LLM Providers

DocIntel supports multiple LLM providers. Configure via the LLM__PROVIDER environment variable.

Supported providers

Provider	LLM__PROVIDER value	Notes
OpenAI	openai	Default; uses ModelRegistry tiers
Azure OpenAI	azure	Set LLM__BASE_URL to your Azure endpoint
Anthropic	anthropic	Claude 3.x via tool-use mode
Google Gemini	gemini	Via Vertex AI or AI Studio
Ollama	ollama	Set LLM__BASE_URL; air-gapped deployments

Model Registry

All profiles reference models from the central ModelRegistry in config.py — never hardcoded strings. Upgrading the entire model stack is a four-line change.

Tier	Current model	Used by profiles
FLAGSHIP	gpt-5.6-sol	AEC, Contract
DEFAULT	gpt-5.6-terra	Financial, Scientific, General
VISION	gpt-5.6-sol	Image / sketch parsing
NANO	gpt-6-luna	High-volume, cost-sensitive extraction

OpenAI (default)

LLM__PROVIDER=openai
LLM__API_KEY=sk-proj-...

Uses the OpenAI chat completions API. Model selection is automatic per profile via ModelRegistry.

Azure OpenAI

LLM__PROVIDER=azure
LLM__API_KEY=your-azure-key
LLM__BASE_URL=https://your-resource.openai.azure.com/

Point LLM__BASE_URL at your Azure OpenAI resource endpoint. Ensure your deployment names match the model IDs in ModelRegistry, or override LLM__MODEL.

Anthropic

LLM__PROVIDER=anthropic
LLM__API_KEY=sk-ant-...

Uses Claude via the Anthropic messages API with tool-use mode for structured extraction.

Google Gemini

LLM__PROVIDER=gemini
LLM__API_KEY=your-gemini-key

Works with both Vertex AI and Google AI Studio endpoints.

Ollama (air-gapped)

LLM__PROVIDER=ollama
LLM__BASE_URL=http://localhost:11434
LLM__MODEL=llama3.1:70b

For fully air-gapped deployments with no external API calls. Pair with OCR__ENGINE=tesseract and SKETCH__ENGINE=local for a completely offline stack.

WARNING

Extraction quality with local models depends heavily on model size. We recommend ≥ 70B parameter models for production-quality extraction. Smaller models work for testing but may produce significantly more hallucinations.

Advanced settings

Variable	Default	Description
LLM__TEMPERATURE	0.0	Lower = more deterministic extraction
LLM__MAX_TOKENS	4096	Max response tokens per LLM call
LLM__TIMEOUT_S	60	Per-call timeout in seconds

See Configuration → for the full reference.