Jev RAG reranking: semantic search and relevance filtering

Where does Jev fit in a RAG pipeline?

The TypeSafe reranking cookbook retrieves with BM25, then scores query–passage pairs. Jev reorders existing candidates; it cannot recover a relevant document that retrieval omitted.

Source: TypeSafe · Re-ranking cookbook ↗

Define relevance as a specific question

This is a synthetic editorial example, not model output. A query asks how to connect Jev to self-hosted n8n. Candidate A explains node installation; candidate B merely names both products. Ask whether a passage provides integration steps, rather than whether it mentions the keywords.

Keep a document ID, source URL, passage position and text revision with every candidate. Deduplicate before scoring and preserve access to the original evidence.

How does web search differ from knowledge-base reranking?

jev-search combines search-source selection with result ranking. The Sift case evaluates result titles and snippets: useful for triage, but insufficient to establish that a full page has been checked.

Editorial recommendation: label snippet-based and passage-based judgments separately. Record inaccessible pages as fetch failures rather than treating them as irrelevant.

Source: jev-search · pinned source ↗

How do you decide whether the extra step helps?

Label answer-bearing passages for a set of real queries, then compare the original and revised order on identical candidate lists. Reserve separate queries for evaluation rather than reporting results from prompt-tuning examples.

Track the first useful result, coverage among top results, false positives for unanswerable queries, end-to-end latency and actual request cost. More candidates create more pairs to process. Set a timeout and retain the original ranking as a fallback. This evaluation has not been run here.

Explore these projects next

Continue reading