GraphRAG Internals: Structuring Neo4j Knowledge Graphs

Ključne točke

1. Extract entities and relationships with a schema. Unconstrained extraction produces a noisy graph that retrieves worse than vectors.
2. Deduplicate aggressively. The same person or concept appears under many surface forms; merge them before community detection.
3. Detect communities with Leiden. Communities become the unit of summarization and the boundary for multi-hop questions.
4. Summarize each community. A short summary per community gives the model a high-level map before it drills into entities.
5. Store provenance. Every entity and edge must link back to source chunks so answers can be cited and audited.
6. Query with both paths. Combine graph traversal for relationship questions and vector search for topical questions.
7. Rebuild incrementally. Full rebuilds are expensive; track document versions and update only affected subgraphs.
8. Evaluate against a held-out question set. GraphRAG gains are easy to overstate; measure answer accuracy, not just retrieval recall.

Architecture

The pipeline has four stages. Extraction turns chunks into entities and edges. Deduplication merges surface forms. Community detection groups related entities. Summarization produces a navigable map. At query time, the system selects communities, expands to entities, and retrieves source chunks for the final answer.

Production configuration

CREATE CONSTRAINT entity_name IF NOT EXISTS FOR (e:Entity) REQUIRE e.name IS UNIQUE;
CREATE INDEX entity_type IF NOT EXISTS FOR (e:Entity) ON (e.type);

// Link an entity to its source chunk for provenance
MATCH (c:Chunk {id: $chunk_id})
MERGE (e:Entity {name: $name, type: $type})
MERGE (e)-[:MENTIONED_IN]->(c)

from graphdatascience import GraphDataScience
gds = GraphDataScience(driver)
G, _ = gds.graph.project('entities', 'Entity', 'RELATED_TO')
gds.leiden.write(G, writeProperty='community')

for cid in communities:
    members = fetch_entities(cid)
    summary = llm.summarize(members, max_tokens=200)
    db.run('MERGE (c:Community {id: $id}) SET c.summary = $s', id=cid, s=summary)

def answer(question):
    communities = vector_search(question, index='community_summaries', k=3)
    entities = expand_communities(communities, depth=2)
    chunks = fetch_chunks(entities) + vector_search(question, index='chunks', k=5)
    chunks = dedupe(chunks)
    return llm.answer(question, context=chunks, communities=communities)

Edge cases from production

• Extraction hallucinates relationships; validate edges against source text before writing them.
• Community summaries go stale when the corpus changes; version them and rebuild on document updates.
• Multi-hop questions exceed context limits; cap expansion depth and summarize intermediate nodes.
• Entity resolution fails on abbreviations and acronyms; maintain an alias table seeded from the corpus.

Često postavljana pitanja