The Investigator -- Spreading Activation at Scale
1,200 city case files with 5 hidden investigation chains. Each chain has 4 cases in different domains (building inspection, health, shipping, environmental) connected by a shared location or entity. Flat retrieval finds only the direct query match. Entity-linked spreading activation discovers the full chain.
What It Demonstrates
- Spreading activation with dual-path expansion (FAISS + entity links)
- Entity linking via spaCy NER
- Cross-domain association that pure embedding similarity cannot capture
- Scale -- 1,200 memories in the FAISS index
Results
| Chain | Flat Retrieval | With Spreading | Improvement |
|---|---|---|---|
| Thornfield (industrial contamination) | 1/4 | 4/4 | +3 |
| Ravenswood (financial fraud) | 2/4 | 4/4 | +2 |
| Westbrook (prescription ring) | 1/4 | 4/4 | +3 |
| Harborview (construction corruption) | 1/4 | 3/4 | +2 |
| Greenfield (data theft) | 3/4 | 4/4 | +1 |
| Average | 1.6/4 | 3.8/4 | +2.2 |
Entity-linked spreading activation found 2.4x more connections than flat retrieval.
Key Finding
Entity-linked spreading traverses "Industrial Way" across a warehouse inspection, a hospital cluster, shipping records, and environmental readings -- four completely different city departments that no single query could connect. The LLM synthesized a coherent investigation brief from the scattered evidence.
Architecture Lesson
The original 384D embedding model (all-MiniLM-L6-v2) could not connect cross-domain cases -- "warehouse inspection" and "hospital patients" shared only 0.18 similarity. Upgrading to 768D (all-mpnet-base-v2) raised this to 0.41, and adding entity linking via spaCy NER connected all cases through the shared "Industrial Way" entity.