Crate mnem_extract 

mnem-extract

Statistical, embedding-based entity + relation extraction for mnem.

This crate is the default path for experiment E3 of the GraphRAG research track: it replaces LLM-driven NER with a KeyBERT-style candidate-scoring pass over the chunk embedding that mnem-ingest already computes. That keeps extraction deterministic, fully offline, and cost-free at ingest time.

Scope

• traits::Extractor - pluggable extractor surface. One default implementation (keybert::KeyBertExtractor) ships with the crate; callers can swap in authored or LLM-backed extractors by implementing the trait themselves.
• keybert::KeyBertExtractor - KeyBERT-style n-gram ranking against a supplied chunk embedding, with MMR (Maximal Marginal Relevance) diversification and deterministic tiebreaks.
• cooccurrence::mine_relations - PMI-weighted co-occurrence relation miner that emits one traits::Relation per sentence- local entity pair whose pointwise mutual information exceeds a configurable threshold.

Determinism

Every public extractor in this crate is deterministic: same input text + same embedder → byte-identical traits::Entity and traits::Relation streams across runs. The proptest suite under tests/proptest_determinism.rs enforces this as a first-class property.

Non-goals

• No LLM calls. No network. No tokio.
• No training, no fine-tuning: the extractor consumes whatever mnem_embed_providers::Embedder the caller already configured.
• No HTTP / MCP / CLI wiring lives in this crate; mnem-ingest exposes the integration and mnem-cli surfaces the flag.

Re-exports

pub use cooccurrence::CoOccurrenceMiner;

pub use cooccurrence::mine_relations;

pub use keybert::KeyBertExtractor;

pub use traits::Entity;

pub use traits::ExtractionSource;

pub use traits::Extractor;

pub use traits::Relation;

pub use inference::InferenceBudget;

pub use inference::InferenceMethod;

pub use inference::TypedRelation;

pub use trust::AuthorFingerprint;

pub use trust::AuthorRateLimiter;

pub use trust::Candidate;

pub use trust::PPR_AMPLIFICATION_FLOOR;

pub use trust::TrustBoundary;

Modules

cooccurrence

Co-occurrence relation miner - PMI-weighted edges between entities that share a sentence.

inference

Optional typed-relation inference (gap 03). Gated behind the typed-relations Cargo feature. Default OFF per solution.md R3. Optional typed-relation inference for mnem-extract.

keybert

KeyBERT-style statistical keyword / entity extractor.

traits

Public traits and value types for mnem-extract.

trust

Adversarial trust-boundary gate for opt-in typed-relation inference (gap 03). Gated behind the typed-relations Cargo feature. Default OFF. Trust-boundary gate for opt-in typed-relation inference.