server.py

#!/usr/bin/env python3
"""FastAPI reference-voice search server. Free-text query -> BM25 or embedding (vector) similarity ->
top-N voices with their single best-of-three audio. Serves the demo page + audio + a /search endpoint.
Run: CUDA_VISIBLE_DEVICES=<g> uvicorn server:app --host 0.0.0.0 --port 8778"""
import os, json, re, pickle
import numpy as np
from fastapi import FastAPI, Query
from fastapi.responses import HTMLResponse, FileResponse, JSONResponse
DS = os.environ.get("DS_DIR", "/run/user/1001/vcluster_annot/ds")
meta = json.load(open(f"{DS}/meta.json"))
emb = np.load(f"{DS}/emb.npy")
bm25 = pickle.load(open(f"{DS}/bm25.pkl", "rb"))
info = json.load(open(f"{DS}/index_info.json"))
def tok(s): return re.sub(r"[^a-z0-9 ]", " ", s.lower()).split()
_model = None
def model():
    global _model
    if _model is None:
        from sentence_transformers import SentenceTransformer
        _model = SentenceTransformer(info["embed_model"], trust_remote_code=True,
                                     device=os.environ.get("EMB_DEVICE", "cuda"))
    return _model
app = FastAPI()

@app.get("/search")
def search(q: str = Query(...), mode: str = "embed", k: int = 5):
    if mode == "bm25":
        sc = np.asarray(bm25.get_scores(tok(q)))
    else:
        qv = model().encode([q], normalize_embeddings=True)[0]
        sc = emb @ qv
    idx = np.argsort(-sc)[:k]
    return JSONResponse([{**meta[i], "score": float(sc[i]), "rank": r+1} for r, i in enumerate(idx)])

@app.get("/audio/{cid}.mp3")
def audio(cid: str):
    p = f"{DS}/audio/{cid}.mp3"
    return FileResponse(p) if os.path.exists(p) else JSONResponse({"error": "not found"}, 404)

@app.get("/", response_class=HTMLResponse)
def home():
    return open(f"{DS}/index.html").read()