| # DocuAudit AI — environment template (see docs/DOCUAUDIT_AI_REQUIREMENTS.md) | |
| # LLM Provider: ollama | anthropic | openai | huggingface | |
| LLM_PROVIDER=ollama | |
| # OpenAI (optional) | |
| OPENAI_API_KEY= | |
| OPENAI_MODEL=gpt-4o | |
| OPENAI_EMBEDDING_MODEL=text-embedding-3-small | |
| # Anthropic (optional) | |
| ANTHROPIC_API_KEY= | |
| ANTHROPIC_MODEL=claude-3-5-sonnet-20241022 | |
| # Hugging Face Inference API (when LLM_PROVIDER=huggingface — typical on Hugging Face Spaces) | |
| # Use a fine-grained token with "Make calls to Inference Providers" / Inference API where required. | |
| HUGGINGFACE_API_KEY= | |
| # Use a model your Hub gates allow (e.g. Llama 3.8B under “Meta Llama 3”, or Mistral instruct). Llama 3.1 needs its own gate. Chat: hf-inference then router auto. | |
| #HUGGINGFACE_MODEL=mistralai/Mistral-7B-Instruct-v0.3 | |
| #HUGGINGFACE_MODEL=meta-llama/Meta-Llama-3.1-8B-Instruct | |
| HUGGINGFACE_MODEL=meta-llama/Meta-Llama-3-8B-Instruct | |
| HUGGINGFACE_EMBEDDING_MODEL=sentence-transformers/all-MiniLM-L6-v2 | |
| # Optional: huggingface_hub InferenceClient provider. Leave unset: primary hf-inference, then router auto for chat (Mistral instruct ids also try Novita). | |
| # Use `auto` for router-only primary client (may pick Novita and break some models). | |
| HUGGINGFACE_INFERENCE_PROVIDER= | |
| # On Hugging Face Spaces you can omit HUGGINGFACE_API_KEY if the Space provides HF_TOKEN (mapped | |
| # automatically when LLM_PROVIDER=huggingface). For local .env you can set HF_TOKEN instead. | |
| # Ollama (recommended local default) | |
| OLLAMA_BASE_URL=http://localhost:11434 | |
| OLLAMA_CHAT_MODEL=llama3.1:8b | |
| OLLAMA_EMBEDDING_MODEL=nomic-embed-text | |
| # App | |
| APP_NAME=DocuAudit AI | |
| APP_VERSION=1.0.0 | |
| DEBUG=false | |
| MAX_FILE_SIZE_MB=50 | |
| # Spec name alias (optional; mapped to MAX_FILE_SIZE_MB in settings) | |
| MAX_UPLOAD_SIZE_MB= | |
| # ChromaDB | |
| CHROMA_PERSIST_DIRECTORY=./data/chroma | |
| CHROMA_PERSIST_DIR= | |
| CHROMA_COLLECTION_NAME=docuaudit_docs | |
| # Chunking | |
| CHUNK_SIZE=1000 | |
| CHUNK_OVERLAP=200 | |
| # Retrieval default (overridable per request on /query/ask via top_k) | |
| TOP_K_RESULTS=5 | |
| # Audit + jobs SQLite | |
| AUDIT_DB_PATH=./audit.db | |
| JOBS_DB_PATH=./data/jobs.db | |
| # Limits | |
| MAX_DOCUMENTS_PER_BATCH=100 | |
| # URL ingest (POST /ingest/url). SEC.gov blocks undeclared bots — use "Company Name you@email.com". | |
| # INGEST_USER_AGENT=DocuAudit AI you@example.com | |
| # Streamlit → API (Streamlit process reads these when set in the shell / OS env) | |
| STREAMLIT_BACKEND_URL=http://localhost:8000 | |
| DOC_AUDI_API_BASE=http://127.0.0.1:8000 | |
| # Read timeout (seconds) for Ask/Summarise HTTP calls; default in code is 3600 if unset | |
| DOC_AUDI_HTTP_READ_TIMEOUT=3600 | |
| # --- Docker Compose (Milestone 12) --- | |
| # Copy this file to `.env` before `docker compose up` (Compose loads `.env` for substitution and `env_file`). | |
| # | |
| # Persistent paths below are overridden in docker-compose.yml to a single volume mount at /data: | |
| # CHROMA_PERSIST_DIRECTORY=/data/chroma, AUDIT_DB_PATH=/data/audit.db, JOBS_DB_PATH=/data/jobs.db | |
| # You do not need to duplicate those in .env for compose unless you use a custom override file. | |
| # | |
| # Ollama from the API container cannot reach localhost on your machine; default in compose is: | |
| # OLLAMA_BASE_URL=http://host.docker.internal:11434 | |
| # (extra_hosts host-gateway is set for Linux.) Run `ollama serve` on the host, or start the bundled | |
| # Ollama service: docker compose --profile ollama up -d | |
| # When using the compose `ollama` profile, set in .env: | |
| # OLLAMA_BASE_URL=http://ollama:11434 | |
| # | |
| # Compose sets DOC_AUDI_API_BASE / STREAMLIT_BACKEND_URL to http://api:8000 for the Streamlit service | |
| # so server-side HTTP calls reach the API on the Docker network (do not override for UI in compose). | |
| # | |
| # Optional port overrides: API_PORT=8000, STREAMLIT_PORT=8501, OLLAMA_HOST_PORT=11434 | |
| # --- Hugging Face Spaces --- | |
| # Recommended for CPU Spaces (no Ollama): set in Space Settings → Repository secrets → Variables | |
| # LLM_PROVIDER=huggingface | |
| # HUGGINGFACE_API_KEY=<token> OR rely on built-in HF_TOKEN (same value as a Hub token secret) | |
| # HUGGINGFACE_MODEL / HUGGINGFACE_EMBEDDING_MODEL as needed | |
| # If the API runs in a second Space or external URL, set for the Streamlit Space: | |
| # DOC_AUDI_API_BASE=https://your-api....hf.space (or your FastAPI public URL) | |
| # Streamlit on Spaces must listen on port 8501 (default). Entry file: app.py (see docs/HUGGING_FACE_SPACES.md). | |
| # On Streamlit SDK Spaces, only Streamlit starts by default; app.py auto-starts uvicorn on 127.0.0.1:8000 when | |
| # SPACE_ID is set (built-in Hub env). Set DOC_AUDI_EMBED_API=0 to disable if you use a separate API URL above. | |
| # Repository secrets (HF_TOKEN / HUGGINGFACE_API_KEY) are copied from st.secrets into the API subprocess env. | |