LLM protocol, so the pipeline only needs generate() and
generate_streaming() methods.
Install a provider
Local providers
preload_model=False when startup should not send a
warm-up request.
Hosted providers
Use an adapter directly
Adapters may be called without a RAG pipeline:RAGPipeline or
AgenticPipeline. Conversation state is explicit; use the history argument or create a session.
Custom endpoints and models
OpenAILLM accepts api_base and api_key, which makes it suitable for compatible endpoints.
For a fully custom backend, implement the small LLM protocol instead of inheriting from a
RAGLight base class. See Configuration and extensions.