DSN / env.example
nexusbert's picture
Enhance Docker configuration by adding HUGGING_FACE_HUB_TOKEN as a build argument in docker-compose.yml and Dockerfile. Update Dockerfile to install CPU-only PyTorch and adjust thread counts for reduced memory usage during builds. Modify docker_build_assets.py to support both HF_TOKEN and HUGGING_FACE_HUB_TOKEN for model caching. Update README.md to clarify token usage and build instructions.
be705e8
# Copy to .env and fill in values. Never commit .env.
#
# Azure endpoint is the resource base URL only (no /deployments/... path), e.g.
# https://YOUR_RESOURCE.openai.azure.com/
# https://YOUR_RESOURCE.cognitiveservices.azure.com/
AZURE_OPENAI_ENDPOINT=https://YOUR_RESOURCE.openai.azure.com/
AZURE_OPENAI_API_KEY=
AZURE_OPENAI_API_VERSION=2024-06-01
# Task A — fine-tuned chat deployment name (exact string from Azure / Foundry)
AZURE_OPENAI_DEPLOYMENT_TASK_A=yelp-reviewer-v1
TASK_A_MAX_TOKENS=1024
TASK_A_TEMPERATURE=0.35
# Optional: Azure embeddings for catalog build / Task B retrieval.
# Leave blank to use local sentence-transformers (TASK_B_EMBEDDING_BACKEND=auto picks local).
AZURE_OPENAI_DEPLOYMENT_EMBEDDINGS=
# Task B Azure reranking only: chat deployment name. If blank, Task A deployment is used.
AZURE_OPENAI_DEPLOYMENT_CHAT=
# Task B retrieval: azure | local | auto (auto = azure if AZURE_OPENAI_DEPLOYMENT_EMBEDDINGS set)
TASK_B_EMBEDDING_BACKEND=auto
TASK_B_LOCAL_EMBEDDING_MODEL=all-MiniLM-L6-v2
# Task B reranking: azure | local (local uses Hugging Face model below; no Azure chat quota needed)
TASK_B_RANK_BACKEND=local
TASK_B_LOCAL_LLM_MODEL=Qwen/Qwen2.5-1.5B-Instruct
TASK_B_EMBEDDED_CATALOG=data/business_catalog_embedded.jsonl
# Hugging Face Hub token — optional for public models; use either name (Docker build sees HF_TOKEN if passed as build-arg).
# HF_TOKEN=
# HUGGING_FACE_HUB_TOKEN=
# Optional: bind-mount Yelp business JSON inside image build context at this path to bake a real catalog.
# YELP_BUSINESS_JSON=/code/yelp_dataset/extracted/yelp_academic_dataset_business.json
# Docker catalog size cap when Yelp present:
# DOCKER_CATALOG_MAX_ROWS=15000
# Container listens on 7860 (Hugging Face Spaces default); override locally if needed.
# PORT=7860
HF_TOKEN=hf....