Graph RAG for Multi-Hop Questions
- Python
- Pydantic
- Groq
- NumPy
- Question
- When the facts for an answer sit in different news articles, does walking a knowledge graph find it where similarity search does not?
- Setup
- Relation extraction over 110 Indian insurance articles, with deal status and ownership periods; 222 names merged into 136 companies and people; cited answers.
- Result
- Graph RAG 10/11, vector RAG 9/11, the model alone 3/11.
- 115 articles
- Extract
- Resolve · 222 → 136
- Graph · 166 edges
- 2-hop walk
- Cited answer
11 questions, one run each, over 110 of 115 articles (5 didn’t extract). A direction, not a proof.
4/4 vs 3/4cross-article questions · graph vs vector
One question apart. Read it as a direction, not a proof.
- Every relation has a status (asserted, prospective, denied), so a pending deal is never reported as done.
- No fuzzy entity matching. Every alias was checked by reading the source articles.
- Round-robin evidence, after one deal filled 14 of 30 context slots.
- A fair baseline: vector RAG gets the same prompt, model and citation format; only retrieval differs.