Solutions / AI engineers

Test retrieval on your own data.
Replay any read by its commit.

Vector search, graph patterns and reasoning run in one SPARQL query. A read names its commit. Evals check the model calls.

[01] In your code

Call it from your agent.

Use the SDKs or the MCP tools. Results come with the rows, the commit and the plan.

The Query page: a SPARQL query with search:similarTo joined with the customer of each issue, five rows with scores

Plans you can read.

A search query reports its plan: nearest, filter_first or search_first.

Pinned reads.

Reads report served_at_seq. Pin a read with as_of_commit_seq to replay it.

Open interfaces.

MCP tools, TypeScript and Python SDKs, and plain HTTP.

The hosted MCP configuration with the server URL for the environment — Paste this into your agent's MCP settings.

[02] MCP

Your agent reads and writes the graph.

Run the server with npx for Claude Code, Cursor or Codex. Or connect the hosted server with OAuth sign-in.

lbb_inspect, lbb_query, lbb_commit and lbb_rdf for the graph.
lbb_embeddings, lbb_evals and lbb_models for search and models.
Open source under Apache-2.0.

MCP server

[03] What is included

SDKs, MCP and safe writes.

A hosted service in the EU, with your data in object storage.

@littlebigbrain/client

The TypeScript SDK for Node 18 and later, browsers and edge runtimes.

littlebigbrain

The Python SDK for Python 3.10 and later.

@littlebigbrain/mcp

The MCP server, over stdio or hosted with OAuth.

SHACL

A write that breaks your rules returns 400 with the reason.

Idempotency-Key

A retried write returns replayed: true and writes nothing twice.

compare

Compare what a query returns at two commits.

[04] How it works

Load, query, evaluate.

Step 01

Load.

Define classes and commit JSON records, or import RDF.

Step 02

Query.

SPARQL with search:similarTo, paths and entailment.

Step 03

Evaluate.

Goldens are scored by commit. Trials test cheaper models.

[05] FAQ

Questions from AI engineers.

Is it a vector database?

It is a knowledge graph with search built in. Vector search runs through POST /v1/search and inside SPARQL.

Which formats can I load?

JSON records, NDJSON up to 32 GiB per job, and N-Triples, Turtle, N-Quads or TriG.

How do I replay an answer?

Store served_at_seq with the answer. Run the query again with as_of_commit_seq.

Can I read my own writes?

Yes. Read with strong consistency, or wait for min_indexed_seq.

Next step

Test retrieval on your own data.Replay any read by its commit.