jev-recall
Retrieve by relevance, not resemblance: filter an AI assistant's memories with TypeSafe's Jev.
Added to Jevfast
How it uses Jev
Given a situation and memories, Jev Recall asks a yes/no relevance question per memory. It batches or shards questions into requests within configured limits; pointer and shortlist modes rank or verify candidates.
What Jev decides
Inline memory relevance requestExample answers · not a recorded Jev response · Source ↗
Question 1 · q0
YOUR APP
INSTRUCTION
Does this candidate memory matter for the situation?
STATE
The situation is shared state. In inline mode, each candidate memory is included in its own Noul instruction; q0 is the first question in a batch.
JEV · NOUL
YesNo
App workflow
- Collect the query and memories
Pass the user's retrieval query together with the memory collection to the selected retrieval method.
- Judge relevance
Ask Jev an independent yes/no relevance question for each memory, or use a ranking-and-verification mode.
- Return relevant memories
Use the relevance judgments to select memories for the downstream response.
Why it is interesting
It addresses retrieval failures caused by relevant memories using different wording from the query, and compares multiple retrieval approaches over the same memory set.