extractor
diapason.memory.extractor
¶
LLM-backed extraction of durable facts from a conversation turn.
The extractor takes a single (user, assistant) exchange and asks a small
local model to distill any long-term, user-specific facts worth remembering.
It is deliberately defensive: extraction runs on a background thread far from
the request path, so any failure — a dropped Ollama connection, a timeout, a
BrokenPipeError when the client went away, or simply unparseable output —
must degrade to "no facts" rather than propagate.