Skip to content

latent.agents.guided.context_builder

Prompt assembly and guideline matching strategies for GuidedAgent.

Functions

batch_match_guidelines

batch_match_guidelines(guidelines: list[Guideline], messages: list[Message], model: str) -> list[GuidelineMatch]

Single structured-output LLM call to evaluate all guideline conditions.

build_system_prompt

build_system_prompt(base_prompt: str, matched_guidelines: list[Guideline], mode: str = 'fluid', glossary: Glossary | None = None, messages: list[Message] | None = None, journey_state: JourneyState | None = None, canned_responses: list[CannedResponse] | None = None) -> str

Assemble the full system prompt for a single turn.

embedding_match_guidelines

embedding_match_guidelines(guidelines: list[Guideline], messages: list[Message], model: str, threshold: float = 0.75, guideline_embeddings: dict[str, list[float]] | None = None) -> list[GuidelineMatch]

Semantic similarity matching using litellm embeddings.

match_always

match_always(guidelines: list[Guideline]) -> list[GuidelineMatch]

Include all guidelines every turn — no LLM call.

resolve_guidelines

resolve_guidelines(matched: list[GuidelineMatch]) -> list[Guideline]

Apply exclusion rules and sort active guidelines by priority (descending).

Exclusions are applied in priority order: a high-priority guideline excludes lower-priority ones, but a low-priority guideline cannot exclude a higher-priority guideline that has already been accepted.