EPIC. MemGraphRAG is demonstrated as a transferable graph constructor that improves four downstream GraphRAG frameworks. Fathom's analog is a curated, versioned rule pack that acts as the deterministic governance stage in a GraphRAG construction pipeline: an external extractor emits candidate triples; fathom denoises (schema-frequency), detects conflicts, and emits allow/deny/escalate per triple before it enters the graph — replacing non-deterministic LLM filtering with auditable rules. Zero core changes.
Composition (sub-work).
Where.
- New
src/fathom/rule_packs/graphrag_consistency/{templates,modules,rules} (follow hipaa/ pack layout).
- Register in
pyproject.toml [project.entry-points."fathom.packs"] :144-149; discovered via packs.py RulePackLoader.discover :18 / load :44; mounted via Engine.load_pack (engine.py:856).
- Module ordering via the focus stack (
evaluator.py _setup_focus_stack :79).
Templates. candidate_triple(head, relation, tail, source_passage, support_count), stable_schema(relation_type), mutual_exclusion(relation, tail_a, tail_b), conflict(...).
Acceptance criteria.
Notes. Depends on the two linked features landing first. τ and mutual-exclusion lists are domain-specific. Relates to #73.
Source: arXiv:2606.00610v1 — "MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation". Distilled from arXiv-research/2606.00610v1/analysis.md; file refs verified against current main by the analysis pass.
EPIC. MemGraphRAG is demonstrated as a transferable graph constructor that improves four downstream GraphRAG frameworks. Fathom's analog is a curated, versioned rule pack that acts as the deterministic governance stage in a GraphRAG construction pipeline: an external extractor emits candidate triples; fathom denoises (schema-frequency), detects conflicts, and emits allow/deny/escalate per triple before it enters the graph — replacing non-deterministic LLM filtering with auditable rules. Zero core changes.
Composition (sub-work).
graphrag_consistencypack with ordered denoise→detect→(escalate) modules + external-extractor wiring over MCP.Where.
src/fathom/rule_packs/graphrag_consistency/{templates,modules,rules}(followhipaa/pack layout).pyproject.toml [project.entry-points."fathom.packs"]:144-149; discovered viapacks.pyRulePackLoader.discover:18 /load:44; mounted viaEngine.load_pack(engine.py:856).evaluator.py_setup_focus_stack:79).Templates.
candidate_triple(head, relation, tail, source_passage, support_count),stable_schema(relation_type),mutual_exclusion(relation, tail_a, tail_b),conflict(...).Acceptance criteria.
Notes. Depends on the two linked features landing first. τ and mutual-exclusion lists are domain-specific. Relates to #73.
Source: arXiv:2606.00610v1 — "MemGraphRAG: Memory-based Multi-Agent System for Graph Retrieval-Augmented Generation". Distilled from
arXiv-research/2606.00610v1/analysis.md; file refs verified against currentmainby the analysis pass.