KrishokTech Advisory System

A safety-aware, retrieval-grounded, Bengali agricultural advisory system for smallholder farmers in Bangladesh. The system answers farming questions in Bengali through a four-stage agent pipeline (safety screening, retrieval, grounded generation, verification) and diagnoses crop diseases from photos using a crop-classifier + per-crop vision model workflow. It is a research capstone demo prototype, not a production service.

🌐 Interactive Demo Preview: https://krishoktech.vercel.app
🎥 Demonstration Screencast: https://krishoktech.vercel.app/screencast
▶️ Screencast on YouTube (2:24): https://youtu.be/a4cibXlvGdQ

🚀 EACL 2027 Reviewer Quick Start (5-Minute Reproduction Guide)

KrishokTech provides an independently runnable, 100% offline (zero API keys required, CPU-only) replication suite for all quantitative tables, safety gates, and empirical claims reported in the EACL 2027 paper.

1. One-Command Paper Results Reproduction

git clone https://github.com/RaiyaanReza/KrishokChat-Agricultural-Advisory-System.git
cd KrishokChat-Agricultural-Advisory-System

# Run the turnkey offline paper verification runner:
python reproduce.py
# (or: bash reproduce.sh / reproduce.bat)

Expected terminal output (< 500 ms total runtime):

======================================================================
  KRISHOKTECH — PAPER RESULTS REPRODUCIBILITY BENCHMARK (EACL 2027)  
======================================================================
  Release / Git Commit : v1.0.0-eacl2027
  Execution Mode       : 100% Offline (0 API Keys, Local CPU)
  Deterministic Seed   : 42
----------------------------------------------------------------------
Benchmark / Paper Claim                          Result           Status
------------------------------------------------------------------------
T0 Deterministic Precheck (Banglish Red-Teaming) 85/85 adv, 15/15 ben [PASS]
T1 Pre-Retrieval Crop Gate (Halt-Before-Retrieval) 30/30 halted     [PASS]
C2 Photo Fence & Contradiction Badge Gate        453/454 caught   [PASS]
T4 Hardened Dosage Verifier Entailment           114/118 caught   [PASS]
C2 Crop-Fence Retrieval Isolation & Simulation   0/400 violations [PASS]
------------------------------------------------------------------------
✓ ALL PRINCIPAL PAPER METRICS SUCCESSFULLY REPRODUCED! (5/5 PASS)

A machine-readable execution audit is automatically generated at reproduce_results.json.

---

2. Artifact & Reproducibility Matrix

Artifact / Component	Reported Paper Claim	Offline Reproducible	Test / Execution Command	Status
T0 Safety Precheck	85/85 Banglish/phonetic attacks intercepted, 0/15 false alarms	✓	python reproduce.py	PASS
T1 Crop Gate (ASK)	30/30 crop-less symptom requests halted before retrieval	✓	python reproduce.py	PASS
C2 Contradiction Badge (CONFIRM)	453/454 text-image conflicts detected (53/54 farmer, 400/400 PRISM)	✓	python reproduce.py	PASS
T4 Hardened Verifier (DROP)	114/118 dosage mutations caught, 0/38 clean false alarms	✓	python reproduce.py	PASS
C2 Retrieval Crop-Fencing	Wrong-crop advice drops 36.25% → 30.00% ($p=5.96\times 10^{-8}$), 0/400 violations	✓	python reproduce.py	PASS
Core Safety Unit Tests	74 regression tests across gates, boundaries, registries, & redaction	✓	pytest backend/tests/safety_gates	PASS
Five Demo Scenarios (S1–S5)	Interactive verification of ASK, BIND, CONFIRM, REFER, DROP	✓	python backend/tests/run_e2e.py	PASS

---

3. Five Canonical Demonstration Scenarios (S1–S5)

The paper structures KrishokTech's user-facing authorization boundaries across five observable states:

• S1: ASK (Crop Disambiguation Intercept) — Farmer asks পাতায় হলুদ দাগ হয়েছে, কী বিষ দিব? without naming a crop. The system halts retrieval immediately, saving 88% token context, and presents quick-reply chips: [ ধান ] [ আলু ] [ টমেটো ].
• S2: BIND (Photo-Bounded Retrieval) — Uploading a potato leaf photo binds retrieval to potato evidence only, preventing cross-crop chemical contamination.
• S3: CONFIRM (Contradiction Badge) — Farmer uploads a potato leaf photo but types a message about brinjal (বেগুনের পাতা). The system intercepts with a [সন্দেহজনক বৈপরীত্য] confirmation badge before recommending any chemical.
• S4: REFER (16123 Helpline Referral) — Farmer asks for a banned chemical (e.g. প্যারাকোয়াট or phonetic Banglish parakwat). The T0 precheck halts generation in 0.32 ms and surfaces the national agricultural emergency hotline (16123).
• S5: DROP (Dosage-Claim Sanitization) — If the answer generator outputs an ungrounded or mutated dosage rate, the narrow relational verifier strips the sentence while preserving safe cultural practices.

---

Table of Contents

• EACL 2027 Reviewer Quick Start
• Overview
• Features
• Architecture
• Tech Stack
• Repository Layout
• Getting Started
• Configuration
• API Reference
• Data & Models
• Frontend Routes
• Testing & Verification
• Known Limitations
• Documentation
• Status & License

---

Overview

KrishokTech combines three already-trained research artifacts into one demo system:

1. Bengali agri Q&A assistant — hybrid retrieval (BM25 sparse + FAISS dense) over a precomputed corpus of 2,120 knowledge nodes, feeding a fine-tuned LLM with grounded, source-cited answers.
2. Crop disease advisory workflow — a crop classifier routes an uploaded photo to a crop-specific disease model, which returns a Bengali diagnosis and treatment advice.
3. Safety-aware agentic pipeline — every query is classified before retrieval. Unsafe queries are stopped with a canned redirect to the government Krishi Call Center (16123); every classification decision is written to a local audit log.

A research/benchmark panel in the frontend displays precomputed retrieval-evaluation stats from the author's prior work. No numbers are computed live, and no claim is made that the system itself is evaluated beyond the verified checks in backend/ml_assets/vision/verification_report_live.md.

Features

• Hybrid RAG in Bengali — BM25Okapi (k1=2.2, b=0.4) + FAISS IndexFlatIP over mE5-small (384-dim) dense embeddings, with per-query top-k fusion.
• Four-stage agent pipeline — Safety/Router -> Retrieval -> Generation -> Verifier. Each stage emits a trace event consumed by the frontend stepper UI.
• Six-way safety classification — safe_agri, banned_or_restricted_chemical, self_harm_or_poisoning_risk, off_topic, prompt_injection, low_confidence. Terminal categories stop the pipeline before any retrieval happens.
• Local audit trail — every query, category, action, and timestamp appended to backend/app/logs/safety_audit.jsonl; surfaced via GET /api/safety/metrics.
• Streaming chat — SSE transport (POST /api/qa/stream) emitting stage events, token chunks, and a final typed response.
• Photo-based diagnosis — POST /api/classify (crop) and POST /api/detect (crop -> disease -> treatment advice, with confidence and quality warnings). Current artifacts are classification models; the API reports detection_mode: "classification" and never claims bounding boxes.
• Weather + helpline extras — token-efficient Bengali weather summary with a one-line agri tip, and a local-only helpline registration endpoint (no external telemetry).
• Replaceable LLM adapters — the pipeline depends on a small LLMClient port, not a provider SDK. Switching between OpenRouter, Gemini, Ollama, or a stub is a configuration change (LLM_PROVIDER), not a code change.

Architecture

The backend is organized as a layered composition root, not a script of module calls:

HTTP routers (app/api)                    -- transport only, no business logic
        |
        v
application use cases (app/application)   -- QA pipeline, vision pipeline
        |
        v
ports / protocols (app/ports)             -- LLM, retriever, verifier, vision,
                                             audit, session contracts
        |
        v
infrastructure adapters (app/infrastructure)
      -- OpenRouter/Gemini/Ollama client, BM25/FAISS retriever,
         Ultralytics inference, JSONL audit sink, in-memory sessions

Rules enforced by the refactor contract (docs/refactor/ARCHITECTURE.md):

• The application layer never imports google.genai, httpx, ultralytics, pickle, or filesystem paths directly. Those belong in infrastructure.
• The legacy app/agents/ and app/services/advisory/ directories are compatibility shims for offline scripts. No new behavior may be added there.
• One FastAPI process, one Next.js process. No auth, queues, microservices, or live index building.

QA pipeline

QARequest
  -> SafetyClassifier.classify          (deterministic pre-check, then LLM; parse
                                         failure => low_confidence, stop)
  -> QueryBuilder.build                 (only for safe_agri)
  -> Retriever.retrieve                 (BM25 + dense, top-k)
  -> GenerationModel.generate           (streamed or one-shot)
  -> Verifier.verify                    (grounding check; flags ungrounded claims,
                                         esp. chemical dosages)
  -> SessionStore.append / AuditSink.record (exactly once)
  -> QAResponse | SSE event stream

off_topic, prompt_injection, banned_or_restricted_chemical, self_harm_or_poisoning_risk, and low_confidence are terminal outcomes. They receive a canned response (for harmful categories: a calm redirect to the Krishi Call Center at 16123, plus a note to seek in-person medical help for self-harm/poisoning framings) and never reach retrieval or generation.

Vision pipeline

uploaded image
  -> crop classifier (6 families)
  -> route to crop-specific disease model (rice 8 / corn 4 / potato 3 /
     brassica 11 / wheat 11 classes)
  -> disease + Bengali info + treatment advice (grounded in the disease
     knowledge map, `backend/ml_assets/advisory/disease_knowledge_map.json`)
  -> verifier flags + agent trace returned to the UI

All checked-in .pt weights are verified task: classify. The ONNX export path exists for future detection artifacts but nothing in the UI or API claims bounding boxes today.

Tech Stack

Layer	Choice
Frontend	Next.js 16 (App Router), React 19, TypeScript 5.9
Styling	Tailwind CSS 4 + shadcn/ui
Animation	Motion 13 (formerly Framer Motion)
Chat UX	Vercel AI SDK 7 (streaming)
Backend	FastAPI 0.141 (Python 3.11-3.13, target 3.13.15 — python.org 2026-08-05), uvicorn 0.52.3 (uvicorn.dev 2026-08-13), single service
LLM serving	Ollama (local fine-tuned Gemma 4-bit); OpenRouter/Gemini adapters available
Vision	Ultralytics YOLO (classify), ONNX Runtime
Retrieval	rank-bm25 + FAISS (CPU), loaded in-process from disk
Package mgmt	pnpm (frontend), uv (backend)
Lang tooling	bnunicodenormalizer (Bengali Unicode normalization)

Repository Layout

├── AGENTS.md                  Agent instructions, hard rules, locked stack
├── .env.example               All backend/frontend env vars, documented
├── backend/
│   ├── app/
│   │   ├── api/               HTTP transport: qa, vision, benchmark, extras
│   │   ├── application/       Use cases: qa_pipeline, vision_pipeline, container
│   │   ├── domain/            Enums, contracts, safety policy
│   │   ├── ports/             Provider protocols (LLM, retriever, verifier, ...)
│   │   ├── infrastructure/    Adapters: llm, retrieval, vision, audit, sessions
│   │   ├── agents/            Compatibility shims (legacy, do not extend)
│   │   ├── services/          Compatibility shims (legacy, do not extend)
│   │   ├── logs/              Audit trail output (gitignored)
│   │   └── main.py            App factory + composition root
│   ├── ml_assets/
│   │   ├── rag_index/         Corpus, indexes, provenance, eval results
│   │   ├── vision/            Crop classifier + per-crop disease weights
│   │   ├── advisory/          Disease knowledge map (Bengali)
│   │   └── gemma/             GGUF target for local inference
│   ├── scripts/               Offline build/eval/test scripts
│   └── tests/                 test_api, test_pipeline, test_vision
├── frontend/
│   └── src/
│       ├── app/               (app) chat/detect/analytics, (marketing) pages
│       ├── components/        chat, detect, layout, ui
│       └── lib/               api client, Bengali helpers, constants
├── demo-assets/               Demo images for the live investor demo
├── docs/                      Plans, architecture, vision-pipeline docs
└── scripts/                   Offline build/eval/test scripts (repo root)

Getting Started

Prerequisites

Verified on Windows 11 (see SETUP_REPORT.md):

Tool	Version
Node.js	v24 LTS Krypton (24.19.0 LTS 2026-08-03 — nodejs.org, local 24.11.0)
pnpm	11.x
Python	3.13 (maintained 3.13.15 — python.org 2026-08-05; compatible 3.11-3.12)
uv	0.11+
Ollama	optional (only if using the ollama LLM provider)

Backend

cd backend
uv sync
uv run uvicorn app.main:app --reload

Health check: http://localhost:8000/health returns {"status": "ok", ...}. The FastAPI docs (Swagger) are served at http://localhost:8000/docs.

Frontend

cd frontend
pnpm install
pnpm dev

Open http://localhost:3000. In dev, next.config.ts rewrites /api/* and /health to the backend, so the frontend and backend share an origin.

First run notes

1. Copy .env.example to .env in the repo root and set LLM_API_KEY (or install Ollama and switch LLM_PROVIDER=ollama with OLLAMA_MODEL_NAME pointing at your local tag).
2. Without a key, run with LLM_PROVIDER=stub: intent classification fails closed (safe behavior for tests) and generation returns stub output. Useful for API contract work, not for a real answer.
3. RAG indexes are precomputed and loaded from backend/ml_assets/rag_index/ at startup. Nothing is built at request time.

Configuration

Settings are read from the repo-root .env, then overridden by backend/.env.local (local secrets win). See .env.example for the full list.

Variable	Default	Purpose
LLM_PROVIDER	openrouter	Adapter: openrouter, gemini, ollama, stub
LLM_MODEL_NAME	krishoktech-4b	Model for both stages unless overridden
LLM_BASE_URL	(empty)	Endpoint override for HTTP-compatible servers
LLM_API_KEY	(empty)	Provider key (prefer local secret file)
LLM_TIMEOUT_SECONDS	30	Request bound for remote adapters
LLM_TEMPERATURE	0.2	Generation temperature
LLM_MAX_OUTPUT_TOKENS	1000	Max tokens per generation call
INTENT_MODEL_NAME / GENERATION_MODEL_NAME	(empty)	Role-specific model overrides
OLLAMA_BASE_URL	http://localhost:11434	Ollama server URL
OLLAMA_MODEL_NAME	krishoktech-4b	Ollama model tag
OPENROUTER_MODEL	google/gemini-2.5-flash-lite	OpenRouter model
GEMINI_MODEL	gemini-2.5-flash-lite	Gemini model
BACKEND_HOST / BACKEND_PORT	0.0.0.0 / 8000	Uvicorn bind address
FRONTEND_ORIGIN	http://localhost:3000,...	CORS allow-list (comma-separated)
RETRIEVAL_TOP_K	5	Passages fused per query
RAG_BACKEND	FAISS	Retrieval store (Chroma supported by contract)
AUDIT_LOG_PATH	backend/app/logs/safety_audit.jsonl	Audit trail location
SESSION_MAX_TURNS	10	History length cap
SESSION_TTL_SECONDS	1800	Session expiry
VISION_CROP_CONFIDENCE_THRESHOLD	0.60	Minimum crop-class confidence
VISION_DISEASE_CONFIDENCE_THRESHOLD	0.55	Minimum disease-class confidence
VISION_MAX_IMAGE_BYTES	10000000	Upload size limit (10 MB)
DEMO_MODE	true	Enables the demo answer cache (exact-replay lane)
DEMO_CACHE_PATH	demo-assets/cached_responses.json	Cached demo responses (verified pipeline outputs only)
DEMO_CACHE_MAX_ENTRIES	100	Max entries kept in the demo answer cache
QUERY_REWRITE_ENABLED	true	Follow-up queries rewritten into standalone retrieval queries (history-aware)
NEXT_PUBLIC_BACKEND_URL	http://localhost:8000	Frontend -> backend URL

The frontend additionally reads frontend/.env.local (NEXT_PUBLIC_API_BASE), though in dev the rewrite proxy makes it unnecessary.

API Reference

Base URL: http://localhost:8000. All request/response payloads are UTF-8; Bengali is preserved with ensure_ascii=False.

Method	Path	Description
GET	/health	Liveness + app version
POST	/api/qa	Full QA pipeline, single JSON response
POST	/api/qa/stream	QA pipeline as SSE stream (stages, tokens, final)
GET	/api/models	Generation model availability (gemini / krishoktech-4b)
GET	/api/safety/metrics	Audit-derived counts by category (no fabrication)
POST	/api/classify	Crop classification of an uploaded image
POST	/api/detect	Crop -> disease -> treatment advisory for an image
GET	/api/benchmark	Precomputed retrieval-benchmark stats (see Limitations)
POST	/api/weather	Bengali weather summary + agri tip for a district
GET	/api/soil/dataset	Frozen soil-moisture dataset stats (722 imgs, 6 soil types)
POST	/api/soil/analyze	Soil photo analysis — locked until model verification, returns honest trace
POST	/api/helpline/register	Local-only helpline registration (JSONL)

POST /api/qa

Request:

{
  "query": "ধান গাছের পাতা হলুদ হয়ে যাচ্ছে কেন?",
  "session_id": "demo-1",
  "crop": null,
  "disease": null,
  "history": []
}

Response shape: query, category, answer, sources[] (id, crop, disease, question, score, answer excerpt, treatment, source, expert_verified), confidence (verified / flagged-unverified / low_confidence / blocked), agent_trace[] (stage, status, detail), verifier_flags[], model.

POST /api/qa/stream (SSE)

Events are lines of the form <event>: <json>\n\n:

• data: — AgentStageEvent ({"stage": "safety", "status": "complete", ...})
• token: — {"text": "..."} partial generation chunks
• final: — the complete QAResponse object

The frontend renders the stage events as the "agent trace" stepper (Checking safety -> Retrieving sources -> Generating answer -> Verifying).

Data & Models

Local model — KrishokTech-4B (optional)

The chat UI offers two generation models: Gemini 2.5 Flash-Lite (online, default) and KrishokTech-4B (local, via Ollama). The local option is disabled in the UI until the model is actually registered in Ollama.

Activate it in one command once you have the fine-tuned GGUF:

# GGUF at backend\ml_assets\gemma\model.gguf (default location)
powershell -ExecutionPolicy Bypass -File scripts\local_model.ps1

# Or: point at a GGUF elsewhere / download it automatically
powershell -ExecutionPolicy Bypass -File scripts\local_model.ps1 -GgufPath "D:\models\krishoktech-4b-q4.gguf"
powershell -ExecutionPolicy Bypass -File scripts\local_model.ps1 -DownloadUrl "https://huggingface.co/<org>/<repo>/resolve/main/<file>.gguf"

What it does: ensures the GGUF exists, starts Ollama if needed, registers the tag krishoktech-4b from scripts/Modelfile, restarts the backend, and verifies via GET /api/models. Manual equivalent:

ollama serve   # or start the Ollama app
cd backend\ml_assets\gemma
ollama create krishoktech-4b -f ..\..\..\scripts\Modelfile

Until the tag exists, choosing the local option in the UI fails closed: the pipeline returns the 16123 referral with low_confidence instead of a plausible-but-fake answer.

RAG corpus (backend/ml_assets/rag_index/)

Item	Value
Knowledge nodes	2,120 (cleaned, normalized, Gemini-refined)
Source documents	2,946 markdown files
Source institutions	13 Bangladeshi agricultural bodies (BARC, BARI, DAE, CABI, ...)
Sparse index	BM25Okapi, k1=2.2, b=0.4
Dense index	mE5-small embeddings (384-dim), FAISS IndexFlatIP (exact)
Provenance	Node-to-QA and MD-to-QA mapping manifests under provenance/

Vision models (backend/ml_assets/vision/)

All weights verified against the actual artifact metadata (see backend/ml_assets/vision/verification_report_live.md):

Model	Task	Classes
crop_classifier	classify	6 families (Brassica, Corn, GourdGuava, Potato, Solanacea, Wheat)
rice_disease	classify	8 (BLB, Brown Spot, Healthy, Leaf Blast, Leaf Scald, Narrow Brown Spot, Rice Hispa, Sheath Blight)
corn_disease	classify	4
potato_disease	classify	3 (Early Blight, Healthy, Late Blight)
brassica_disease	classify	11 (cabbage x4, cauliflower x7)
wheat_disease	classify	11

Verified live results: wheat 11-class sweep 44/50 (88%) correct with 100% Bengali treatment info; 437-image crop-library sweep returned Bengali disease info for 436/437 images (99.8%).

Safety dataset

A 20,112-record Bengali safety evaluation dataset (refusal + requery splits, 6 dialects, 12 categories) lives in KrishokTech/dataset_release/safety/ alongside generation scripts and a design doc in docs/pipeline/scripts/.

Frontend Routes

Path	Page
/	Marketing home
/chat	Q&A assistant with agent-trace stepper
/detect	Photo upload -> crop/disease/advisory
/soil	Soil moisture field console — dataset showcase + locked analyzer
/analytics	Safety metrics panel (from audit log)
/research, /research/benchmark, /research/methodology, /research/safety	Research/benchmark panel
/about, /team, /data, /library, /contact, /auth	Marketing/support pages

Testing & Verification

Backend smoke checks (no external services required):

cd backend
uv run python -m compileall -q app
uv run python -c "from app.main import app; print(app.title, app.version)"

Integration tests live in backend/tests/ (test_api.py, test_pipeline.py, test_vision.py, plus the T0/P0 hardening suites). Run the offline scripts under backend/scripts/ for retrieval, classifier, and pipeline checks (e.g. replay_golden.py --assert-invariants, test_bm25_retrieval.py).

Frontend checks:

cd frontend
pnpm build

Production Deployment (single box)

Phase 0 hardening is documented in docs/PRODUCTION_ROLLOUT_PLAN.md (full roadmap with Phase 1–3). The ready-to-use single-box path:

# 1. One-time build + env
cd backend;  uv sync
cd ..\frontend;  pnpm build

# 2. Start both services (uvicorn with proxy-headers + graceful shutdown +
#    bounded concurrency; Next production server)
powershell -ExecutionPolicy Bypass -File scripts/start_prod.ps1

Runtime controls added in Phase 0:

Surface	What it does
GET /health	Liveness (always 200 when the process is up)
GET /readyz	Readiness per-check (bm25 index, corpus, dense index, sqlite/audit dirs). Informational by default; READINESS_STRICT=true → 503 on failure
DOCS_ENABLED=false	Hides /docs, /redoc, /openapi.json
SSE heartbeat	/api/qa/stream emits : keepalive every 15 s of silence (proxy/NAT safe)
500 envelope	Uniform {"error","request_id","detail"} — no traceback in bodies
LOCAL_LLM_MAX_CONCURRENCY	Queues local-lane generations (default 2) instead of overloading llama.cpp
CORPUS_VERSION	Bumps demo-cache keys after an index rebuild
Frontend headers	nosniff, no-referrer, X-Frame-Options: DENY, scoped Permissions-Policy

Security/reliability notes: dependency audit report in docs/production_readiness/dependency_audit_2026-08-17.md; audit-log retention policy in docs/production_readiness/retention_policy.md; sample log rotation in docs/production_readiness/logrotate.krishoktech.

Known Limitations

• GET /api/benchmark is a placeholder returning {"status": "not_implemented"}. Precomputed stats exist under backend/ml_assets/rag_index/eval/ and are wired into frontend copy, but the endpoint itself is not filled in.
• The demo answer cache (demo-assets/cached_responses.json) is populated by uv run python scripts/prewarm_demo_cache.py or self-populates on the first live ask in DEMO_MODE; until prewarmed, first-time curated questions run the live pipeline (5–15 s) exactly as before.
• Vision artifacts are classification-only. Object detection (bounding boxes) is supported by the ONNX export path but no detection weights are checked in, so nothing claims boxes.
• Ollama is not installed in the dev environment; the default adapter is OpenRouter.

Documentation

Document	Content
AGENTS.md	Hard rules, locked tech stack, demo constraints
docs/high_level_plan.md	7-day roadmap, latency engineering, demo tactics
docs/00_preprocessing_plan.md	Corpus preprocessing
docs/01_chatbot_rag_plan.md	RAG + chatbot pipeline
docs/03_safety_aware_agentic_pipeline.md	Four-stage agent pipeline
docs/04_router.md	Crop classifier + YOLO pipeline
docs/refactor/ARCHITECTURE.md	Backend contracts, dependency direction
docs/refactor/REFACTOR_PLAN.md	Refactor milestones and gates
docs/refactor/PROJECT_HANDOFF.md	Persistent handoff for future agents
docs/vision-pipeline/	Vision architecture, API reference, verified stats
docs/advisory_workflow/	Advisory workflow plans
backend/README.md	Backend run/replacement details
SETUP_REPORT.md	Environment bootstrap report (versions, verification)

Status & License

Version 0.2.0. Research prototype built for a 7-day capstone demo; not intended for production use. No license is declared. Related research assets are referenced from the frontend (frontend/src/lib/constants.ts): the dataset on Hugging Face (RaiyanKhaan/krishokChat). The authoritative research papers are the local files in paper/done papers/ (see docs/PAPER_POLICY.md — the old arXiv listing is deprecated and must not be cited).