Thesis: Vector RAG confuses name similarity with fact. A knowledge graph enforces entity identity and tenure; vector search only supports the pack; the language model may not invent employers or dates.
1. Problem
Executive track records live in unstructured proxy biographies and 8-K appointment notices. Nearest-neighbor retrieval attaches the wrong employer or wrong decade because embeddings optimize semantic proximity, not temporal or entity structure.
2. Pipeline
1. Real EDGAR corpus (fixed 10 large-caps) · 2. Hand gold graph (25 executives) · 3. Local extraction vs gold · 4. Neo4j + Qdrant · 5. Hybrid retrieval (1–2 hop + ticker filter) · 6. Citation-bound memo + post-hoc verifier.
| Metric | Bar | Measured |
|---|---|---|
| Entity resolution (hybrid) | ≥90% | 100% (8 fixtures) |
| Temporal accuracy (hybrid) | ≥85% | 100% |
| Hallucination-free (hybrid) | 100% | 100% (vector-only 62.5%) |
| Memo unlinked claims | 0 | 0.0 rate |
