Skip to main content
Knowledge-graph pipelines (neo4j-graphrag, LlamaIndex, Graphiti, and friends) ingest text, extract entities and relationships, resolve/dedupe the entities, and write a graph. goldenmatch is the entity-resolution layer, not a KG builder, so the goldenmatch-kg package drops it in as that resolve stage wherever a framework exposes a seam.
This is an ER-stage swap, not “goldenmatch builds your graph”. You still run the rest of your pipeline (extraction, relationship building, graph store). goldenmatch resolves the entities, and ER-KG-Bench measures how much better than each framework’s built-in default that is.

​
Install

​
neo4j-graphrag

A real GoldenMatchResolver that subclasses the library’s BasePropertySimilarityResolver and overrides the resolution step, so goldenmatch decides which nodes merge and Neo4j persists them. It replaces FuzzyMatchResolver as a true in-pipeline plugin.
On ER-KG-Bench’s ghsuite corpus, neo4j-graphrag’s own fuzzy resolver scores F1 0.322; goldenmatch zero-config scores F1 0.969 (+64.7pp), at zero LLM calls.

​
LlamaIndex PropertyGraphIndex

LlamaIndex ships no fuzzy entity resolver (its default is exact name+label upsert), so goldenmatch is additive here: a GoldenMatchEntityResolver transform canonicalizes entity names before upsert, so variant mentions collapse into one node.
LlamaIndex’s exact-upsert default behaves like the exact-match family (F1 0.0 to 0.24 on ghsuite); goldenmatch scores F1 0.969.

​
Graphiti

Graphiti exposes no public resolver seam, so goldenmatch runs as a post-ingestion maintenance pass: propose_entity_merges re-resolves the graph’s existing entity nodes and returns the duplicate groups its floor + LLM missed.
Graphiti’s real deterministic dedup floor scores F1 0.379 on ghsuite; goldenmatch scores F1 0.969 (+59pp), at zero LLM calls (Graphiti’s full path escalates unresolved nodes to an LLM).

​
How the lift stays honest

The per-framework numbers above are read straight off the ER-KG-Bench board (each framework’s default ER row versus the goldenmatch row), not re-scored here. Each adapter is proven by a parity test against real dedupe_df plus an integration test that runs the framework’s real code with the goldenmatch shim wired in. GraphRAG, mem0, LightRAG, and Cognee have no clean resolver seam and are out of scope for v1.