CogniGraph / Database
Four ways in.
One file on disk.
CogniGraph is an evidence-linked multi-model database written in Rust: JSON documents, typed edges, full-text and vector search in one process, backed by one file on disk, with room in the model for the text behind every relationship. CGQL brings document queries and graph traversal into the same request.
01 / Connected records
A document keeps its context.
An account record can keep its handle, display name and the case it belongs to. A handset is another record. A registered_on edge connects the two, so an application can ask about the account and follow the relationship to the handset it was found on.
Documents live in collections and use stable _key and _id identifiers. Edge collections use _from and _to to record direction. Your application decides which relationships belong in its model.
02 / Find your starting point
Words, meaning and paths.
Different questions need different entry points. An exact identifier can find one record. Full-text search finds wording. Vector search finds nearby embeddings. Graph traversal follows the relationships you stored.
| Approach | Useful when |
|---|---|
| CGQL | You know a collection, a condition or an edge to follow. |
| BM25 full-text | The words in a document matter. No embedding provider is required. |
| Vector search | You have an embedding and want nearby records. |
| Hybrid search | You want text and vector rankings combined through reciprocal rank fusion. |
| Graph-augmented retrieval | You want semantic starting points followed by connected records. |
CGQL supports vector search as a query source. Full-text, semantic and hybrid retrieval also have dedicated HTTP endpoints. See how the API fits together.
03 / The operating model
One process. Nothing beside it.
The Native backend uses redb as the durable primary store, with resident or paged operation and rebuildable search derivatives. The whole database is one server process next to your application: no cluster to run, no coordinator, no separate search service.
Text and graph workloads run without an embedding provider. Semantic workflows take embeddings supplied directly or produced by a configured local or remote provider.
Community ships with authentication, roles, rate limits, metrics, snapshot backup and the HTTP API, plus Docker packaging and a single-node Helm chart. Today it is a single-writer system; read replicas and standby are on the roadmap.
Run your product on it.
The Community database is free to use in production under the Functional Source License, including inside commercial products, subject to its terms. Governed construction and multi-tenancy belong to Enterprise.
Understand the two editions04 / Before you choose
A few direct answers.
Is CogniGraph a graph database or a vector database?
Both, and a document store. Use the parts your workload needs: a document or graph application never has to call a language model.
Can I use my AQL queries unchanged?
CGQL shares familiar concepts with AQL, but has its own grammar and supported features. Review the AQL migration introduction against your actual queries.
Does every edge have evidence and approval attached?
Ordinary edges follow your application's model, and the demo on the homepage shows the pattern: keep the source chunk as a record and point at it. Evidence-bound construction and signed governance are the Enterprise workflows that build and sign those edges for you.