not begin by deleting the lockfile or ignoring failing checks. Solve a problemA long prompt still performs poorly: how do you retain essential requirements and remove distractions?Prompt length should follow the constraints the task needs, not the assumption that more words are better. Keep a working short baseline and add only guidance or examples that affect results. Solve a problemAn article has a recent update date but outdated content: how can substantive updates and dates be checked?For recent dates on stale content, inspect changing facts and separate editorial revision, source checking and build time. Refreshing a timestamp does not update facts. Solve a problemThe input fits the context window, so why does the answer stop early?An accepted input only shows that the request did not immediately exceed its input limit; it does not guarantee enough room for the answer. Read the service's stop reason first, then distinguish an output cap, a stop string and a network interruption. Solve a problemEmail draft recipients and body do not match: how do you verify mapping before sending?For mismatched recipients and body, verify fields by stable source record instead of only repairing the greeting. Address, name and business content must belong to the same confirmed object. Solve a problemEmbedding dimensions no longer match after changing models: how do you migrate the index?After changing an embedding model, check the new vector length against the index field definition, then verify that documents and queries use the same model. Even equal dimensions do not make different models' vector spaces interchangeable. Solve a problemThe same questions are reused for prompt tuning: how do you detect evaluation overfitting?Repeated prompt tuning on the same questions can raise scores through overfitting rather than general improvement. Track examples used in edits or demonstrations and reserve independent final tests. Solve a problemExcel processing removes identifier leading zeros: how do you check field types and exports?For lost leading zeros, separate identifiers from quantities and inspect automatic numeric conversion on import. Display formatting may not restore lost values, and long identifiers can lose precision. Solve a problemThe structured result parses but its values are wrong: how do you check structure and facts separately?For parseable objects with wrong values, separate syntax, Schema and factual support. A valid number does not establish the correct amount, and a string type does not establish a source-supported name. Solve a problemFAQ markup contains questions absent from the page: how do you align structured data with visible content?If FAQ markup includes invisible questions, align each answer with actual content, removing unsupported markup or adding useful visible material. Valid JSON-LD does not establish semantic compliance or display. Solve a problemA model still gets the format wrong after examples: how do you check the examples for ambiguity?When examples do not produce the correct format, check their consistency and coverage of realistic exceptions. Incidental example details may be learned as rules, so distinguish variable values from fixed structure. Solve a problemFiltered spreadsheet updates hit the wrong rows: how do you locate records by stable ID?For wrong-row writes after filtering, distinguish visible numbering, physical position and business identity. Sorting and hiding alter display order, making positional mappings unsafe. Solve a problemCharacters or products change across AI video shots: how can inconsistent segments be located using keyframes?For generated-video identity or product drift, inspect shots and keyframes rather than only the cover or endpoints. Fix references and required features and examine changes at action transitions. Solve a problemAI suggests a nonexistent library function: how do you verify the API for your installed version?When an AI-proposed function does not exist, check the installed package, import path and version before consulting that version’s official API. A plausible name is not grounds for inventing more parameters. Solve a problemEnglish search results open Chinese content: how do you verify language versions and hreflang?For English results opening Chinese pages, inspect the actual English URL, body and redirects before hreflang and canonical tags. Language annotations do not translate Chinese, and all versions canonicalizing to Chinese can affect interpretation. Solve a problemProduct codes and order IDs retrieve the wrong records: how should keyword and vector search cooperate?When exact identifiers retrieve poorly, first verify that parsing and indexing preserved their full values, then separate exact matching from semantic retrieval. Vector similarity represents meaning but does not guarantee correct distinction between codes differing by one character. Solve a problemCompression changes image colors or transparent edges: how do you compare export settings?When compression changes colors or transparent edges, preserve the original and inspect color management, format conversion and resizing separately. Smaller files do not establish visual equivalence; exports may discard profiles or transparency. Solve a problemGenerated posters contain Chinese typos: when should you regenerate or add editable text?For Chinese typos in generated posters, verify the intended copy independently before choosing localized repair or a separate text layer. Letter-like pixels do not guarantee correct characters, and OCR can miss errors. Solve a problemPrompt edits change too much of an image: how do you isolate effective wording?When prompt edits change too much, fix controllable model, reference and output settings and alter one description at a time. Random generation varies, so one before-and-after pair cannot establish a word’s effect. Solve a problemAI-generated tests merely repeat the implementation: how can examples that expose real errors be added?When tests repeat implementation logic, both may share the same bug. State observable requirements and derive expectations from manually explainable samples or independent evidence, then add relevant boundaries and properties. Solve a problemInpainting changes products outside the mask: how do you inspect masks and output differences?When inpainting changes unmasked content, inspect mask polarity, dimensions and coordinates and establish whether the tool promises pixel-level preservation. Some pipelines re-encode the entire image and alter unmasked regions. Solve a problemA requested JSON response includes explanations or wrong fields: how do you validate it reliably?Valid JSON establishes parseable syntax, not correct fields or business meaning. Distinguish explanatory wrappers, missing fields and wrong types, then check whether the service supports the required schema constraints. Solve a problemA knowledge base still answers from old documents: how do you verify version updates?When a knowledge base cites an old version, inspect the new record’s indexing state and the query’s version filters, then identify the actual record cited in the answer. Changing a source file does not guarantee that old records were updated or deleted. Solve a problemA file uploaded but cannot be retrieved: how do you check parsing, indexing and querying?A successful upload does not establish that content is retrievable. Record the document identifier, locate a known passage in parsed output, confirm that its chunk exists in the current index, and then inspect query conditions. Solve a problemLLM judge scores fluctuate: how do you test rubrics and evaluation bias?For unstable model judging, fix rubric, judge version and input order and test preferences for position, length or style. A confident rationale does not establish reliable scoring. Solve a problemAdding llms.txt did not increase AI citations: how does it relate to search-index eligibility?If llms.txt does not increase citations, separate its navigation purpose from Google eligibility. It is a proposal for adopting consumers, not a Google indexing or ranking switch. Solve a problemA local model runs out of memory: should you change quantization or shorten the input?First distinguish memory exhaustion while loading weights from failure after the input grows longer. For the former, check weight size and quantization support; for the latter, also inspect the KV cache, batch size and generation length. Solve a problemLong transcription loses several minutes: how do you compare segment coverage and merging?When a long transcript loses content, compare source duration, segment manifests and each task result to locate segmentation, recognition or merging failure. A successful final task does not establish complete coverage. Solve a problemA long-document summary omits key passages: was the input truncated or the extraction request unclear?When a long summary omits important material, first confirm that the complete source reached the model, then check whether the task explicitly identifies facts that must be retained. Truncation and failure to extract a point are different problems; adding “read carefully” does not distinguish them. Solve a problemMCP is connected but a tool fails: how do you locate the failing stage?An MCP connected status only establishes that a connection stage completed; it does not establish that the business API behind a tool works. Check initialization, tool discovery and execution separately, starting with one minimal request without side effects. Solve a problemMeeting minutes turn a suggestion into a personal commitment: how can attribution be checked against the recording?When minutes turn a suggestion into someone’s commitment, trace the sentence to audio and the verbatim transcript and check speaker and tone. Speaker labels do not establish real identity, and discussing a plan does not mean accepting a task. Solve a problemMerged cells break imports: how do you restore structure before AI processing?For misaligned imports from merged cells, identify merged ranges and stored values before assigning headings or group values to later rows. Unmerging or filling every blank can fabricate missing information. Solve a problemDate or department filters remove every result: how do you diagnose metadata and conditions?When date or department filters produce no results, inspect field values and types on a known record, then combine conditions one at a time. Text dates, timezone differences and missing fields can exclude apparently valid material. Solve a problemAI-generated code fails on another machine: how do you check dependencies and startup instructions?When AI-generated code fails elsewhere, check declared runtime dependencies, runtime versions and startup steps before comparing the error. Packages coincidentally installed on the author’s machine do not establish complete project dependencies. Solve a problemA field is absent from the source: how do you make AI mark it missing instead of guessing?Fields absent from the source should be marked missing rather than completed from general knowledge. Distinguish information not stated, extraction loss and mapping errors before defining the unknown-value representation. Solve a problemUnit tests pass but the real API fails: how do you cover contracts missed by mocks?If unit tests pass but the real API fails, compare mock responses against the actual contract, including fields, status codes and timeouts. An always-successful mock verifies only an idealized path. Solve a problemThe first model request is slow: how do you separate loading from generation time?For a slow first response, time process startup and model loading separately, then record the first token and generation completion rather than calling the whole wait inference time. Compare cold and already-loaded requests with identical input to locate the dominant cost. Solve a problemHow do you compare models fairly with matching prompts and sampling settings?For model comparisons, fix tasks and inputs, inspect comparable sampling, output budgets and tools, and score with the same rules. Identical parameter names need not have identical effects, and defaults can differ. Solve a problemA model keeps repeating sentences: how do you distinguish decoding settings from input problems?First determine whether repetition originates in the input, decoding settings or application code that appends previous output twice. Fix one failing input, inspect whether the raw response already repeats, and compare sampling and repetition controls one at a time. Solve a problemA model upgrade breaks a working task: how do you preserve evidence and verify a rollback?Compare the upgraded model with the previous version using identical input and settings before attributing a failure to the model rather than prompts, tools or API defaults. Save the version identifier and replayable failing examples before deciding on rollback. Solve a problemMixed Chinese and English becomes the wrong language: how do you verify recognition settings?For wrong-language mixed speech, compare audio with recognition settings and distinguish transcription, translation and term normalization. Automatic detection does not guarantee correct identification of every short word, and extra candidate languages may not help. Solve a problemEnglish results are strong but Chinese tasks fail: how do you evaluate languages separately?Evaluate languages and actual tasks separately rather than infer Chinese performance from English totals. Translation can alter meaning or difficulty, and cultural knowledge and output budgets affect results. Solve a problemNews details are noindex: how do you build indexable in-depth pages for important stories?Important noindex news can support a separate in-depth page addressing lasting questions, but it needs independent value rather than a retitled duplicate. Define mechanisms, impacts and usage conditions readers need and add direct evidence. Solve a problemAn agent picks the wrong overlapping tool: how do you clarify tool selection?When overlapping tools cause wrong selections, first define the objects and effects of each tool, then improve names and descriptions. More tools do not automatically make available actions clearer. Solve a problemA page opens but is not indexed: how do you check crawling, eligibility and canonical URLs?For accessible but unindexed articles, inspect access and crawling separately from eligibility and canonical selection. Browser success establishes one working access path, not search processing. Solve a problemParallel agents overwrite shared state: how do you prevent lost updates?When agents share files or state, determine whether updates were lost during writes based on concurrent reads or during merging. Separate branches can isolate code but do not automatically isolate a shared database or external resource. Solve a problemPDF text is visible but cannot be extracted: how do you distinguish scans and broken text layers?When visible PDF text cannot be extracted, distinguish scans, true text and incorrect hidden layers. Correct rendering does not establish extractable characters, and copyable output need not be correct. Solve a problemPDF table rows become misaligned in RAG: how do you validate parsed evidence?When answers misalign PDF table values, first inspect whether extraction retained headers, row and column relationships and footnotes, then adjust generation prompts. A visually orderly table may become merged neighboring rows in linear text. Solve a problemAn AI product image changes colors or parts: how can each detail be checked against the actual product?When AI changes product colors or components, establish required features from real photographs and formal specifications and check each generated image. Attractive imagery and reference inputs do not guarantee accurate product representation. Solve a problemA document tells AI to ignore the user: how do you verify that it cannot redirect the task?A document's instruction to ignore the user is content to process, not new authorization. Check whether the model or application promoted source text into higher-priority instructions and allowed it to affect tool execution. Solve a problemA model ignores output requirements: how do you find conflicting prompt instructions?When requirements are ignored, clarify conflicting or ambiguous conditions and inspect the order of instructions actually sent by the application. More text does not mean clearer constraints. Solve a problemA prompt change worsens results: how do you compare template versions and keep a recovery record?When a template change worsens results, associate outputs with specific versions and compare on the same test set rather than judging the latest answer alone. Save the recoverable old template and current diff and ensure the model and evidence did not change simultaneously. Solve a problemA model confuses yuan with ten-thousand-yuan units, or percentages with percentage points: how do you check units?Unit checks require the original value, unit and conversion rule rather than a bare number. Retain original units, define target units and rounding, then verify results through reproducible calculations. Solve a problemQuantization increases number and format errors: how do you isolate the quality loss?If quantization introduces number or formatting errors, compare with an unquantized version using the same source model, prompt and decoding settings. Confirm that the backend supports the format before attributing differences to weights rather than runtime parameters. Solve a problemA recording is quiet or noisy: how can recognition be compared before and after volume adjustment?For quiet or noisy recordings, establish whether intelligible speech exists before comparing gain and noise reduction. Amplification also increases noise, and aggressive denoising can remove speech detail. Solve a problemRAG cites a document the user cannot access: how do you verify retrieval permissions?If a knowledge-base answer includes unauthorized content, stop treating it as a trusted result and check that access constraints apply before retrieval supplies generation. Hiding a link in the interface does not prevent body-text disclosure. Solve a problemRAG invents answers without evidence: how do you test refusal and requests for missing material?When a knowledge base answers without evidence, check whether retrieval supplied irrelevant material and whether generation treated similar text as support. Build clearly answerable and unanswerable questions, requiring explicit limits and requests for specific missing material. Solve a problemRAG retrieved documents but answered incorrectly: how to trace the causeTrace incorrect RAG answers through retrieved passages, document versions, chunking and unsupported generation, with a clearly labeled synthetic example. Solve a problemThe model refuses normal requests: how do you distinguish legitimate limits from false refusals?For excessive refusals, define permitted business scope and distinguish safety refusal, insufficient evidence and interface failure. Not every absent answer is a false refusal, and improved pass rates must not remove legitimate limits. Solve a problemSearch results are wrong: how do you distinguish poor recall from incorrect reranking?For poor retrieval, first check whether the correct passage appeared among candidates before reranking. Reranking cannot invent a candidate that was never retrieved. Solve a problemRetrieval hits are high but answers are untrustworthy: how do you evaluate recall and groundedness separately?When retrieval hits are high but answers are unreliable, evaluate evidence retrieval separately from whether actual context supports answer claims. A relevant document in candidates does not prove use of the correct passage, and citations do not prove conclusions. Solve a problemThe same question gets different answers: how do you control randomness without losing useful variety?Control randomness by recording the model version, sampling configuration and input, not just a seed. For stable-format tasks, compare a nonsampling baseline first; for creative tasks, define which parts may vary. Solve a problemScheduled summaries run at the wrong time: how do you check time zones and daylight saving?For wrong-time summaries, inspect timezone, schedule and actual triggers and separate intended from actual start. Server defaults may differ from the target region, and fixed offsets do not represent every daylight-saving change. Solve a problemA transcript mixes two speakers: how do you assess diarization reliability?When diarization mixes speakers, replay transitions, interruptions and overlap and compare labels and ranges. Automatic labels represent distinguished voices rather than proven names. Solve a problemSpreadsheet and calendar dates differ by a day: how do you check timezones and date serials?For a one-day discrepancy, separate dates, local times and absolute instants before checking serial-date and timezone conversion. Birthdays and all-day items should not be arbitrarily treated as UTC instants, while timed events require a region. Solve a problemAn AI spreadsheet update replaces formulas with fixed values: how can formulas be protected and results verified?When updates replace formulas with values, inspect whether reads returned expressions or cached calculations and identify the write range. Equal displayed results do not establish preserved dependencies. Solve a problemA streaming answer disconnects midway: how do you distinguish completion, failure and truncation?For interrupted streams, inspect final business events and stop reasons rather than equate connection closure with completion. Network chunks need not contain complete SSE events, and errors may arrive after an initial success status. Solve a problemStreamed JSON parses only sometimes: how do you buffer fragments and validate completed output?For intermittent JSON parsing, distinguish network chunks, event-level JSON and the final business object. Chunks can end inside strings or multibyte characters and are not complete objects. Solve a problemSubtitles drift progressively out of sync: how do you distinguish timebase and segment offsets?For progressive subtitle drift, record offsets near the beginning, middle and end and distinguish fixed shifts, proportional drift and jumps after splices. Repeated global shifting does not repair an incorrect speed ratio. Solve a problemEvaluation includes prices, versions or dates: how do you avoid grading with stale answers?For time-sensitive facts such as pricing rules, versions and dates, specify an as-of time and scope and verify reference answers before grading. Disagreement between an old reference and a new response is not automatically model error. Solve a problemThe right tool was selected but its arguments are wrong: how should validation work?When the correct tool still fails, compare its raw arguments with the current schema and separate missing fields, type mismatches and invalid business values. A correct tool name does not establish correct object selection. Solve a problemAn agent reads only the first page: how do you verify pagination completeness?When an agent processes only the first page, check that the next-page marker reaches the subsequent call and stop according to the service’s own termination rule. Data on the first page does not establish completeness, and a short page is not necessarily the last page. Solve a problemA connected tool reports permission denied: how should access be diagnosed?When a tool reports insufficient permission, check its actual account, target resource and authorization scope before expanding access. Successful authentication establishes identity, not permission for a specific operation. Solve a problemA tool succeeds but the agent cannot see its result: how do you repair feedback?If a tool succeeds but the agent reports no result, check that the result enters the next request in the correct format and under the correct call ID. A successful tool log does not establish that the model received its content. Solve a problemA write tool times out: how do you retry without creating duplicates?After a write timeout, mark the outcome unknown rather than assuming nothing executed. Check business records or a request identifier before deciding to retry. Solve a problemAn image shows a transparency checkerboard but has no transparent background: how can its alpha channel be checked?A checkerboard does not establish transparency; its pattern may be painted into the pixels. Place the image on a solid background, then inspect valid transparency data and background opacity. Solve a problemTTS mispronounces Chinese words and English abbreviations: how do you correct and review them?For mispronounced words and abbreviations, establish the intended reading before using supported dictionaries, phonemes or pronunciation markup. Changing spelling may corrupt captions, and repeated generation may not fix the issue reliably. Solve a problemSVG or design exports look blurry at small sizes: how can rasterization be distinguished from scaling problems?For blurry SVG or design exports, establish whether the source is truly vector content and inspect output pixels and final display scaling. An SVG can embed a low-resolution bitmap, and a larger container creates no detail. Solve a problemVertical crops hide subtitles or subjects: how do you check framing and platform safe areas?When vertical conversion hides subjects or subtitles, establish canvas and platform overlays before adjusting picture and text separately. Center cropping often removes side content, and original caption positions may not fit vertical screens. Solve a problemBatch rendering leaves missing images or black clips: how can resource loading and timeline coverage be checked?For missing images or black shots in batch renders, map failures to resource requests and time ranges to separate incomplete loading, missing files and timeline gaps. An open page does not establish ready images and fonts, and a successful renderer exit does not verify every shot. Solve a problemExported captions are too small: how do you check readability at the final display size?For tiny exported captions, inspect font size, line length and background contrast at actual playback size rather than an enlarged editor preview. High resolution does not ensure large enough on-screen text. Solve a problemVideo works locally but fails on phones or websites: how do you diagnose encoding compatibility?When video works locally but fails elsewhere, inspect actual container, video and audio codecs rather than infer compatibility from its extension. Player support, pixel format and media requests all matter. Solve a problemVideo cuts truncate words: how do you verify audio boundaries instead of relying on caption timestamps?For clipped word endings or half-sentences, replay the original and inspect actual cut boundaries before treating subtitle times as precise speech boundaries. Automatic timestamps can drift, and punctuation-based cutting can remove breath and endings. Solve a problemExported video has lip-sync errors: how do you check frame rate and timebase?For export lip-sync errors, compare timestamps, audio starts and durations before inspecting frame-rate conversion and speed changes. Relabeling frame rate need not preserve playback time. Solve a problemA video summary confuses observation and inference: how do you verify its sources?When a video summary presents inference as fact, map claims to audio, visible footage or other sources and separate observation from interpretation. Captions alone do not establish a visible event, and inferred emotion does not prove intent. Solve a problemVideo understanding misses brief events: how do you test sampling and input coverage?For missed brief events, establish actual processed ranges and sampling before judging recognition. Sparse frames from a long video may never include the event. Solve a problemVoice exports crackle or clip: how do you inspect peaks, mix gain and encoding?For distorted exports, compare original tracks, the mix bus and the final file to locate the changing stage. Individually valid tracks can exceed limits when summed. Solve a problemAI rewrites remove Word headings and table formatting: how do you verify structure and exports?For lost Word formatting, check plain-text reconstruction and whole-paragraph replacement that discards run-level formatting. Correct content does not establish preserved headings, hyperlinks or tables. Solve a problemAutomation reruns create duplicate records: how do you define business keys and idempotency?For duplicate automation records, define a stable key per business event and inspect trigger replay, retries and concurrent creation. Lookup-then-create can still race without atomicity.