Apppublic

LuckmanShaik/AP_MSME_Backend

sourceHugging Facemitupdated 4mo agoView on Hugging Face

0likes

App README

AP MSME Chatbot API

FastAPI backend for the AP MSME ONE Portal chatbot. Deployed on Hugging Face Spaces with Docker.

Endpoints

—GET /health — pipeline readiness
—POST /auth/request-otp — send OTP to a 10-digit mobile (demo OTP: 123456)
—POST /auth/verify-otp — verify OTP and receive bearer token
—POST /chat — main RAG chat endpoint (requires Authorization: Bearer <token>)
—POST /speech/stt — speech-to-text (multipart audio upload)
—POST /speech/tts — text-to-speech (returns audio/wav or audio/mpeg)
—GET /docs — Swagger UI

Secrets (HF Space Settings → Variables and secrets)

The Docker build only installs dependencies. You must add runtime secrets on the Space (Settings → Variables and secrets). Copy values from your local .env in the main project.

Secret	Required	Notes
`MILVUS_URI`	Yes	Zilliz Cloud public endpoint (not `your-cluster.zillizcloud.com`)
`MILVUS_TOKEN`	Yes	Zilliz API key
`COLLECTION_NAME`	Yes	e.g. `apmsme_chunks_v3`
`LLM_PROVIDER`	Yes	e.g. `azure`
`AZURE_OPENAI_*` or other LLM keys	Yes	Match your provider
`FRONTEND_URL`	Yes	Vercel URL for CORS (e.g. `https://your-app.vercel.app`)
`BACKEND_URL`	Optional	Auto-detected from `SPACE_HOST` if omitted
`HF_TOKEN`	Optional	Hugging Face token — faster model downloads

If startup fails with your-cluster.zillizcloud.com, MILVUS_URI / MILVUS_TOKEN were not set on the Space.

Local run

bash

pip install -r requirements.txt
cp .env.example .env   # fill in secrets
python run.py

Docker

requirements.txt is pinned for fast HF builds (CPU PyTorch, no optimum/CUDA wheels). The Dockerfile installs torch from the PyTorch CPU index before other deps.

bash

docker build -t ap-msme-backend .
docker run -p 7860:7860 --env-file .env ap-msme-backend

Windows local dev (Python 3.13 ONNX embeddings only):

bash

pip install -r requirements.txt
pip install optimum[onnxruntime]==1.17.1 onnxruntime==1.23.2

BM25 corpus (optional)

Place output/clean_chunks.json in the repo for keyword search. Vector search works without it.