Architecture

borg_traverse()

A recursive CTE function for 1-2 hop graph traversal. Cycle-safe via path tracking. Scoped to a single namespace. Used by the graph_neighborhood retrieval strategy.

SELECT * FROM borg_traverse(
  p_entity_id := 'a1b2c3...',
  p_max_hops  := 2,
  p_namespace := 'product-engineering'
);
-- Returns: entity_id, entity_name, entity_type,
--          fact_id, predicate, evidence_status,
--          hop_depth, path

Core

Namespace Management

Admin

Why LLM in the write path?

Borg extracts structured knowledge (entities, typed facts, procedures) — not just text blobs. This requires an LLM. The trade-off is extraction cost and latency, but it runs offline so it never blocks queries. The alternative (embedding-only, like Ogham) gives you similarity search but not a queryable knowledge graph.

Why not Neo4j / FalkorDB?

Adding a graph database means syncing between two systems. Sync means drift. PostgreSQL recursive CTEs handle 1-2 hop traversal at the expected scale (hundreds of entities, thousands of facts). A dedicated graph DB is a measured-bottleneck escape hatch, not a starting point.

Why three-pass resolution instead of always-merge?

Collision is catastrophic — two different things merged corrupt every attached fact. Fragmentation is recoverable. The 0.92 semantic threshold is deliberately high. The 0.03 ambiguity gap flags uncertain matches for human review. You can always merge entities manually; you can never safely un-merge them.

Why task-specific memory weights instead of one ranking?

A debug task and a compliance audit need fundamentally different memory. Debug needs episodic recall (what happened?) and procedures (what patterns do I follow?). Compliance needs episodic evidence and semantic facts, but should never surface unverified procedures. One ranking can't serve both.

Why two output formats, not a universal one?

Claude handles structured XML with metadata attributes well. GPT prefers compact JSON. Sending XML to GPT wastes tokens on tags it doesn't need. Sending flat JSON to Claude loses the metadata Claude can reason about. Two formats, chosen by model parameter.