Expand description
Turning finalized user speech into candidate observations.
This is a different question from retrieval planning, asked by a different call with a different schema: did the user just reveal something worth keeping? Conflating the two produces a model that stores what it was asked to recall.
Extraction never runs on a partial transcript. A partial may be revised, and evidence that can be revised is not evidence.
Structs§
- Bounded
Observation Extractor - Runs an extractor under a deadline.
- Observation
Extraction Context - What the observation extractor is given.
- Rule
Based Observation Extractor - A rule-based extractor covering explicit statements and memory commands.
Constants§
- OBSERVATION_
EXTRACTION_ INSTRUCTION - The system instruction for the observation extractor.
Traits§
- Memory
Observation Extractor - Extracts candidate memories from an utterance.
Functions§
- observation_
schema - The JSON Schema a structured-output extraction should be constrained to.