Upload backend/main.py with huggingface_hub
Browse files- backend/main.py +188 -0
backend/main.py
ADDED
|
@@ -0,0 +1,188 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
main.py β Sentinel-AI
|
| 3 |
+
=========================
|
| 4 |
+
FastAPI application entry point.
|
| 5 |
+
|
| 6 |
+
Start the server (loopback by default; use --host 0.0.0.0 only with
|
| 7 |
+
SENTINEL_API_KEY and a TLS reverse proxy):
|
| 8 |
+
uvicorn main:app --host 127.0.0.1 --port 8000 --reload
|
| 9 |
+
|
| 10 |
+
Swagger UI: http://localhost:8000/docs
|
| 11 |
+
ReDoc: http://localhost:8000/redoc
|
| 12 |
+
"""
|
| 13 |
+
|
| 14 |
+
from __future__ import annotations
|
| 15 |
+
|
| 16 |
+
import os
|
| 17 |
+
from contextlib import asynccontextmanager
|
| 18 |
+
from pathlib import Path
|
| 19 |
+
|
| 20 |
+
from dotenv import load_dotenv
|
| 21 |
+
|
| 22 |
+
# Load .env β check backend dir first, then fall back to project root.
|
| 23 |
+
_backend_env = Path(__file__).parent / ".env"
|
| 24 |
+
_root_env = Path(__file__).parent.parent / ".env"
|
| 25 |
+
|
| 26 |
+
# Load backend .env first (takes precedence), then root .env as fallback.
|
| 27 |
+
load_dotenv(_backend_env)
|
| 28 |
+
load_dotenv(_root_env)
|
| 29 |
+
|
| 30 |
+
from fastapi import FastAPI
|
| 31 |
+
from fastapi.middleware.cors import CORSMiddleware
|
| 32 |
+
|
| 33 |
+
from api.routes.analyze import router as analyze_router
|
| 34 |
+
from api.routes.generate import router as generate_router
|
| 35 |
+
from api.routes.search import router as search_router
|
| 36 |
+
from api.schemas import HealthResponse
|
| 37 |
+
from core.faiss_store import (
|
| 38 |
+
get_faiss_order,
|
| 39 |
+
get_index,
|
| 40 |
+
get_metadata,
|
| 41 |
+
warm_attribute_index,
|
| 42 |
+
)
|
| 43 |
+
from core.models import (
|
| 44 |
+
detector_ready,
|
| 45 |
+
embedder_ready,
|
| 46 |
+
load_adaface,
|
| 47 |
+
load_face_detector,
|
| 48 |
+
)
|
| 49 |
+
from core.paths import resolve_path
|
| 50 |
+
from core.security import install_security
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
def _cors_origins() -> list[str]:
|
| 54 |
+
"""
|
| 55 |
+
Read CORS_ORIGINS as a comma-separated list.
|
| 56 |
+
|
| 57 |
+
Defaults to the local Vite dev server. Set ``CORS_ORIGINS=*`` to allow
|
| 58 |
+
every origin (the old hard-coded behaviour) when explicitly desired.
|
| 59 |
+
"""
|
| 60 |
+
raw = os.getenv(
|
| 61 |
+
"CORS_ORIGINS",
|
| 62 |
+
"http://localhost:5173,http://127.0.0.1:5173,"
|
| 63 |
+
"http://localhost:3000,http://127.0.0.1:3000,"
|
| 64 |
+
"http://localhost:8080,http://127.0.0.1:8080,"
|
| 65 |
+
"http://localhost:4173,http://127.0.0.1:4173",
|
| 66 |
+
)
|
| 67 |
+
origins = [origin.strip() for origin in raw.split(",") if origin.strip()]
|
| 68 |
+
return origins or ["http://localhost:5173"]
|
| 69 |
+
|
| 70 |
+
|
| 71 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 72 |
+
# Lifespan: warm-up all heavy singletons at startup
|
| 73 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 74 |
+
|
| 75 |
+
@asynccontextmanager
|
| 76 |
+
async def lifespan(app: FastAPI):
|
| 77 |
+
"""Load models and FAISS index once at startup."""
|
| 78 |
+
print("[Sentinel-AI] Warming up models and index ...")
|
| 79 |
+
|
| 80 |
+
# 1. Initialise cascadeflow harness β must run before any cascadeflow.run() scope
|
| 81 |
+
import cascadeflow
|
| 82 |
+
|
| 83 |
+
cf_report = cascadeflow.init(mode="enforce")
|
| 84 |
+
version = getattr(cascadeflow, "__version__", "unknown")
|
| 85 |
+
print(
|
| 86 |
+
f" [OK] cascadeflow {version} "
|
| 87 |
+
f"mode={cf_report.mode} instrumented={cf_report.instrumented}"
|
| 88 |
+
)
|
| 89 |
+
|
| 90 |
+
# 2. Load local CV models (MTCNN detector + AdaFace embedder)
|
| 91 |
+
try:
|
| 92 |
+
load_face_detector()
|
| 93 |
+
print(" [OK] MTCNN face detector loaded")
|
| 94 |
+
except Exception as exc:
|
| 95 |
+
print(f" [FAIL] MTCNN detector not available: {exc}")
|
| 96 |
+
|
| 97 |
+
try:
|
| 98 |
+
load_adaface()
|
| 99 |
+
print(" [OK] AdaFace embedder loaded")
|
| 100 |
+
except Exception as exc:
|
| 101 |
+
print(f" [FAIL] AdaFace not available: {exc}")
|
| 102 |
+
|
| 103 |
+
# 3. Load FAISS index + metadata into RAM and build the vectorised
|
| 104 |
+
# attribute matrix used by semantic search (one-off, ~8 MB).
|
| 105 |
+
try:
|
| 106 |
+
idx = get_index()
|
| 107 |
+
get_metadata()
|
| 108 |
+
get_faiss_order()
|
| 109 |
+
warm_attribute_index()
|
| 110 |
+
print(f" [OK] FAISS index loaded ({idx.ntotal:,} vectors)")
|
| 111 |
+
print(" [OK] Attribute index warmed")
|
| 112 |
+
except FileNotFoundError as exc:
|
| 113 |
+
print(f" [FAIL] {exc}")
|
| 114 |
+
except Exception as exc:
|
| 115 |
+
print(f" [FAIL] Attribute index warm-up failed: {exc}")
|
| 116 |
+
|
| 117 |
+
# 4. Pre-warm provider clients (validates keys, keeps first request fast)
|
| 118 |
+
from core.gemini_agent import warm_providers
|
| 119 |
+
|
| 120 |
+
warmed = warm_providers()
|
| 121 |
+
if warmed:
|
| 122 |
+
print(f" [OK] LLM clients warmed: {', '.join(warmed)}")
|
| 123 |
+
else:
|
| 124 |
+
print(" [WARN] No cloud LLM clients configured; using offline fallbacks")
|
| 125 |
+
|
| 126 |
+
print("[Sentinel-AI] Ready.")
|
| 127 |
+
yield
|
| 128 |
+
print("[Sentinel-AI] Shutting down.")
|
| 129 |
+
|
| 130 |
+
|
| 131 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 132 |
+
# App
|
| 133 |
+
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 134 |
+
|
| 135 |
+
app = FastAPI(
|
| 136 |
+
title="Sentinel-AI",
|
| 137 |
+
description=(
|
| 138 |
+
"Hybrid edge-cloud security agent. "
|
| 139 |
+
"Detects faces via MTCNN, embeds them with AdaFace, "
|
| 140 |
+
"searches 200k CelebA vectors in FAISS, and drafts alerts with Gemini."
|
| 141 |
+
),
|
| 142 |
+
version="1.0.0",
|
| 143 |
+
docs_url="/docs",
|
| 144 |
+
redoc_url="/redoc",
|
| 145 |
+
lifespan=lifespan,
|
| 146 |
+
)
|
| 147 |
+
|
| 148 |
+
# Install security *before* CORS: Starlette makes the last-added middleware
|
| 149 |
+
# outermost, and CORS must wrap auth so preflight OPTIONS requests are answered.
|
| 150 |
+
install_security(app)
|
| 151 |
+
|
| 152 |
+
app.add_middleware(
|
| 153 |
+
CORSMiddleware,
|
| 154 |
+
allow_origins=_cors_origins(),
|
| 155 |
+
allow_methods=["GET", "POST", "OPTIONS"],
|
| 156 |
+
allow_headers=["Content-Type", "Authorization", "X-API-Key"],
|
| 157 |
+
)
|
| 158 |
+
|
| 159 |
+
# ββ Routers βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 160 |
+
app.include_router(analyze_router, prefix="/api/v1", tags=["Inference"])
|
| 161 |
+
app.include_router(search_router, prefix="/api/v1", tags=["Search"])
|
| 162 |
+
app.include_router(generate_router, prefix="/api/v1", tags=["Generation"])
|
| 163 |
+
|
| 164 |
+
|
| 165 |
+
# ββ Health check ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
|
| 166 |
+
@app.get("/health", response_model=HealthResponse, tags=["System"])
|
| 167 |
+
async def health() -> HealthResponse:
|
| 168 |
+
"""Liveness probe with real index/model readiness telemetry."""
|
| 169 |
+
index_path = resolve_path(
|
| 170 |
+
os.getenv("FAISS_INDEX_PATH"),
|
| 171 |
+
"./index_artifacts/celeb_index.faiss",
|
| 172 |
+
)
|
| 173 |
+
index_ready = index_path.exists()
|
| 174 |
+
index_vectors: int | None = None
|
| 175 |
+
if index_ready:
|
| 176 |
+
try:
|
| 177 |
+
index_vectors = int(get_index().ntotal)
|
| 178 |
+
except Exception:
|
| 179 |
+
index_ready = False
|
| 180 |
+
|
| 181 |
+
return HealthResponse(
|
| 182 |
+
status="ok",
|
| 183 |
+
index_ready=index_ready,
|
| 184 |
+
index_vectors=index_vectors,
|
| 185 |
+
detector_ready=detector_ready(),
|
| 186 |
+
embedder_ready=embedder_ready(),
|
| 187 |
+
model=os.getenv("GEMINI_MODEL", "gemini-2.5-flash"),
|
| 188 |
+
)
|