"""FastAPI application for the template-agnostic v2 backend.""" from __future__ import annotations import json import logging from contextlib import asynccontextmanager from pathlib import Path from fastapi import FastAPI, Request from fastapi.middleware.cors import CORSMiddleware from fastapi.responses import FileResponse from fastapi.staticfiles import StaticFiles from backend.config import settings from backend.domain import template_discoverer from backend.pii import scrubber as pii_scrubber from backend.rag.store import get_rag_store from backend.rag.types import TIER_MASTER, TIER_REFERENCE from backend.startup import run_startup_ingest from backend.utils.runtime_paths import ensure_data_drive_runtime_dirs ensure_data_drive_runtime_dirs() logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) _FRONTEND_DIR = Path(__file__).resolve().parent.parent / "frontend" _STARTUP_SUMMARY: dict = {} _PII_SCAN_PATHS = {"/api/report/generate", "/api/report/preview"} @asynccontextmanager async def lifespan(app: FastAPI): global _STARTUP_SUMMARY # Warm jina-reranker-v3 FIRST, before startup ingest touches the embedder, so # the host-RAM weight load happens before the embedder + ingest stack up their # own allocations. Off by default (reference_cross_encoder_warmup); the lazy # path loads it on the first reference mapping otherwise. # Non-fatal: degrades to multi-signal on failure or under the RAM guard. try: from backend.llm.reranker import warmup_cross_encoder if warmup_cross_encoder(): logger.info("jina-reranker-v3 warmed at startup.") except Exception: # noqa: BLE001 - warmup must never block startup logger.warning("jina-reranker-v3 warmup skipped (non-fatal).", exc_info=True) _STARTUP_SUMMARY = run_startup_ingest() from backend.storage import document_library recovered = document_library.recover_all_tenants_stale_processing() if recovered: logger.info("Startup recovered %d stale processing document(s)", recovered) yield app = FastAPI( title="RICS Template-Agnostic Backend (v2)", version="2.0.0", lifespan=lifespan ) app.add_middleware( CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"], ) @app.middleware("http") async def pii_request_middleware(request: Request, call_next): """Log PII in inbound notes for awareness (notes are not scrubbed; RAG docs are).""" if request.method == "POST" and request.url.path in _PII_SCAN_PATHS: body = await request.body() async def _receive(): return {"type": "http.request", "body": body, "more_body": False} request._receive = _receive # re-inject consumed body try: raw_notes = json.loads(body or b"{}").get("raw_notes", "") audit = pii_scrubber.detect_pii(raw_notes) flagged = {k: v for k, v in audit.items() if k in ("EMAIL", "POSTCODE")} if flagged: logger.warning( "Inbound notes contain PII (notes kept verbatim; scrub RAG uploads only): %s", flagged, ) except Exception: # noqa: BLE001 pass response = await call_next(request) return response # Routers from backend.api.routes import ( # noqa: E402 admin, auth, legacy_api_routes, photos, report, schema, upload, ) app.include_router(auth.router) app.include_router(legacy_api_routes.router) app.include_router(schema.router) app.include_router(report.router) app.include_router(photos.router) app.include_router(upload.router) if settings.master_template_upload_enabled: app.include_router(admin.router) @app.get("/metrics/summary", tags=["observability"]) def metrics_summary(limit: int = 50) -> dict: """Aggregate recent per-report telemetry from the local metrics sink. Reads the JSONL records written under the metrics dir (no live state), so it is cheap and safe to poll. Returns token/cost totals, average grounding pass rate + retrieval confidence, status/violation breakdowns and per-label LLM call latency over the most-recent reports. """ from backend.observability import tracing as observability return observability.recent_summary(limit=max(1, min(limit, 500))) @app.get("/health", tags=["health"]) def health() -> dict: tenant = settings.default_tenant_id loaded = template_discoverer.load_schema(tenant) is not None store = get_rag_store() master_faiss_count = store.count(tenant, TIER_MASTER) reference_faiss_count = store.count(tenant, TIER_REFERENCE) # Reference-only model: readiness depends on the (canonical) schema being # installed, not on a shared master being seeded. When the operator master is # explicitly enabled, its chunks must be present for an "ok" status. master_ready = (not settings.master_template_auto_ingest) or master_faiss_count > 0 from backend.llm.embeddings import get_embedder try: emb = get_embedder() embedder_name = type(emb).__name__ embed_dim = int(getattr(emb, "embed_dim", 0)) embedder_ok = True embedder_error = None except Exception as exc: # noqa: BLE001 - surface load failure in health embedder_name = None embed_dim = 0 embedder_ok = False embedder_error = str(exc) return { "status": "ok" if loaded and master_ready and embedder_ok else "degraded", "embedder": embedder_name, "embed_dim": embed_dim, "embedder_ok": embedder_ok, "embedder_error": embedder_error, "master_loaded": loaded, "master_faiss_count": master_faiss_count, "reference_faiss_count": reference_faiss_count, "reference_ready": reference_faiss_count > 0, "startup": _STARTUP_SUMMARY, "admin_override_enabled": settings.master_template_upload_enabled, "ui_path": "/", } if _FRONTEND_DIR.is_dir(): app.mount("/static", StaticFiles(directory=str(_FRONTEND_DIR)), name="static") @app.get("/", include_in_schema=False) def root_ui() -> FileResponse: index = _FRONTEND_DIR / "index.html" if index.is_file(): return FileResponse(index) return FileResponse(_FRONTEND_DIR / "v2.html") @app.get("/v2.html", include_in_schema=False) def v2_ui() -> FileResponse: return FileResponse(_FRONTEND_DIR / "v2.html") @app.get("/index.html", include_in_schema=False) def legacy_index() -> FileResponse: return FileResponse(_FRONTEND_DIR / "index.html") @app.get("/favicon.ico", include_in_schema=False) def favicon() -> FileResponse: path = _FRONTEND_DIR / "favicon.ico" if path.is_file(): return FileResponse(path) return FileResponse(_FRONTEND_DIR / "v2.html")