Phase 6 — NIM / Earth-2 Inference

Duration: ~7 working days Critical-path predecessor: Phase 3 Status: in progress

Deliverables (§5.5 / §5.7 / §11)

1. Production FourCastNet + CorrDiff client with Redis cache and R2 persistence
2. Vision NIM — SAM-2 (segment_glof_inundation) and OWL-ViT (detect_avalanche_debris)
3. Scientific RAG corpus — pgvector store of IPCC / ICIMOD / NDMA / forensic
4. Multilingual narrator — Llama-3.1 70B via NIM, EN / HI / UR / DOI / PAH
5. CAP 1.2 bulletin generator — OASIS-compliant XML for NDMA / SDMA feeds
6. Hardened Netlify edge proxy — origin allow-list + per-IP rate limit
7. Frontend NarratorPanel integrated into the Provenance section

Acceptance Gate

Sub-2 s FourCastNet 6-h forecast latency; multilingual SMS auto-generation verified against IMD bulletin format.

Run:

cd snow-ir-backend
pytest tests/nim -v
uvicorn snow_ir.api.main:app --reload

# Cache-cold
time curl -s -X POST http://localhost:8000/nim/forecast/fcn \
  -H "Content-Type: application/json" \
  -d '{"lead_hours": 6}' | jq

# Cache-hot (should return < 50 ms server-side)
time curl -s -X POST http://localhost:8000/nim/forecast/fcn \
  -H "Content-Type: application/json" \
  -d '{"lead_hours": 6}' | jq

# Multilingual narrator
curl -s -X POST http://localhost:8000/nim/narrate \
  -H "Content-Type: application/json" \
  -d '{"language":"hi","assessment":{"hazard_class":"GLOF","sub_basin_id":"kishtwar","probability":0.62,"ci_low":0.55,"ci_high":0.69,"forecast_horizon_h":72,"severity":"warning"}}' | jq

Architecture

Layer	Module	Responsibility
Forecast	nim/forecast/earth2_production.py	FCN + CorrDiff with cache, retries, synthetic fallback
Vision	nim/vision/segmentation.py	SAM-2 + OWL-ViT prompted segmentation
RAG	nim/rag/corpus.py	Chunking, embedding, pgvector search
Narrator	nim/narrator/llm_narrator.py	Llama-3.1 multilingual bulletin generation
CAP	nim/cap/bulletin.py	OASIS CAP-1.2 XML serialiser
Edge	netlify/edge-functions/nim-proxy.ts	Server-side key handling + rate limit
API	api/nim_routes.py	FastAPI router for all NIM operations
UI	components/narrator/NarratorPanel.tsx	Per-alert multilingual UI

Discipline

• API key security: API key never reaches the browser. All /api/nim/* calls are intercepted at the edge and signed server-side.
• Deterministic fallback: Synthetic fallback is deterministic — same input → same payload hash → same bulletin text. Non-credentialed staging environments still produce every artefact.
• Length budgets: Bulletin length budgets are enforced in the system prompt and clipped at the response boundary in narrator.py.
• Validation: CAP XML is always validated against the schema in tests/nim/test_cap.py.
• Cache isolation: Earth-2 cache keys are bbox-pinned to the Chenab basin geometry, so a neighbour's identical issued_at request would not collide.