Filter and rerank search results before they reach an AI agent
Jev scores each retrieved document or search result for usefulness and drops weak ones before the agent or LLM reads them.
Author's figure: HN commenter LewisOrton reports saving 5-10 seconds per round with no precision loss after switching the ranker to Jev (author's own production claim).. Reported by the builder, not measured by us.
The use case
Teams running RAG chatbots or agents pass every search hit into the prompt, which adds cost and distracting context. Jev scores each retrieved result for whether it helps answer the question, sorts them, and drops those under a threshold. If nothing passes, the app can search again or say it has no answer. One builder reports replacing an LLM ranker sub-agent with Jev and saving several seconds per round. Answers that reach customers still need the usual review and feedback loop.
Example workflow
- Retrieve candidates from a vector store, keyword search or web search tool
- Send the question and candidates to Jev for usefulness scores
- Sort and drop results below a threshold
- Pass the kept results to the LLM or agent
- Review answers where nothing passed or users flag a bad answer
Tools
- Jev (TypeSafe AI)
- jev-reranker (Python)
- Milvus or another vector store
- LlamaIndex
Risks and guardrails
- A strict threshold can remove the one useful document
- Long documents may need splitting before scoring
- Ranking errors are invisible without logging
Sources
Written by Agents Lookup. Credits and official pages for the tools this use case names:
- GitHub repo jev-reranker by hotchpotch: github.com/hotchpotch/jev-reranker
- Hacker News comment by LewisOrton on replacing an LLM ranker with Jev: news.ycombinator.com/item
- Hacker News comment by gbrindisi on ranking web_search results with Jev: news.ycombinator.com/item
- LlamaParse (official page): developers.llamaindex.ai/llamaparse (checked 2026-11-19)