Retrieval that can show where its answer came from.

A retrieval pipeline usually stitches a vector store, a search engine and a graph together, and loses the trail between them. In Tessera DB the text index, the vectors and the graph read the same records, and each record carries its origin.

What it does.

Hybrid search on every tier

BM25 keyword search and HNSW vector search, run together and fused with Reciprocal Rank Fusion, so records strong in either rise to the top.

An embedder built in

A fresh server makes text vector-searchable with no setup. The built-in embedder is lexical: it matches spelling, not meaning. Set TESSERA_EMBEDDING_URL to use a semantic model of your choice.

Answers with a source

Every record carries who wrote it, through which door and from which load, so a retrieved passage leads back to where it came from.

GraphRAG (Enterprise)

Document chunking, embedding and retrieval in the database, with graph context around each hit.

Agent memory (Enterprise)

The database decides which fact is current: a new value for the same subject and predicate replaces the old one as current, and history is a separate call. Scored 100% on our own published test, which is not comparable with other vendors' scores.

Natural-language query (Enterprise)

POST /api/nlq turns a question into TQL using your schema and a language model you configure. It returns the TQL for review and does not run it. No model ships with the product.

Search, two ways, from the kit.

# Keyword (BM25)
curl -s -H "Authorization: Bearer $TESSERA_KEY" "$TESSERA_URL/api/search?q=invoice&limit=5"

# Meaning: plain text in, nearest records out
curl -s -X POST -H "Authorization: Bearer $TESSERA_KEY" -H 'Content-Type: application/json' \
  "$TESSERA_URL/api/search/vector-text" -d '{"text":"loan repayment","top_k":3}'

Results are scoped to the caller's tenant, and for end users to the rows their policies allow.

Not yet.

MCP server

A local MCP server exists but is not in any kit, because it reads the data directory below the HTTP layer, where licence checks, roles and query defence do not apply. No release date is set.

Approximate by nature

HNSW vector search is approximate. Exact-match and keyword lookups are exact; nearest-neighbour search finds very nearly all neighbours, and the kit documents its limits.