Module llm 

Model-backed extraction.

Both extractor seams are structured-output calls over a BaseLlm, constrained by a schema derived from the decoding type. Deriving rather than hand-writing matters here: a schema that can drift from the struct it decodes into is a bug waiting for a model to find, and the wire types use the domain enums directly so constrained decoding can only produce values the domain already understands.

Everything the model returns is still a proposal. Caps are re-applied, confidences are clamped, instruction-shaped statements are dropped, and speaker attribution comes from the runtime — never from the model, because a model cannot know who was in the room.

Structs

GeminiObservationExtractor

An observation extractor backed by a Gemini model.

GeminiPlanExtractor

A retrieval-plan extractor backed by a Gemini model.

Constants

DEFAULT_EXTRACTION_MODEL

The default extraction model — the one retrieval planning uses.

DEFAULT_TRANSCRIPT_MODEL

The default model for extracting observations from a transcript.

Functions

extraction_llm

Build a Gemini LLM for out-of-band extraction from the environment.

observation_context

A turn’s worth of context, for callers driving the extractors directly.