Module extracting_backend 

LLM fact-extraction wrapper around any oxicode_agent::tools::MemoryBackend.

When settings.memory_llm_extract = true, every put(content, …) is sent through a structured extraction prompt and the model’s JSON response is fanned out into one inner.put(fact_text, …) per atomic fact.

Gap-2 design: the LLM call must run inside the async runtime. The LLM extractor is implemented directly as an async function that awaits the provider stream. The sync FactExtractor trait is retained only for the heuristic path (pure CPU, no I/O).

Failure policy: when the extractor errors, returns no facts, or returns malformed JSON, the wrapper falls back to storing content verbatim so memory_retain never silently drops user input.

Structs

ExtractedFact

One atomic fact returned by the LLM extractor.

ExtractingMemoryBackend

Wrapper around any MemoryBackend.

HeuristicFactExtractor

Default heuristic — splits on sentence boundaries.

LlmExtractor

Async LLM fact extractor. Owns a Provider and a Model.

Constants

STAGE_ONE_SYSTEM_PROMPT

omp stage_one_system.md, truncated to the parts that drive our extraction contract (we don’t need rollout_slug here — we only store a flat list of facts).

STAGE_ONE_USER_TEMPLATE

omp stage_one_input.md user template ({{content}}).

Traits

FactExtractor

Sync, no-IO extractor (heuristic). Safe to call anywhere.

Functions

try_make_llm_extractor

Resolve the LLM extractor from the wired Oxicode engine. Returns None when the user hasn’t configured memory_llm_extract_model or the pattern doesn’t resolve.

wrap_with_extractor

Wire the wrapper on top of an existing memory backend.