mempalace.kg_llm_extractor

Source: mempalace/kg_llm_extractor.py

LLM-based triple extractor for the async KG worker.

Produces (subject, predicate, object) triples from drawer text by calling an OpenAI-compatible chat-completions endpoint (llama-server, Ollama, vLLM, etc.). Pure module — no DB, no AGE imports — so it can be unit-tested with a mocked httpx.AsyncClient.

The extraction prompt template lives in PROMPT_TEMPLATE below and mirrors the spec at docs/specs/kg-triple-extraction.md lines 87-102.

The function returns [] on any parse failure; the worker decides retry policy. Validation drops triples where subject == object, the predicate is empty, or either endpoint is a meta-stopword.

Classes

class Triple

Functions

extract_triples

async def extract_triples(client: Any, endpoint: str, model: str, drawer_text: str, *, timeout: float = 60.0, use_response_format: bool = True) -> list[Triple]

Call the LLM and parse out validated triples.

Args: client: An httpx.AsyncClient (or any object exposing a compatible async post method — useful for tests). endpoint: Base URL of the OpenAI-compatible server (e.g. http://familiar:11436). The function appends /v1/chat/completions if the path isn't already present. model: Model alias (e.g. phi-4-mini). drawer_text: Raw drawer document. Tail-truncated to the last TEXT_TAIL_LIMIT characters before prompting. timeout: Per-request timeout in seconds. use_response_format: First try response_format={"type":"json_object"} for structured output (llama-server supports this via GBNF); on 4xx/5xx for that field, retry without it and fall back to regex parsing.

Returns [] on any failure (network, JSON parse, validation). The worker decides retry behavior — this function never raises.