What is a vector database? Similarity search, metadata, and source text
Understand vectors, identifiers, metadata, exact and approximate neighbors, model migrations, and access updates through manual retrieval.
Understand the concept
A vector database stores vectors and associated records and supports similarity search, often for RAG evidence. Records can include source IDs and metadata, depending on the product. It is not an answer-generating model or truth judge. Search depends on compatible representations and indexes; updates, permissions, and recovery need separate management.
Vectors and record identifiers link to actual material
Embed each manual passage with a stable ID and source position, retrieve similar IDs, and read the text for evidence. Vectors alone generally cannot restore the full source or verify quoted sentences.
Qdrant and Pinecone describe vectors and associated fields with different capabilities. Check whether embeddings are integrated, external, or accompanied by stored text. This explains roles without creating databases or benchmarking production services.
Assess speed together with coverage in approximate search
Approximate indexes reduce work for large collections but may differ from exhaustive nearest-neighbor results. Products can offer exact modes too. Evaluate evidence coverage rather than equating fast responses with best results.
Match dimensions and metrics to the embedding model. Equal dimensions do not align model spaces. Plan re-embedding and index switching with versions and sample comparisons so new queries do not silently search an incompatible old space.
Verify actual index state after updates or permission changes
Confirm index synchronization after source edits, deletions, or permission changes. Metadata helps filtering when maintained by trusted programs; do not retrieve unauthorized content and rely only on model concealment.
Test known data for insertion, stale-version removal, and recovery records. This is acceptance design without real database operations. Successful retrieval is infrastructure evidence rather than proof of factual correctness.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.