← @bgSponge

Jev changes a lot in memory & context engineering. Here's exactly how.

Summary

Dhravya Shah shows Jev is a system 1 model for fast decisions and structured outputs that outperforms BM25 in reranking and sets a new state-of-the-art for multi-lingual chunking.

  1. Jev prioritizes decisions over generation.

    The model functions as a system 1 tool designed for rapid decision-making and structured probabilities rather than producing fluid text.

    it's really good at taking decisions really fast, and can output structured choices and probabilities. Not text generation.

  2. Jev improves reranking quality.

    Across SciFact, NFCorpus, and TREC-COVID, Jev beats base BM25 search by 0.05 to 0.17 nDCG@10.

    it beat base BM25 search every time

  3. Jev dominates multi-lingual chunking.

    The model achieves state-of-the-art performance on messy data where semantic boundaries are poorly defined across different languages.

    Jev is pretty clearly the SOTA at chunking

  4. Filtering with Jev cuts costs.

    Using Jev as a pre-observer filter reduces content tokens by 58% on internal benchmarks.

    Jev was able to save us 58% of content tokens!!

  5. Jev fails as a compactor.

    Removing sentences independently damages contextuality and semantic meaning, making it unreliable for memory compaction.

    Jev is not a very good compactor for memory!

Highlights

  • Basically a cheap filter before things go into the observer model.

  • Jev wins quality on the 3 overlapping sets (mean 0.612 vs 0.564) and costs ~$0.00037/q vs ~$0.00002/q

  • With Jev, we can ask whether each sentence continues the previous thought or starts a new one

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →