P13 Roll-Up — What works, what’s complex, what’s tested

Area / ComponentWorksComplexityWhat we tested / Evidence (real data only)Notes & Links Embeddings (GTR-T5 768D, offline)✅Med10k vectors encoded from data/teacher_models/gtr-t5-base; L2-normed; NPZ schema validated (vectors/ids/doc_ids/concept_texts/tmd_dense/lane_indices/cpe_ids).artifacts/fw10k_vectors_768.npz; offline flags set; doc updated. NPZ Contract & Boot Gates✅LowStartup checks: required keys present, dim=768, shapes consistent, zero-vector kill switch.docs/architecture.md (NPZ schema); enforced in src/api/retrieve.py. FAISS Index (IP, 768D, 10k, ID-mapped)✅MedBuilt with IndexIDMap2, nlist=128, nprobe=16; /admin/faissreports dim=768, vectors=10k; returns stable doc_id.artifacts/fw10k_ivf_768.index, artifacts/faiss_meta.json. Retrieval API (FastAPI, lane-aware)✅MedServer up on localhost:8092; /search 200 + hydrated fields (doc_id, cpe_id, concept_text, tmd_code, lane_index). Smoke tests pass.tests/test_search_smoke.py; src/api/retrieve.py. LightRAG Embedder Adapter (offline GTR)✅LowNew adapter exposes embed_batch() + dim=768; smoke shows (N,768) float32, ~1.0 norms.src/adapters/lightrag/embedder_gtr.py, configs/lightrag.yml. Knowledge Graph Build (LightRAG)✅MedBuild runs with real chunk text; exports nodes/edges/stats; Neo4j load path wired.src/adapters/lightrag/build_graph.py; artifacts/kg/{nodes,edges,stats}.json(l). GraphRAG Query Pipeline (LightRAG)✅HighEnd-to-end runner: vector top-k → graph slice (PPR/BFS) → prompt pack → LLM call → Postgres + JSONL.src/adapters/lightrag/graphrag_runner.py; eval/graphrag_runs.jsonl. Local LLM (Llama, no cloud)✅MedProvider local_llama enforced; metrics captured (latency, bytes in/out); empty responses treated as failures.src/llm/local_llama_client.py; docs/howto/how_to_access_local_AI.md. Vec2Text (JXE / IELab, steps=1)🟨 ReadyMedOffline embedder path verified; round-trip hooks ready to run with --steps 1.how_to_use_jxe_and_ielab.md; next: run on 3 real IDs & log cosine. Eval & Reporting✅Med20-query harness; latency probe; report scaffold populated.scripts/run_graphrag_eval.sh; eval/day13_graphrag_report.md. 3D Semantic Cloud (real vectors)✅LowPCA 768→3D HTML (Plotly) generated from real embeddings; zero-vector fallback removed.tools/generate_semantic_cloud.py; artifacts/semantic_gps_cloud_visualization.html. CI/Hardening✅MedFail-fast boot, dim checks, zero-vector guards, API smoke; FAISS/NPZ parity checks.Documented in docs/architecture.md; tests in tests/*. Docs & Run Logs✅LowArchitecture, local model usage, LLM policy; run log updated.eval/day13_graphrag_report.md; docs/run_log.md.

quick legend

✅ = implemented & exercised on real data

🟨 Ready = path implemented; final evaluation pending

Complexity: an estimate of ongoing maintenance/operational effort (Low/Med/High)