littlebigbrain 0.6.0

Python client for the little big brain graph + hybrid search HTTP API

little big brain Python SDK

The Python client for little big brain. Use it to ingest records, build BM25 + vector + graph indexes, search with authorization filters, traverse the graph, and turn retrieval feedback into training data.

pip install littlebigbrain

Five-minute start

from lbb import LbbClient

with LbbClient(
    "https://db.eu.littlebigbrain.com",
    api_key="lbb_sk_live_...",
    graph="main",
) as lbb:
    graph = lbb.graph("main")

    graph.facts.create({
        "triplets": [{
            "source": {"type": "CONCEPT", "name": "handbook", "key": "doc:42"},
            "relation": "RELATED_TO",
            "target": {"type": "CONCEPT", "name": "vacation policy", "key": "passage:42:1"},
            "evidence": "Employees receive 25 days of annual leave.",
        }],
    }, idempotency_key="doc:42:v1")

    lbb.indexes.run(wait=True)
    results = lbb.graph_search({
        "query": "how much annual leave do employees get?",
        "targets": ["entities"],
        "top_k": 10,
    })

Stack keys belong on a server, worker, or secret-backed notebook—not in a browser bundle. Safe reads and idempotency-keyed writes retry transient errors and honor Retry-After.

Enterprise-search integration

For an enterprise-search retrieval adapter, keep the application database for users, connectors, tasks, and migration cursors. Put searchable documents, passages, graph facts, embeddings, BM25/ANN/adjacency indexes, ontology review, and retrieval feedback in LBB.

The production sequence is:

1. map a connector batch to stable-keyed document, passage, provenance, and edge records;
2. call graph.facts.import_ndjson(..., index=False, idempotency_key=...);
3. submit one durable index job with index_submit, then reconnect with index_job;
4. translate the product's ACL/scope filter into native set overlaps filters;
5. call graph_search with projected fields and hydrate only the final top-k;
6. submit grade-3 feedback for sources cited in the grounded answer;
7. evolve ontology changes through draft → validate → promote/reject.

For an LLM query planner, use lbb.context.suggest(...) to fill grounded schema/value prefixes and lbb.context.resolve(...) to snap free-text guesses onto real vocabulary. resolve uses the graph's managed embeddings when configured. Record adopted suggestions and accepted/rejected/corrected plans so the feedback can train a smaller planner and suggest ranker.

See the complete enterprise-search integration guide for the record model, migration plan, acceptance gates, and capability mapping.

Useful surfaces

# Bulk ingestion (flat or generated typed property values, including sets).
graph.facts.import_ndjson(records, strict=True, index=False,
                          idempotency_key="connector:batch:17")

# Durable indexing and training.
job = lbb.index_submit({}, idempotency_key="index:head:147")
status = lbb.index_job(job.job_id)
cancelled = lbb.cancel_index_job(job.job_id)
gc = lbb.index_gc_submit({"dry_run": False}, idempotency_key="gc:2026-07-15")
gc_status = lbb.index_gc_job(gc.job_id)
train = lbb.train_submit({"kind": "fusion", "force": True},
                         idempotency_key="fusion:gate:7")
progress = lbb.train_job(train.job_id).progress

# Typed namespaces.
answer = lbb.context.ask({"question": "what changed?"})
ontology = lbb.ontology.view(counts=True)
rows = lbb.query.sparql({"query": "SELECT ?s WHERE { ?s ?p ?o }"})

# Cursor-safe iteration.
for entity in lbb.entities.iter(fields=["text", "acl"]):
    print(entity.name, entity.attributes)

# Branch deletion protects the final live branch; graph deletion removes all branches.
lbb.graph("main", branch="review").delete_branch(confirm="review")
lbb.graph("main").delete(confirm="main")

LbbClient and AsyncLbbClient expose the same capabilities. Preferred namespaces return generated Pydantic models; compatibility helpers return parsed dictionaries. LbbError includes HTTP status, structured code, parameter, request ID, and documentation URL. raw_request(...) exposes attempt count, elapsed time, response headers, build commit, and replica.

Major capability areas

• graph(...).facts: commit, dry-run, retract, NDJSON/RDF import
• search / graph_search: lexical, BM25, vector, hybrid, filters, facets
• indexes: full build, durable submit/status, delta, garbage collection
• entities: projected reads, native filtering, cursor-safe iteration
• ontology / schema: define, evolve, induce, draft review, SHACL lifecycle
• query: SPARQL, structured query, analytics, SHACL, inference, conflicts
• context: grounded ask, suggest, resolve, decode, groundability
• feedback/training: labels, export/summary, durable trainer jobs and progress; typed suggestion/planner supervision helpers with validation before transport, automatic idempotency keys, and durable receipt/trainability acknowledgements
• temporal graph: traversal, state, history, lineage, snapshot pins

Generated models come from the bundled OpenAPI contract and are available in lbb.models.

Develop

python3 -m venv .venv
.venv/bin/pip install -e ".[dev]" httpx pydantic
ruff check lbb tests
mypy lbb
pytest tests

lbb/models.py is generated. Change the Rust API types and regenerate clients instead of editing it by hand.