Vector database (vector index)
A vector database is a store built to hold embeddings and answer nearest-neighbour queries over them quickly — given a query vector, return the closest records out of millions. It is the retrieval layer underneath semantic search, RAG pipelines, and most AI shopping assistants, and it is only as good as the product text that was embedded into it.
What it does that a normal database does not
Relational databases answer exact and range queries: rows where voltage = 480. A vector index answers a different question: which stored vectors are closest to this one. Comparing a query against every vector would be too slow at catalog scale, so these systems use approximate nearest-neighbour algorithms such as HNSW or IVF, which trade a small amount of recall for large speed gains.
The category spans dedicated engines (Pinecone, Weaviate, Qdrant, Milvus), extensions to existing databases (pgvector for Postgres), and vector capabilities inside search engines you may already run (Elasticsearch, OpenSearch, Vespa, Solr). For most teams the last option is the cheapest starting point, because the infrastructure is already there.
Hybrid retrieval, and why B2B needs it
Pure vector retrieval fails on the two query types that dominate industrial catalogs: exact part numbers and numeric constraints. HB-150-50K-U has no semantic content to embed, and "22 kAIC" sits close to "10 kAIC" in vector space because they are the same kind of statement.
The standard fix is hybrid search: run lexical matching (BM25 or exact-match) and vector similarity in parallel, then fuse the result sets, often with reciprocal rank fusion, and apply structured filters on typed fields as a hard gate. The filter step is the important one for distributors. It only works if voltage is stored as a number with a unit rather than as text inside a description, which puts the burden back on normalization and fill rate.
What you feed it
The unit stored is usually a chunk plus its metadata: the vector, the source text, the SKU, the category, and whatever attributes you want to filter on. Metadata is where the leverage sits, because a filter on branch availability or ETIM class narrows the candidate set before any similarity math runs.
A common mistake is embedding the marketing description and nothing else. A better input for technical products is a composed text that concatenates manufacturer, part number, category, and the spec block into one passage, so the vector encodes what the product actually is rather than how it was advertised.
Your index versus theirs
Two different projects get confused under one word. Building a vector index over your own catalog powers on-site search and internal assistants; you control chunking, hybrid weighting, and refresh. Being retrieved by someone else's index — ChatGPT's, Perplexity's, a customer's procurement assistant — gives you control over exactly one thing: the content they can reach.
The encouraging part is that the same preparation serves both. Complete attributes, self-contained passages, clean identifiers, and server-rendered pages improve your own retrieval quality and everyone else's at the same time.
Frequently asked questions
Do we need a vector database to show up in AI search?
No. Public AI engines build their own indexes from your pages and feeds. You would run a vector database to power your own site search or an internal assistant. The overlap is that both depend on the same underlying content quality.
What is hybrid search?
Running keyword and vector retrieval together and merging the results, usually with structured filters applied as hard constraints. It exists because vector similarity handles intent well but handles part numbers and numeric specs badly, and keyword matching has the opposite profile.
Can we just add vector search to our existing search engine?
Usually yes. Elasticsearch, OpenSearch, Vespa, Solr and Postgres via pgvector all support vector fields alongside lexical indexes, which makes hybrid retrieval simpler to build than running a separate vector store and stitching results together.
What most often makes catalog vector search disappointing?
Embedding thin text. If the indexed passage is a title and a two-line marketing blurb, every record in a category embeds to nearly the same place and relevance is arbitrary. Enriched, spec-bearing text is what gives records distinguishable positions.