Jev Retrieval

A dual-system architecture for autonomous research that couples fast evidence processing with deliberative planning: System Two sets evidence objectives, System One locates support. On fixed-corpus QA and document ranking it reports leading performance with lower cost than comparable agents.

An audit of request independence in seven open Jev-style checkpoints, using 276 three-passage states built from 92 TREC retrieval queries. Paired interventions show request expansion flips 9.1%/8.0% of Nimble's judgments and changes up to 21.4% of selection-or-abstention actions.

1,601,157 training and 66,482 test decision questions for JevEmbed, covering choice, score and noul types across ten Parquet shards of at most 200,000 rows each.

897,035 search-agent decision rows in unified JEV format across train, validation and test, covering retrieval judgment, query understanding, evidence judgment and search actions.

Local-first knowledge search with seven selectable pipelines: BM25 plus Jev reranking is the default vector-free path, plus agentic, embedding-hybrid, taxonomy, gate and line-search modes. Reranks up to 30 candidates in one batched set of Jev Noul relevance questions and streams cited answers from a configurable model.

Graph RAG framework over Milvus vector search with no graph database required: Jev Noul judgments select graph relations and rerank full source passages alongside vector-search candidates, reaching 86.07% average Recall@5 on 1,000 MuSiQue and 1,000 2Wiki queries in the two-stage evaluation.