Vector database A store that finds the passages whose meaning is closest to a question.
Short answer. A vector database stores embeddings, lists of numbers that stand for the meaning of text, together with the text and its metadata, and returns the stored items closest to a query vector. Exact search compares the query with every vector, which is perfectly accurate and fine for small collections. Approximate indexes such as HNSW, a layered graph, trade a little recall for speed on large ones. One hundred thousand passages at 1,024 dimensions take about 410 MB as 32-bit numbers. You may not need a separate product: PostgreSQL’s pgvector extension does exact and approximate search inside an ordinary database.
01
How it works
Each record holds an id, a vector, the original text and metadata such as file name, page, date and who may read it. To search, the question is embedded with the same model, and the database returns the k records whose vectors are nearest by cosine distance, inner product or Euclidean distance.
pgvector’s documentation puts the choice simply: by default it performs exact nearest neighbour search, which provides perfect recall, and you can add an index to use approximate search, which trades some recall for speed. Its HNSW index builds a multilayer graph and has a better speed and recall balance than its alternative, IVFFlat, but builds more slowly and uses more memory. The HNSW paper reports that the structure allows logarithmic complexity scaling.
02
Exact and approximate search
Approach
Recall
Speed as the collection grows
Memory and build
Fits
Exact scan
Perfect
Slows in proportion to size
No extra index
Small collections, or filters that leave few rows
HNSW index
High, tunable (pgvector default ef_search is 40)
Stays fast
More memory, slower build
Large collections with many queries
IVFFlat index
Lower than HNSW for the same speed
Fast
Less memory, faster build
Large collections where build time matters
03
Permissions and filtering
An office rarely wants everyone to see every file. The usual design stores an access group with each chunk and filters on it. There is a trap. pgvector’s documentation warns that with approximate indexes, filtering is applied after the index is scanned: if a condition matches 10% of rows, then with HNSW and the default ef_search of 40, only 4 rows will match on average. A strict permission filter can silently return too few passages, and the answer may be built on the little that remains.
The remedies it lists include iterative scans, partial indexes for a few distinct values, and partitioning when filtering by many values. Whatever the product, enforce permissions in the database query, not in the prompt, and test with a user who can see only a small share of the files.
04
Do you need one?
A few thousand passages: exact search in an ordinary database, or even a file, is enough.
Tens of thousands to millions: an approximate index in PostgreSQL with pgvector, or a dedicated vector store.
Exact words matter, such as invoice numbers: add keyword search alongside the vectors.
Confidential material: protect the index like the documents. Embeddings can be partly inverted to recover text, and the store holds the original passages too. ANVE keeps everything in one database on the box, and full-disk encryption and secure boot, part of the appliance image, are not on ANVE yet (see its security page). For who may see which documents, ask any vendor, ANVE included, to put the design in writing before a paid pilot.
Operations: plan backups, encryption at rest and a way to delete a document’s chunks when the document is withdrawn.
05
Questions people ask
Is a vector database the same as a normal database?
No, but they can be combined. A vector database is built to find nearest neighbours among embeddings. Extensions such as pgvector add that ability to an ordinary database, keeping text, metadata and vectors together.
What is HNSW?
Hierarchical Navigable Small World, a graph index for approximate nearest neighbour search. It links each vector to near neighbours across several layers, so a search can hop quickly toward the closest items.
How much memory does a vector index need?
At least the raw vectors: chunks times dimensions times 4 bytes for 32-bit numbers. Approximate indexes add overhead, and HNSW uses more than IVFFlat.
Can I search by words and meaning together?
Yes. Many systems combine keyword matching with vector search, then re-rank the results. That helps with exact identifiers, which vectors match poorly.
06
What is not real yet
No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.
Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.
