| --- |
| title: KCC AgriAdvisor |
| emoji: 🌾 |
| colorFrom: green |
| colorTo: yellow |
| sdk: docker |
| pinned: false |
| app_port: 7860 |
| --- |
| |
| # 🌾 KCC AgriAdvisor |
|
|
| AI-powered agricultural advisory for Indian farmers. The most comprehensive |
| open Indian agriculture AI stack — RAG over 16.5M Kisan Call Center records, |
| national pest forecast for 700 districts × 20 crops, block-level satellite |
| features, ICAR-cited dose recommendations. |
|
|
| ## 📊 Performance |
|
|
| Independent benchmark — **100-query stratified sample** from |
| `IndiaAgriBench-498`, Indian farmer questions across pest, disease, nutrient, |
| irrigation, scheme, price, and crop-selection problem types. |
|
|
| | Metric | This Stack | Vanilla Gemini | Vanilla Groq | |
| |---|---|---|---| |
| | Citation rate | **0.91** | 0.00 | 0.00 | |
| | Banned-chemical leakage | **0.00** | 0.00 | 0.04 | |
| | Latency p50 (ms) | **787** | 9,978 | 1,015 | |
| | Judge — Citation (1-10) | **9.12** | 1.14 | 0.26 | |
| | Judge — Safety (1-10) | **8.48** | 8.67 | 5.50 | |
|
|
| Smoke re-run on the v4-merged stack (5-query sample, 2026-05-11) confirmed |
| citation rate 1.0, banned leak 0.0, latency p50 627 ms — at or above |
| baseline. Full benchmark: see the in-app **/proof** tab or |
| `eval/benchmark_full_2026-05-10.json` in the IP bundle (NDA-gated). |
|
|
| ## Features |
|
|
| - **RAG chatbot** over 16.5M KCC Q&A — Hindi / English / 8 regional languages |
| - **Pest Early Warning** — district stacking model (AUC 0.937), 1-month lead time |
| - **Price Forecast** — presow_v4 P25/P50/P75, 290 crops |
| - **National Pest Heatmap** — 700 districts × 20 crops × 3 months, interactive |
| - **Citation Guard** — post-generation safety filter, negation-aware |
| - **Enterprise Dashboard** — B2B pest + price analytics |
| |
| ## Architecture (v4) |
| |
| ``` |
| client → /query |
| ↓ |
| topic guard → classify → v2 multi_step_retrieve (FAISS+BM25+rerank) |
| ↓ |
| prompt build (system + safety + few-shot + retrieved context) |
| ↓ |
| kcc_core.llm cascade: Groq → Gemini → local Llama (first success wins) |
| ↓ |
| kcc_core.citation_guard.review (negation-aware banned-chem strikethrough, |
| hard overrides, [1][2] citation check) |
| ↓ |
| JSON { answer, sources, safety_warnings, backend, ... } |
| ``` |
| |
| ## HF Spaces Secrets required |
|
|
| | Secret | Purpose | |
| |---|---| |
| | `GROQ_API_KEY` | Primary LLM (Groq llama-4-scout, fast) | |
| | `GEMINI_API_KEY` | Fallback LLM + image diagnosis | |
| | `SECRET_KEY` | JWT signing (≥32 hex chars) | |
| | `B2B_DEMO_PASSWORD` | Demo enterprise login | |
| | `HF_TOKEN` | (Optional) Pull large data files from HF Dataset at boot | |
|
|
| ## What's in this v4 merge (vs v2) |
|
|
| | Area | v2 (original) | v4 (this) | |
| |---|---|---| |
| | LLM cascade | inline in api.py | `kcc_core.llm` (one canonical path) | |
| | Backend tracking | always "groq" lie | actual model that answered | |
| | Banned chemicals | 26-item list, substring | 33-item, negation-aware | |
| | Citation guard | none post-gen | runs on every /query + /query/stream | |
| | Hard overrides | none | BPH ≠ Imidacloprid | |
| | Rate limiting | none | slowapi (120/min, 2000/hr default) | |
| | Request tracing | none | X-Request-ID middleware | |
| | `/pest-risk` state-only | open | requires B2B JWT | |
| | Frontend | Farmer + B2B tabs | + new **/proof** tab with benchmark charts + India pest heatmap | |
| | Tests | none | 19/19 passing | |
|
|
| Full IP bundle (LoRA v2, BGE-M3 index, 498-query eval, ICAR JSON) is |
| NOT in this Space — kept local for NDA-gated buyer sharing. |
|
|
