LLM providers¶

LitCurate supports two backends. You choose them with provider on the global llm: block or on each LLM stage (search, abstract_filter, fulltext_filter, extraction).

Provider	Use for	Auth
`anthropic`	Claude models (Anthropic API)	`ANTHROPIC_API_KEY`
`openai_compatible`	Any OpenAI-style Chat Completions API	`OPENAI_API_KEY` (or a custom env via `api_key_env`)

Always set an explicit model: on every LLM stage. For openai_compatible, also set base_url.

Global llm: is a required fallback

Every config must include a global llm: block with provider: anthropic or provider: openai_compatible. There is no implicit default. Config load fails if it is missing.

Stages that set their own provider / base_url / api_key_env ignore the matching global fields. Stages that omit those fields fall back to the global llm: values.

How resolution works¶

Start from the required global llm: block (the fallback).
Overlay any per-stage provider, base_url, api_key_env, or timeout_seconds.
Use that stage’s own model: string when calling the API.

Example: if abstract_filter sets provider: openai_compatible and base_url, it uses those. If search only sets model: and no provider, it falls back to global llm.provider.

You can mix backends in one config — for example Anthropic for query generation and a local or cloud OpenAI-compatible model for screening.

Anthropic (Claude)¶

search:
  provider: anthropic
  model: claude-sonnet-4-6

# .env
ANTHROPIC_API_KEY=sk-ant-...

Omit base_url for Anthropic. LitCurate uses the official Anthropic SDK.

OpenAI-compatible APIs¶

openai_compatible posts to {base_url}/chat/completions. That covers OpenAI, DeepSeek, Grok (xAI), Ollama, vLLM, and most cloud proxies.

Service	Example `base_url`	Example `model`	Key
OpenAI	`https://api.openai.com/v1`	`gpt-4.1`	Real OpenAI key
DeepSeek	`https://api.deepseek.com/v1`	`deepseek-chat`	DeepSeek key
Grok (xAI)	`https://api.x.ai/v1`	`grok-3`	xAI key
Ollama (local)	`http://localhost:11434/v1`	`qwen3:14b`	Any non-empty value
vLLM / proxy	`http://HOST:PORT/v1`	Host-specific id	Host-specific

Use the exact model id your host documents. LitCurate does not rename models.

OpenAIDeepSeekGrok (xAI)Ollama

extraction:
  provider: openai_compatible
  base_url: https://api.openai.com/v1
  model: gpt-4.1

# .env
OPENAI_API_KEY=sk-...

extraction:
  provider: openai_compatible
  base_url: https://api.deepseek.com/v1
  model: deepseek-chat

# .env — store the DeepSeek key under OPENAI_API_KEY, or rename with api_key_env
OPENAI_API_KEY=sk-...

extraction:
  provider: openai_compatible
  base_url: https://api.x.ai/v1
  model: grok-3

# .env
OPENAI_API_KEY=xai-...

ollama pull qwen3:14b
ollama serve

abstract_filter:
  provider: openai_compatible
  base_url: http://localhost:11434/v1   # or http://GPU_HOST:11434/v1
  model: qwen3:14b

# .env — any non-empty value is fine for local Ollama
OPENAI_API_KEY=ollama

Custom API key env name¶

By default, openai_compatible reads OPENAI_API_KEY. To keep keys separate:

extraction:
  provider: openai_compatible
  base_url: https://api.deepseek.com/v1
  model: deepseek-chat
  api_key_env: DEEPSEEK_API_KEY

# .env
DEEPSEEK_API_KEY=sk-...

Global fallback (required)¶

Every config must declare a global llm: block. There is no implicit provider.

llm:
  provider: openai_compatible   # or anthropic
  base_url: http://localhost:11434/v1

This is the fallback for stages that omit provider / base_url / api_key_env. A stage can still override with its own provider and model.

Mixed example¶

llm:
  provider: openai_compatible
  base_url: http://localhost:11434/v1

search:
  provider: anthropic
  model: claude-sonnet-4-6

abstract_filter:
  provider: openai_compatible
  base_url: http://localhost:11434/v1
  model: qwen3:14b

extraction:
  provider: openai_compatible
  base_url: https://api.openai.com/v1
  model: gpt-4.1