Knowledge Graph

The RVBBIT knowledge graph is a SQL-visible memory layer for semantic systems. It stores canonical nodes, aliases, edges, evidence, extraction runs, and merge history.

It does not hide a graph database behind Postgres. It makes model-derived facts inspectable, queryable, mergeable, and tied to receipts.

Core Tables#

Table	Purpose
rvbbit.kg_nodes	Canonical entities and concepts.
rvbbit.kg_aliases	Alternate labels for nodes.
rvbbit.kg_edges	Directed facts between nodes.
rvbbit.kg_evidence	Provenance for nodes or edges.
rvbbit.kg_merge_candidates	Review queue for possible duplicate nodes.
rvbbit.kg_extraction_runs	Bulk extraction audit trail.
rvbbit.kg_extraction_errors	Per-row extraction failures.

Evidence rows can record query_id, which lets a UI connect KG facts to semantic receipts and MCP invocations.

Assert Facts#

SELECT rvbbit.kg_assert_node('customer', 'Acme Corp');
SELECT rvbbit.kg_assert_node('issue', 'late shipment');

SELECT rvbbit.kg_assert_edge(
  'customer',
  'Acme Corp',
  'reported',
  'issue',
  'late shipment',
  confidence => 0.92,
  evidence => '{"text":"Acme reported late shipments after the warehouse move.",
                "source":"support_ticket"}'::jsonb
);

Read Context#

SELECT *
FROM rvbbit.kg_context(
  'customer',
  'Acme Corp',
  max_depth => 2,
  include_evidence => true
);

kg_context is the main read primitive for graph-backed RAG and UI panels. Use kg_neighbors for local traversal and kg_paths when the user asks how two entities are connected.

Extract Triples#

SELECT *
FROM rvbbit.triples_rows(
  'Acme delayed renewal after repeated fulfillment misses.',
  'customer risk'
);

Ingest any triple-shaped query:

SELECT rvbbit.kg_ingest_triples($$
  SELECT *
  FROM rvbbit.triples_rows(
    'Acme delayed renewal after repeated fulfillment misses.',
    'customer risk'
  )
$$);

Or extract from a table:

SELECT *
FROM rvbbit.kg_ingest_table(
  'support_tickets'::regclass,
  pk_col => 'ticket_id',
  text_col => 'body',
  graph => 'support'
);

Extraction runs and errors are stored so bulk ingestion can be retried or reviewed.

Merge And Resolve#

SELECT *
FROM rvbbit.kg_suggest_merges('customer', threshold => 0.86, limit_count => 100);

SELECT rvbbit.kg_accept_merge(42);
SELECT rvbbit.kg_reject_merge(43);

For fuzzy resolution, enable a Lance index per graph/kind:

SELECT rvbbit.kg_lance_enable('customer', graph => 'support', specialist => 'embed');
SELECT rvbbit.kg_lance_refresh('customer', graph => 'support');

Resolution order is exact alias, ready Lance index, then slower embedding scan.

Trust Model#

The KG preserves provenance when you write evidence and query_id as facts come from model calls or tool output. That makes it possible to explain where graph context came from instead of treating it as an unreviewable memory blob.