Spaces:
Paused
Paused
| """PII detection — server-side CPU inference. | |
| LiquidAI/LFM2.5-Encoder-350M-PII-Detector (LFM2.5 encoder + token-classification head + hybrid regex | |
| decode) runs on the Space's CPU. The browser POSTs {text} to /api/detect and | |
| gets back char-anchored PII spans (type/domain/label) + a redacted string, then | |
| highlights them. Nothing is downloaded to the visitor. | |
| The model repo is private, so an HF_TOKEN Space secret with read access is | |
| required to pull it. | |
| """ | |
| import os | |
| import time | |
| import torch | |
| from flask import Flask, request, send_from_directory | |
| import pii as P | |
| APP_DIR = os.path.dirname(os.path.abspath(__file__)) | |
| # Match threads to the container's cgroup CPU quota, not the host core count. | |
| _CPUS = P.effective_cpus() | |
| torch.set_num_threads(_CPUS) | |
| try: | |
| torch.set_num_interop_threads(1) | |
| except RuntimeError: | |
| pass | |
| app = Flask(__name__) | |
| print(f"[boot] loading pii-detect on {_CPUS} CPU thread(s), f32…", flush=True) | |
| MODEL = P.PiiDetector() | |
| MODEL.detect("Warm up: email jane@example.com, SSN 123-45-6789.") # warm the graph | |
| print("[boot] detector ready", flush=True) | |
| def index(): | |
| return send_from_directory(APP_DIR, "index.html") | |
| def static_file(name): | |
| if name in ("style.css", "lliquid.gif"): | |
| return send_from_directory(APP_DIR, name) | |
| return ("not found", 404) | |
| def ready(): | |
| return {"ready": MODEL is not None} | |
| def api_detect(): | |
| data = request.get_json(silent=True) or {} | |
| text = (data.get("text") or "").replace("\r", "") | |
| if not text.strip(): | |
| return {"spans": [], "redacted": "", "count": 0, "ms": 0} | |
| t0 = time.perf_counter() | |
| out = MODEL.detect(text) | |
| out["ms"] = round((time.perf_counter() - t0) * 1000) | |
| return out | |
| if __name__ == "__main__": | |
| app.run(host="0.0.0.0", port=int(os.environ.get("PORT", 7860)), threaded=True) | |