Adapters reference

Available adapters

Name	Kind	Use when	Key constraints / notes
themis.adapters.openai(...)	Provider adapter	An OpenAI or OpenAI-compatible endpoint should handle generation while Themis owns evaluation and storage	Install themis-eval[openai] or inject a compatible client
themis.adapters.vllm(...)	Provider adapter	A vLLM or other OpenAI-compatible local endpoint should handle generation	Install themis-eval[vllm] on Linux or inject a compatible client
themis.adapters.langgraph(...)	Graph adapter	A LangGraph workflow should act as the generator	Pass a graph with invoke() or ainvoke()

Use adapters when Themis should still own planning, reduction, parsing, scoring, storage, and inspection, but a provider or graph runtime should perform generation.

Provider notes

Option	Best for	Persistence / runtime behavior	Caveats
OpenAI	Managed or self-hosted OpenAI-compatible endpoints	Themis still owns planning, storage, parsing, and scoring	Requires credentials or an injected client
vLLM	Local or self-hosted inference servers that speak an OpenAI-compatible API	Behaves like another provider endpoint inside Themis runtime controls	The packaged extra is Linux-only
LangGraph	Existing graph-driven generation workflows	Themis treats the graph as the generator while preserving the rest of the runtime	Best trace capture depends on astream_events() availability