vapi-sathi embed space: bge-m3 FastAPI /embed

Dockerfile

@@ -0,0 +1,20 @@
+# Hugging Face Docker Space — runs as user 1000, app must listen on 7860.
+FROM python:3.11-slim
+
+RUN useradd -m -u 1000 user
+USER user
+ENV HOME=/home/user \
+    PATH=/home/user/.local/bin:$PATH \
+    HF_HOME=/home/user/.cache/huggingface
+WORKDIR /home/user/app
+
+COPY --chown=user requirements.txt .
+RUN pip install --user --no-cache-dir -r requirements.txt
+
+# Bake the model into the image so the first request is not a ~2 GB cold pull.
+RUN python -c "from sentence_transformers import SentenceTransformer; SentenceTransformer('BAAI/bge-m3')"
+
+COPY --chown=user app.py .
+
+EXPOSE 7860
+CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860"]

README.md

@@ -0,0 +1,51 @@
+---
+title: Vapi Sathi Embeddings
+emoji: 🔎
+colorFrom: indigo
+colorTo: blue
+sdk: docker
+app_port: 7860
+pinned: false
+---
+
+# vapi-sathi embeddings
+
+FastAPI `/embed` service hosting **BAAI/bge-m3** (1024-dim, multilingual incl.
+Hindi/Gujarati). Offloads embedding off the NestJS API box (see
+`../INFRA-DEPLOYMENT-PLAN.md`). Deployed on a **free CPU Basic** Hugging Face
+Space (2 vCPU / 16 GB).
+
+## API
+
+`GET /` → health: `{ "status": "ok", "model": "...", "ready": true }`
+
+`POST /embed`
+```json
+{ "texts": ["latest mandi prices", "ભાવ સમાચાર"] }
+```
+→
+```json
+{ "embeddings": [[...1024 floats...], [...]], "dim": 1024, "model": "BAAI/bge-m3" }
+```
+
+Vectors are L2-normalized → use cosine / dot directly. bge-m3 needs **no
+query/passage prefix**.
+
+## Auth
+
+If `EMBED_TOKEN` is set (Space secret), every `/embed` call must send
+`Authorization: Bearer <EMBED_TOKEN>`. Leave unset only for private testing.
+The NestJS client reads `EMBEDDING_SERVICE_URL` + `EMBEDDING_SERVICE_TOKEN`.
+
+## Keep-warm
+
+Free Spaces sleep after ~48 h idle. Worker traffic keeps it warm; otherwise
+ping `GET /` on a cron (e.g. a NestJS `@Cron`).
+
+## Local test
+
+```bash
+docker build -t vapi-embed . && docker run -p 7860:7860 vapi-embed
+curl -s localhost:7860/embed -H 'content-type: application/json' \
+  -d '{"texts":["hello","नमस्ते"]}' | head -c 200
+```

app.py

@@ -0,0 +1,55 @@
+import os
+from contextlib import asynccontextmanager
+
+from fastapi import FastAPI, Header, HTTPException
+from pydantic import BaseModel
+from sentence_transformers import SentenceTransformer
+
+MODEL_NAME = os.getenv("EMBED_MODEL", "BAAI/bge-m3")
+EMBED_TOKEN = os.getenv("EMBED_TOKEN")  # optional shared secret
+
+_model: SentenceTransformer | None = None
+
+
+@asynccontextmanager
+async def lifespan(_: FastAPI):
+    global _model
+    _model = SentenceTransformer(MODEL_NAME, device="cpu")
+    yield
+
+
+app = FastAPI(lifespan=lifespan)
+
+
+class EmbedRequest(BaseModel):
+    texts: list[str]
+
+
+class EmbedResponse(BaseModel):
+    embeddings: list[list[float]]
+    dim: int
+    model: str
+
+
+@app.get("/")
+def health():
+    return {"status": "ok", "model": MODEL_NAME, "ready": _model is not None}
+
+
+# sync `def` → FastAPI runs it in a threadpool, so the blocking encode() does
+# not stall the event loop.
+@app.post("/embed", response_model=EmbedResponse)
+def embed(req: EmbedRequest, authorization: str | None = Header(default=None)):
+    if EMBED_TOKEN and authorization != f"Bearer {EMBED_TOKEN}":
+        raise HTTPException(status_code=401, detail="unauthorized")
+    if _model is None:
+        raise HTTPException(status_code=503, detail="model loading")
+    if not req.texts:
+        raise HTTPException(status_code=400, detail="texts required")
+
+    vecs = _model.encode(req.texts, normalize_embeddings=True, batch_size=16)
+    return EmbedResponse(
+        embeddings=[v.tolist() for v in vecs],
+        dim=len(vecs[0]),
+        model=MODEL_NAME,
+    )

requirements.txt

@@ -0,0 +1,3 @@
+fastapi==0.115.*
+uvicorn[standard]==0.32.*
+sentence-transformers==3.*