StillWorking's picture
Upload .env.example with huggingface_hub
e404d39 verified
Raw History Blame Contribute Delete
8.24 kB
# ── Copy this file to .env and fill in the values ────────────────────────────
#
# cp .env.example .env
# # then edit .env and set your API key
#
# NEVER commit the real .env file. If a key was ever committed or shared,
# rotate it in the provider console.
# ─────────────────────────────────────────────────────────────────────────────
# Google Gemini API key (required for alerts / query parsing)
# Obtain from: https://aistudio.google.com/app/apikey
# Either variable name is accepted; GEMINI_API_KEY takes precedence.
# ─────────────────────────────────────────────────────────────────────────────
GEMINI_API_KEY=your_gemini_api_key_here
# GOOGLE_API_KEY=your_gemini_api_key_here # legacy alias, also accepted
# Gemini model to use for reasoning (change to test different models)
GEMINI_MODEL=gemini-2.5-flash
# ─────────────────────────────────────────────────────────────────────────────
# Optional cloud fallbacks (tried in order after Gemini)
# Leave blank to skip a provider; the server degrades to the local keyword
# matcher / offline alert instead of crashing.
# ─────────────────────────────────────────────────────────────────────────────
GROQ_API_KEY=
GROQ_MODEL=openai/gpt-oss-120b
COHERE_API_KEY=
COHERE_MODEL=command-r-plus-08-2024
# ─────────────────────────────────────────────────────────────────────────────
# cascadeflow guardrail budgets
# These cap per-call LLM spend and latency. Adjust for your workload.
# ─────────────────────────────────────────────────────────────────────────────
CASCADEFLOW_ALERT_BUDGET=0.01 # USD cap for generate_alert()
CASCADEFLOW_SEARCH_BUDGET=0.005 # USD cap for extract_attributes_from_query()
CASCADEFLOW_LATENCY_LIMIT_MS=2000 # advisory target recorded by cascadeflow
# cascadeflow does not abort slow providers, so the API enforces its own hard
# deadlines. The whole Gemini β†’ Groq β†’ Cohere chain is bounded by
# LLM_HARD_TIMEOUT_MS; each provider attempt is bounded by
# LLM_PROVIDER_TIMEOUT_MS. Raise these if you prefer answer quality over the
# 2 s target (for example 8000/4000).
LLM_HARD_TIMEOUT_MS=2000
LLM_PROVIDER_TIMEOUT_MS=1200
# ─────────────────────────────────────────────────────────────────────────────
# API security (shared / networked deployments)
# ─────────────────────────────────────────────────────────────────────────────
# Leave SENTINEL_API_KEY empty for the open local demo. When set, every
# /api/v1 request must send either:
# Authorization: Bearer <key> or X-API-Key: <key>
# Set the matching VITE_API_KEY in frontend/.env. Use a demo key here, not a
# provider key: VITE_ variables are visible in the browser bundle.
SENTINEL_API_KEY=
# Per-client token bucket. RATE_LIMIT_PER_MINUTE=0 disables it.
RATE_LIMIT_PER_MINUTE=60
RATE_LIMIT_BURST=20
# Only enable behind a trusted reverse proxy that overwrites X-Forwarded-For.
TRUST_PROXY_HEADERS=0
# ─────────────────────────────────────────────────────────────────────────────
# Hardware / performance
# ─────────────────────────────────────────────────────────────────────────────
# Set to "cuda" for GPU inference, "cpu" for CPU-only fallback
TORCH_DEVICE=cuda
# ─────────────────────────────────────────────────────────────────────────────
# Face detection / matching
# ─────────────────────────────────────────────────────────────────────────────
# MTCNN minimum face size in pixels and minimum detection probability
MTCNN_MIN_FACE_SIZE=20
DETECTION_CONFIDENCE=0.6
# If the first MTCNN pass finds no face, retry once with these relaxed values.
MTCNN_RETRY_MIN_FACE_SIZE=12
MTCNN_RETRY_CONFIDENCE=0.45
# Context margin around the detected face (CelebA-like framing). 1.5 measured
# best on the bundled dataset; lower crops tighter, higher adds more background.
FACE_CROP_MARGIN=1.5
# Cosine similarity required to report a database match. With both the index
# and the query using MTCNN landmark alignment, correct identities scored
# ~0.98 (min 0.90 over a 48-image sample); unrelated portraits stayed < 0.35.
MATCH_SIMILARITY_THRESHOLD=0.50
# Hard upload size cap for /analyze-frame (bytes)
MAX_UPLOAD_BYTES=10485760
# ─────────────────────────────────────────────────────────────────────────────
# /generate-person image model (optional endpoint)
# Imagen 4 was retired from the Gemini API on 17 Aug 2026. The replacement is
# an image-capable Gemini model called through generate_content().
# ─────────────────────────────────────────────────────────────────────────────
IMAGE_MODEL=gemini-3.1-flash-image
IMAGE_COST_USD=0.067
# ─────────────────────────────────────────────────────────────────────────────
# CORS β€” comma-separated origins, or "*" to allow all (not recommended)
# ─────────────────────────────────────────────────────────────────────────────
CORS_ORIGINS=http://localhost:5173,http://127.0.0.1:5173,http://localhost:8080,http://127.0.0.1:8080,http://localhost:4173,http://127.0.0.1:4173
# ─────────────────────────────────────────────────────────────────────────────
# Paths (relative paths are resolved against backend/, then the project root)
# ─────────────────────────────────────────────────────────────────────────────
CELEBA_IMAGES_DIR=../Data/img_align_celeba
CELEBA_ATTR_CSV=../Data/list_attr_celeba.csv
FAISS_INDEX_PATH=./index_artifacts/celeb_index.faiss
METADATA_PATH=./index_artifacts/metadata.json
ADAFACE_WEIGHTS=./models_cache/adaface_ir50_ms1mv2.ckpt