TypeSafe engineers rebuild RAG retrieval and reranking with Jev
On small corpora Jev acts as the similarity metric, judging each document pair for relevance with a probability and no vector store; on large corpora embeddings recall top-k and Jev only reranks, catching documents with zero word overlap that still answer the query.
this is pure f*cking treasure TypeSafe AI engineers developed this themselves: Jev for RAG most agents retrieve by embedding similarity and call it meaning. Jev asks the real question for every doc: is this relevant to the query, yes or no, with a probability two regimes, one rule: small corpus: Jev is the similarity metric > score every (query, doc) pair > rank by p > no embeddings, no vector store, no index to keep > fresh, no chunking to tune > it catches the doc that shares zero words with the query but answers it exactly large corpus: Jev is the reranker > embeddings pull the top k out of millions, cheap and approximate > Jev scores only those k pairs > recall comes from the dot product, precision comes from Jev and the agent never touches the ranking: > it sends a query > Jev decides what is relevant > the top results go straight into its context your agent is only as smart as what it reads stop letting a dot product decide that
This post quotes another post: x.com/gippp69/status/2103394701063737504