Screen news articles for relevance before an LLM writes the briefing
Jev gives each fetched article a relevance probability so only on-topic stories reach the expensive model that writes the briefing.
Author's figure: Author reports about 18% fewer input tokens at a 0.70 threshold with blind-panel quality mostly unchanged, and a net loss when the writing model is a mini-class model.. Reported by the builder, not measured by us.
The use case
Teams that send news search results to a frontier model for daily briefings pay for every off-topic article in the prompt. Keyword search pulls in loose matches. Jev reads each article's title and snippet and returns the probability that it is truly about the query. Articles above a set threshold go on to the writing model; the rest are dropped. The author found savings only at frontier prices and suggests a lower threshold where missing a story is costly. A person still reads the finished briefing before it is shared and spot-checks dropped articles when the threshold changes.
Example workflow
- Pull candidate articles from the Currents News API keyword search
- Send each title and description to Jev with one yes/no relevance question
- Keep articles whose relevance probability clears a threshold such as 0.70
- Pass the kept articles to Claude or another frontier model to write the briefing
- Review the briefing and sample dropped articles before changing the threshold
Tools
- Jev (TypeSafe AI)
- Currents News API
- Claude (Anthropic API)
- Python
Risks and guardrails
- A high threshold can silently drop a relevant story
- Savings disappear if the writing model is already cheap
- Snippets may not show whether the full article is relevant
Sources
Written by Agents Lookup. Credits and official pages for the tools this use case names:
- Currents News API blog post with code and benchmark: currentsapi.services/en/blog/cut-llm-token-costs-typesafe-jev-prefiltering
- Hacker News submission: news.ycombinator.com/item
- Claude (official page): claude.ai (checked 2026-11-09)