5…d6
+2.32, the engine’s eval after 5…d6. The seam rests there.
Result and evidence
- +35% retrieval vs vector-onlyControlled evaluationIndependent handbook and policy archive, not the Monash corpus.
The case
- Problem
- A question about prerequisites has to walk the graph, not only the nearest paragraph.
- Decision
- The graph is the structured path for prerequisites and policy; the vector store is the fallback when no valid graph path is returned.
- Constraint
- Prerequisites and credit-transfer are edges, not another vector-only retrieval dump.
- Example
- Input: a policy question. Output: a neighbourhood of clauses from the independent handbook archive.
- Considered and rejected
- Vector-only retrieval as the primary path.
Architecture
- LangGraphorchestrates
Beside the path
- Neo4jstructured
- Vector DBunstructured
What was built
- Agents write Cypher and check it.
- Global-to-local retrieval for questions that span the whole handbook.
- Ambiguous questions fall back to a broader semantic search.
- Limitations
- The query set, scoring rule and denominator were not recorded; this is not Recall@k.
- What I would change now
- Record them.
LangGraph · Neo4j · Vector DB · Knowledge Graphs · RAG
- The source is in a private repository.
- Back to all work