Skip to content

symfonic.core.learning.entity_extractor_llm

entity_extractor_llm

LLM-backed entity extractor.

Reuses the agent's bound chat model. Best-effort JSON parsing -- a malformed response yields an empty candidate list, never a raise. Caching per-episodic results via a content-hash key is out of scope for v1 (design §9.3 documents the LLM extractor as non-idempotent).

Design reference: §17.1.4 / §13.4 (prompt template).

LlmEntityExtractor 

LlmEntityExtractor(
    chat_model: Any,
    *,
    max_entities: int = 10,
    observability_hook: ObservabilityHook | None = None,
)

Decision 4.5 strategy. Uses the agent's bound chat model.

Source code in src/symfonic/core/learning/entity_extractor_llm.py 
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
def __init__(
    self,
    chat_model: Any,
    *,
    max_entities: int = 10,
    observability_hook: ObservabilityHook | None = None,
) -> None:
    self._chat_model = chat_model
    self._max_entities = max_entities
    # v8.3.2 observability-bug fix (G6): the entity-extractor LLM call
    # emitted ONLY on the merge-dependent callback manager (same bug
    # class as the v8.3.1 critic).  When a hook is supplied it also
    # emits on the always-on ``ObservabilityHook``.  ``None`` resolves
    # to ``NoOpObservabilityHook`` so the default is byte-identical.
    # NOTE: the live caller (background ``SleepConsolidator`` ->
    # ``_phases_entity_helpers``) does NOT yet thread a real hook (it
    # calls ``extract(content)`` with no callback_manager either) --
    # see v8.3.2 report for the plumbing this would need.
    if observability_hook is None:
        from symfonic.core.observability.hooks import NoOpObservabilityHook

        observability_hook = NoOpObservabilityHook()
    self._observability_hook = observability_hook

extract async 

extract(
    text: str,
    *,
    callback_manager: CallbackManager | None = None,
    run_id: str = "",
    model_name: str = "entity_extractor_llm",
) -> list[EntityCandidate]

Extract entity candidates from episodic text.

v7.4.3 (Jarvio Ask 5): when callback_manager is provided and the underlying chat model exposes ainvoke, fires on_llm_end(node_name="entity_extractor_llm") so the gated Phase 11 LLM call is visible to token billing / OTel. None preserves byte-identical pre-7.4.3 behaviour.

Source code in src/symfonic/core/learning/entity_extractor_llm.py 
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
async def extract(
    self,
    text: str,
    *,
    callback_manager: CallbackManager | None = None,
    run_id: str = "",
    model_name: str = "entity_extractor_llm",
) -> list[EntityCandidate]:
    """Extract entity candidates from episodic text.

    v7.4.3 (Jarvio Ask 5): when ``callback_manager`` is provided and
    the underlying chat model exposes ``ainvoke``, fires
    ``on_llm_end(node_name="entity_extractor_llm")`` so the gated
    Phase 11 LLM call is visible to token billing / OTel.  ``None``
    preserves byte-identical pre-7.4.3 behaviour.
    """
    if not text or self._chat_model is None:
        return []
    cleaned = strip_phase11_markers(text)
    if not cleaned:
        return []
    prompt = _EXTRACTION_PROMPT.format(
        episodic_content=cleaned[:50_000],
        max_entities=self._max_entities,
    )
    try:
        response = await self._invoke_model(
            prompt,
            callback_manager=callback_manager,
            run_id=run_id,
            model_name=model_name,
        )
    except Exception:
        logger.debug("entity_extractor_llm: model invocation failed",
                     exc_info=True)
        return []
    return self._parse_response(response)