Natural-language retrieval over Little Big Brain. Use query for one phrasing, queries for reciprocal-rank fusion across phrasings, and follow_paths: true when you want bounded graph paths from text-resolved seed entities. When you can judge returned results, call lbb_commit mode=search_feedback with good=3, partial=1, bad=0 so Little Big Brain can build customer-specific qrels for embedding training.

Ask a natural-language question about the graph and get a grounded answer with citations. The database snaps the question to its real vocabulary (never invented), retrieves against the pinned snapshot, and answers. mode is resident_planner when a small resident model synthesized the prose, or grounding_only when it returns the grounded evidence for your own model to finish. Prefer this over lbb_search when you want a direct, cited answer rather than a ranked list; set execute: false to get only the grounding — the real vocabulary the question maps to — without retrieval. The response explain block reports how much the database narrowed (vocabulary candidates the question snapped to, plus retrieved entity/assertion counts) and the per-stage latency (ground / retrieve / synth / total ms), so you can see the pipeline that produced the answer.

Name the relation between two entities. The database narrows the candidates to the relations its type signatures admit for the (source type, target type) pair; if the pair admits exactly one, the database answers alone (mode: forced, no model call). Otherwise a small model fine-tuned on this graph's own edges picks from the narrowed set (mode: model_narrowed), constrained to real vocabulary so it can only emit a relation that can exist. You can OMIT the types — pass just the names and the database recovers each type by resolving the name to a real entity (echoed in resolved_source/resolved_target). Use it to fill a missing edge label, verify a relationship, or assemble structured triples. Returns the relation, the admissible candidates, and signature_forced.

Ground your terms to the graph's real vocabulary before you query or write, so you never guess a type, relation, or property name. Actions: complete — narrowed autocomplete: completes a prefix against the real vocabulary, optionally narrowed to the relations a (src_type, dst_type) pair actually admits (so you only propose relations that can exist); resolve — snap free text to the single nearest real vocabulary item by embedding/lexical similarity, never fabricating a name; audit — groundability report: signature sparsity, name semantics, sampled narrowing recall, and a narrow / narrow+finetune / lexical recommendation for this graph.

Read graph context and exact graph facts. Actions: guide, ontology, ontology_conformance, schema, schema_preview, schema_audit, rules, ontology_search, metadata, entity, edges, state, history, transitions, why, traverse. entity returns one node's metadata, scalar attributes, current state, edges, history, and observations — its edge/history arrays are a display sample capped at the detail limit (counts holds the true totals), so on a high-degree node the response carries an edge_sample block with the capped totals and ready-to-run paged reads; follow those (or call edges/history directly, paged by row_limit/offset with direction/relation filters and an as_of/as_of_commit_seq pin) to read the full set — total_count tells you when to stop. transitions returns the ordered state-transition log of an entity's relation with dwell time at each value (cycle-time/process analysis). metadata includes temporal_coverage — check it before attempting as-of/daily views: as_of_valid_time_degenerate or single_commit_time means every point-in-time query returns the same snapshot. Use ontology_conformance to check the live data against the ontology's own capped-cardinality rules (derived SHACL sh:maxCount, whole-snapshot, never blocks a write) — distinct from schema_audit, which runs the separately-published SHACL shape bundle.

Analytical and expert reads. Modes: structured (SPARQL-subset JSON body), sparql (SPARQL text), shacl, infer, retrieval_premises, analyze. Relations are https://littlebigbrain.com/r/NAME and types https://littlebigbrain.com/class/NAME (both lowercased); entities are content-addressed, so anchor a named one by its rdfs:label rather than building its IRI (see the sparql/structured field hints, and lbb_inspect action=ontology for exact names). For structured/sparql row paging, MCP owns limit/offset via row_limit/cursor; a cursor reuses the original query/body and pins continuation pages to the original head commit. When a single page's rows are complete server-side (row_page.has_more false) but too large for one MCP result, the envelope reports rows_shown (fewer than row_page.returned), keeps truncated: true, and hands back a next cursor that pages the same result set at a smaller row_limit — so a large fully-returned result is never silently cut without a way to read the rest.

Write graph facts, retract them, or label search results. mode=facts writes triplets/embeddings/properties; mode=retract removes a wrongly-added fact (by edge or by entity) without a full reset; mode=search_feedback labels query/result relevance after lbb_search (Feedback grades: 3=ideal/good, 1=partial, 0=bad; include query, search_id when available, target, rank, score). Explicit idempotency_key wins; when omitted, MCP derives a stable content hash so content-identical retries dedupe. Facts mode defaults edge_idempotency to append; pass skip_unchanged for re-runnable backfills.

Mutate stored graph configuration. Actions: define_ontology (create a new graph ontology), evolve_ontology (evolve an existing graph's ontology in place by name — add_entity_type / add_relation / add_property (a typed scalar field so entity_properties can write it) / widen_relation, rename and set-inverse/cardinality, and data-gated narrow/remove; bumps the ontology version, preserves every record, no migration), publish_schema (activate a previewed RDF/SHACL shape bundle), and define_rules (replace the branch's stored inference rules — body/head terms may be variables or fixed entities, and not_exists combinators add stratified negation for universal conditions).

Build or refresh persisted BM25, vector, and adjacency indexes so recently committed facts become searchable.

Branch lifecycle. Actions: create (fork a new branch off from_branch — the tool's branch argument names the NEW branch) and merge (validate-then-merge: replay from_branch's post-fork commits onto the scoped target branch — its fork parent — as ONE commit with event ids preserved; SHACL-validates the would-be merged state first and refuses with the report on violations; a fact superseded on the target after the fork wins over the branch's version, reported as a supersedure_race conflict; delete_source consumes the merged branch).

Remember a conversation: store the turns verbatim as an EPISODE evidence entity, then anchor + gate the supplied facts on an observe branch (LLM extraction cannot poison the main graph). Facts with both endpoints already in the graph are anchored; unanchored facts need confidence >= 0.8 to mint new entities, else they come back needs_review. auto_merge merges the branch onto the scoped branch when SHACL validation is clean (the validate-then-merge). Server flag-gated (--enable-observe). This build takes caller-extracted facts (each with a structured triplet); bare statements come back needs_review.