FAISS โ Spot Vector Index
| Status (this host) | NOT INSTALLED on this Windows host |
| GPUs detected | 0 |
| Planned host | A6000-1 (192.168.1.155) or A6000-2 (192.168.1.240) โ both have NVIDIA A6000 48 GB |
| Planned index path | /media/melanoscan/SND_1_Dot_55_8TB/melanoscan/faiss/spots.index |
| Planned vectors | ~17.3 M Spot feature vectors derived from segmented_spots + classed_clinical: [meanR, meanG, meanB, area, perim, MEL, BCC, NV, BKL, DF, AKIEC, VASC] โ 12-dim, normalized |
| Reference script | faiss_neo4j_pipeline.py (Drive csc/) โ currently a doc-embedding template, needs adaptation for spot vectors |
When the index is built, this page will provide:
- Free-form vector query (find spots most similar to a given spot's feature vector)
- "Find look-alike spots" โ given a Neo4j Spot or MySQL primary_key, return top-K nearest neighbors
- Constellation matching (geometric fingerprint similarity across patients)
- Cross-MRN error detection (look-alike spots in different patients = potential MRN swap)
Until built, use Neo4j Constellation similarity (already populated, 137,530 nodes) as the closest equivalent. See (:Constellation)-[:EVOLVED_TO]->(:Constellation) for cross-session geometric matches.