Documentation
¶
Overview ¶
Package extract distils conversation prose into durable, classified facts without an LLM — a port of mempalace's heuristic extractor. It pulls the decisions, preferences, and problems worth keeping out of a short-term intake capture, as durable memories that earn trust on recall. Marker-driven and deliberately conservative: a miss just means no extra memory, never a wrong one. Both the backfill importer and write-time extraction share this logic.
Index ¶
Constants ¶
const ClassifyMaxChars = 400
ClassifyMaxChars bounds a classifiable write: one durable fact is terse; longer content is session history even when it contains decision markers.
const MaxEntities = 12
MaxEntities caps the entities extracted from one memory, keeping the first (reading-order) mentions so one long transcript can't flood the entity index.
const MaxPerExchange = 5
MaxPerExchange caps extractions from a single exchange so one rambling turn can't flood the store.
const MinConfidence = 0.3
MinConfidence gates an extraction: confidence is min(1, score/5) where score is the count of distinct markers that match plus a length bonus, so a lone marker on a short segment (0.2) is dropped while two distinct markers, or one in a longer segment, pass.
Variables ¶
This section is empty.
Functions ¶
func Entities ¶ added in v0.5.5
Entities extracts salient named-entity keys from a memory's content without an LLM: maximal runs of capitalized tokens ("Charlotte's Web", "Black Friday", "Postgres"), normalized to lowercase with possessives stripped. Deliberately precision-first — noisy entities poison every consumer — so it only trusts capitalization evidence:
- a sentence-initial span counts only when it has ≥2 tokens, its lead word is also seen capitalized mid-sentence in the same text, or the word carries its own casing signal (internal uppercase or digits, "iPhone", "Qwen3");
- stopwords (function words, greetings, months/weekdays) never join a span; they split it instead;
- purely numeric or single-letter tokens never start an entity.
Lowercase concept nouns ("pottery", "the retry policy") are out of scope by design: without a POS tagger they cannot be told from ordinary prose.
Types ¶
type Kind ¶
type Kind string
Kind labels an extracted memory; it doubles as the memory's first tag.
func Classify ¶ added in v0.5.4
Classify labels an entire write whose caller picked no tier. It is stricter than Typed because the verdict covers the full text rather than an extracted segment: the content must be short, prose-shaped, unhedged, not a transcript, and clear the same marker-confidence gate. ok=false means no confident call — callers fall back to the working-intake default.
Source Files