Skip to main content
Version: Next

mellea.stdlib.frameworks.react

ReACT (Reason + Act) agentic pattern implementation.

Provides the react() async function, which drives a tool-use loop: the model reasons about a goal, selects a tool, receives the result as an observation, and repeats until it calls final_answer or the loop_budget is exhausted. Accepts any list of AbstractMelleaTool instances and a ChatContext for multi-turn history tracking. Raises RuntimeError if the loop ends without a final answer.

Functions​

FUNC react ​

react(goal: str, context: ChatContext, backend: Backend) -> tuple[ComputedModelOutputThunk[str], ChatContext]

Asynchronous ReACT pattern (Think -> Act -> Observe -> Repeat Until Done); attempts to accomplish the provided goal given the provided tools.

Args:

  • goal: the goal to be accomplished or the question to answer
  • context: the context being used; a type of ChatContext
  • backend: the backend used to generate the response.
  • format: if set, the BaseModel to use for constrained decoding.
  • model_options: additional model options, which will upsert into the model/backend's defaults.
  • tools: the list of tools to use
  • loop_budget: the number of steps allowed; use -1 for unlimited
  • compactor: optional sync Compactor invoked once per turn after the tool observation. Use this for strategies that should fire at turn boundaries rather than on every component append (per-add compaction is configured on context itself). Compose with :func:mellea.stdlib.components.react.pin_react_initiator to preserve the goal across compactions. Compactors that need to call the backend (e.g. LLMSummarizeCompactor) hide the async work behind their sync compact method internally.

The Compactor protocol is generic in T bound to Context, so a compactor that returns a non-chat Context (e.g. SimpleContext) type-checks here but breaks the ChatContext assumptions the next aact call relies on. Custom compactors used with react must return a ChatContext.

Returns:

  • A (ModelOutputThunk, Context) if return_sampling_results is False, else returns a SamplingResult.

Raises:

  • RuntimeError: if the loop ends before a final answer is found