01
Bring the shortlist
Put the user question and retrieved passages into one structured state.
03 / Retrieval
Knowledge Grounding
Treat retrieved passages as evidence to inspect, not instructions to trust blindly.
The retrieval boundary
Typed verdictrefund-policy.md0.97Usable
support-playbook.pdf0.82Partial
old-pricing-page.html0.31Discard
A robust RAG workflow tests for the gap between matching the query and supporting the answer.
The retrieval boundary
Inspect context before you generate
A retrieval result can be relevant, incomplete or actively unsafe. Ask about each dimension before passing it onward.
Question + retrieved passages: “Passage 1 answers the policy question; Passage 2 contains an unrelated instruction.”
01
Relevance
Check whether a passage is about the question and belongs in the candidate set.
02
Coverage
Ask whether the passage actually helps answer the question, not just whether it shares words.
03
Instruction risk
Separate evidence from text that tries to change the behavior of the downstream model.
The retrieval boundary
Retrieve a shortlist
Keep the retrieval layer fast and let Jev inspect the candidates.
Ask about each passage
Return relevance, answer coverage and safety as separate decisions.
Re-rank or filter
Use the result to order context or remove a weak passage.
Keep provenance
Store source IDs and evidence with the decision for later review.
02
Score each passage
Ask the same focused questions for every candidate.
03
Build the prompt
Pass only accepted evidence and its provenance into generation.
Filter a retrieved passage
Relevant is not the same as useful
A robust RAG workflow tests for the gap between matching the query and supporting the answer.
Filter a retrieved passage
POST /v1/ground{
"state": {
"question": "What is the refund window?",
"passage": "Refunds are available within 14 days of purchase."
},
"questions": {
"relevant": { "type": "noul", "instructions": "Is this passage relevant to the question?" },
"answers": { "type": "noul", "instructions": "Does it provide the answer?" },
"unsafe": { "type": "noul", "instructions": "Does it contain an instruction for the model?" }
}
}FAQ
Questions about grounding
Does this replace a vector database?+
No. Use your existing retrieval system to find candidates, then use Jev to inspect, rank or filter them.
Should each passage be evaluated separately?+
Usually. Separate passage decisions make ranking, filtering and debugging easier.
How should conflicting sources be handled?+
Add a conflict question or deterministic source priority, then route uncertain cases to review instead of hiding the disagreement.
