BuildSpark AI: practical AI concepts for builders. New here? Start with Tokens.
Home · Retrieval & Grounding · 09 · Vector databases
Concept card · Retrieval & Grounding
A vector database is a store built to keep embeddings and return the nearest ones.
A store built to keep embeddings and return the nearest ones.
Oct 18: First card. The picture ranks three chunks by meaning, then by a product code. Sources are Azure AI Search, pgvector, and Pinecone's current search overview. Pinecone Local is not a source.
Azure AI Search's vector search overview says vector search indexes numeric representations and returns the vectors most similar to the query. It says you can use the service as a pure vector index for a knowledge base, and that a nearest-neighbors algorithm returns the requested k neighbors. pgvector's README describes vector similarity search for Postgres, including exact and approximate nearest neighbor search, and a query that orders rows by distance. Pinecone's search overview says semantic search uses dense embeddings, and that searches consume read units. Pinecone's 2026 changelog says Pinecone Local reached end of life on October 7, 2026. This card cites the current search overview, not Pinecone Local. The picture ranks three labelled chunks. It is not a query against any of those stores.
Azure's overview says vector search is available in all regions and on all tiers at no extra charge, and that generating embeddings might incur charges from the model provider. That "no extra charge" is vendor-reported for the search feature, not for the embedding call. Pinecone's overview says searches consume read units and points to a pricing page. This card does not copy a read-unit rate. pgvector stores the vectors in Postgres.
Azure's overview says vectors can match text that is conceptually alike, and gives "dog" and "canine" as the example. A product code can still miss when the match is by meaning. The hybrid search card is the one that runs words and vectors together.
The nearest chunk is not a decision about who may read it, and it is not a check that the text is current. The store returns neighbors. A filter decides who sees them.
Name the store you already run. If the support rows live in Postgres, pgvector can keep the vectors beside them. If you already use Azure AI Search, the vector fields can live on that index. Write down where the embedding is computed. That call is a separate bill.
Opening a second database before you know the embedding call is the line that grows. Azure's overview separates the search feature from the embedding charge.
Azure's overview says vector search has no extra charge. Which part can still cost money?
Support-docs answer bot uses this concept in one build. The set on this tab is still the note below.
Coming soon. On the support-docs answer bot, the vector column is a local hash, not a hosted store. Write down which store you would use for those eight chunks, and where the embedding call would sit.
Coming soon. How Azure returns k nearest neighbors, how pgvector orders by distance, and which Pinecone search the current overview still documents.