NewZerikai Memory
Jev checks codebase questions and retrieved passages before memory evidence is used.
Added to Jevfast
NewHow it uses Jev
The retrieval pipeline asks Jev to classify and assess a user query, then separately judge retrieved passages. The query request includes the project brief, query, and index summary. The evidence request includes the query and passage IDs, source locations, names, and text. Typed Choice, Noul, and Score answers help decide whether retrieval should proceed and which evidence is relevant, usable, conflicting, or unsafe. If Jev is unavailable, both stages fail open.
What Jev decides
Query assessment (illustrative request)Example answers · not a recorded Jev response · Source ↗
Question 1 · in domain
YOUR APP
INSTRUCTION
Is the query about this indexed codebase?
STATE
Source-defined Stage A state: project brief, current query, and index summary. This example describes request fields, not a captured runtime request.
JEV · NOUL
YesNo
Question 2 · intent
YOUR APP
INSTRUCTION
Choose the query's retrieval intent.
STATE
Source-defined Stage A state: project brief, current query, and index summary. This example describes request fields, not a captured runtime request.
JEV · CHOICE
- lookup_symbol
- explain_behavior
- architecture
- howto_usage
- debug_error
- general_chat
Question 3 · specificity
YOUR APP
INSTRUCTION
How concretely is the query anchored for code retrieval?
STATE
Source-defined Stage A state: project brief, current query, and index summary. This example describes request fields, not a captured runtime request.
JEV · SCORE
0
Question 4 · answerable from code
YOUR APP
INSTRUCTION
Could repository content answer the query faithfully?
STATE
Source-defined Stage A state: project brief, current query, and index summary. This example describes request fields, not a captured runtime request.
JEV · NOUL
YesNo
Retrieved-evidence assessment (illustrative request)Example answers · not a recorded Jev response · Source ↗
Question 1 · answerable
YOUR APP
INSTRUCTION
Taken together, do the retrieved passages provide enough trustworthy evidence to answer the query?
STATE
Source-defined Stage B state: the query and retrieved passages, each represented by ID, source file, line number, name, and text. The four per-passage judgments vary with the retrieved passage IDs; this chart shows the fixed whole-set questions.
JEV · NOUL
YesNo
Question 2 · coverage
YOUR APP
INSTRUCTION
How completely do the passages cover the main parts of the query?
STATE
Source-defined Stage B state: the query and retrieved passages, each represented by ID, source file, line number, name, and text. The four per-passage judgments vary with the retrieved passage IDs; this chart shows the fixed whole-set questions.
JEV · SCORE
0
Question 3 · conflict
YOUR APP
INSTRUCTION
Do the passages materially disagree on facts relevant to the query?
STATE
Source-defined Stage B state: the query and retrieved passages, each represented by ID, source file, line number, name, and text. The four per-passage judgments vary with the retrieved passage IDs; this chart shows the fixed whole-set questions.
JEV · NOUL
YesNo
App workflow
- Assess the query
Send the brief, query, and index summary with four typed questions; compose the answers into a proceed or no-answer decision.
- Retrieve passages
Use the existing retrieval pipeline to collect candidate evidence for the query.
- Assess evidence
Send trimmed passage records to Jev, judge each passage and the set as a whole, then keep and rank qualifying evidence.
- Continue safely
If Jev is unavailable, continue with the pipeline's fail-open behavior.
Why it is interesting
Jev is used at two distinct points: first to assess the question, then to evaluate the retrieved evidence. The second stage checks each passage for relevance, evidential value, contradiction, and injection, then ranks kept passages using those judgments and vector distance as a tie-breaker.