Spaces:
Running on Zero
Running on Zero
Download returnguard/config.py from RajTheBuilder/ReturnGuard: direct link, hf CLI and curl.
- Browser
- Download file 4.78 kB
-
https://huggingface.co/spaces/RajTheBuilder/ReturnGuard/resolve/main/returnguard/config.py
- Command line
-
hf download hf://spaces/RajTheBuilder/ReturnGuard/returnguard/config.py
-
curl -L -o config.py https://huggingface.co/spaces/RajTheBuilder/ReturnGuard/resolve/main/returnguard/config.py
4.78 kB
| """Central configuration. Every threshold lives here so the code, README and | |
| website can't drift apart again. Values are chosen by scripts/evaluate.py.""" | |
| import os | |
| from pathlib import Path | |
| from dotenv import load_dotenv | |
| ROOT = Path(__file__).resolve().parent.parent | |
| load_dotenv(ROOT / ".env") | |
| ANGLES = ("front", "back", "side") | |
| # ββ Storage ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ | |
| DATA_DIR = Path(os.getenv("RG_DATA_DIR", ROOT / "data")) | |
| DB_PATH = DATA_DIR / "returnguard.db" | |
| IMAGE_DIR = DATA_DIR / "images" | |
| CACHE_DIR = Path(os.getenv("RG_CACHE_DIR", ROOT / ".cache")) | |
| RESULTS_PATH = ROOT / "results" / "eval_results.json" | |
| # ββ Models βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ | |
| BG_MODEL = os.getenv("RG_BG_MODEL", "u2net") | |
| EMBED_MODEL = os.getenv("RG_EMBED_MODEL", "facebook/dinov2-base") | |
| EMBED_SIZE = int(os.getenv("RG_EMBED_SIZE", "224")) # input side; patch grid = size / 14 | |
| # gemini-2.5-flash is no longer offered to new API keys, so the default is a current model | |
| LLM_MODEL = os.getenv("RG_LLM_MODEL", "gemini-3.8-flash") | |
| # Google advises keeping Gemini 3 models at their default temperature (1.0); older models run at 0. | |
| LLM_TEMPERATURE = float(os.environ["RG_LLM_TEMPERATURE"]) if os.getenv("RG_LLM_TEMPERATURE") else None | |
| LLM_TIMEOUT_S = float(os.getenv("RG_LLM_TIMEOUT_S", "90")) | |
| LLM_MAX_RETRIES = int(os.getenv("RG_LLM_MAX_RETRIES", "3")) | |
| LLM_MIN_INTERVAL_S = float(os.getenv("RG_LLM_MIN_INTERVAL_S", "0")) # client-side rate limit between calls | |
| LLM_MAX_WAIT_S = float(os.getenv("RG_LLM_MAX_WAIT_S", "75")) # never sleep longer than this between retries | |
| LLM_CACHE = os.getenv("RG_LLM_CACHE", "0") == "1" # on for evaluation runs only | |
| # ββ Upload limits ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ | |
| MAX_UPLOAD_MB = 15 | |
| # Public demo protection: deliveries + returns per IP per hour (0 = unlimited) | |
| RATE_LIMIT_PER_HOUR = int(os.getenv("RG_RATE_LIMIT_PER_HOUR", "0")) | |
| MAX_IMAGE_SIDE = 1024 | |
| # ββ Capture quality gate βββββββββββββββββββββββββββββββββββββββββββββββββββ | |
| MIN_OBJECT_FRACTION = 0.02 # object must cover at least 2% of the frame | |
| # Laplacian variance inside the object. Smooth steel or matte plastic is naturally | |
| # low-texture (5-20), so only near-featureless captures are sent back for a retake. | |
| MIN_SHARPNESS = 2.0 | |
| # ββ Visual screen (local, no LLM) ββββββββββββββββββββββββββββββββββββββββββ | |
| # On the evaluation set, object-only DINOv2 similarity (mean of 3 views) is | |
| # 0.37-0.53 for different products, 0.72-0.85 for another unit of the same or a | |
| # similar model, and 0.89-0.99 for the same item photographed again. | |
| # Fast reject: clearly a different product, no LLM needed. | |
| FAST_REJECT_SHAPE = 0.60 | |
| # Fast accept: every view matches strongly in shape and colour, and the delivery | |
| # fingerprint has no printed identifiers that still need reading. | |
| FAST_ACCEPT_SHAPE = 0.92 | |
| FAST_ACCEPT_MIN_VIEW = 0.85 | |
| FAST_ACCEPT_MAX_COLOR_SHIFT = 3.0 # Delta E (a*b*) between median object colours | |
| # ββ Decision routing on the 0β100 evidence-weighted risk score βββββββββββββ | |
| ACCEPT_BELOW = 50 | |
| REJECT_FROM = 90 | |
| # Customer history makes the policy stricter: each prior flagged return lowers | |
| # the acceptance cut-off by this much (capped). | |
| HISTORY_PENALTY_PER_FLAG = 10 | |
| HISTORY_PENALTY_CAP = 25 | |
| # ββ Return policy ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ | |
| RETURN_WINDOW_DAYS = {"electronics": 7, "apparel": 30, "home_kitchen": 10, "other": 10} | |
| FUNCTIONAL_TEST_CATEGORIES = {"electronics"} | |
| # Wear-and-return ("wardrobing") risk: a worn item is the same unit, so it passes the | |
| # shape, colour and label checks. These categories always get a visual inspection. | |
| WEAR_CHECK_CATEGORIES = {"apparel"} | |
| RETURN_REASONS = { | |
| "changed_mind": "Changed my mind / no longer needed", | |
| "damaged": "Item arrived damaged", | |
| "defective": "Item is defective / not working", | |
| "wrong_item": "Received the wrong item", | |
| "not_as_described": "Not as described", | |
| } | |