GPT Researcher Replaces Embeddings with Jev for 73% Relevant Context

The development team of GPT Researcher has migrated its retrieval-augmented generation system to use a decision model called Jev instead of vector embeddings. In tests involving 28 research tasks from SimpleQA and other open-ended queries, the tool increased the amount of relevant context retrieved from 46% to 73% and saw its reports preferred 15 to 3 over previous methods in blind comparisons. The system now utilizes Jev by default and no longer requires embeddings, while maintaining the same cost per report. Further findings from developer Assaf Elovic indicate that raising Jev's average cost limit from $0.115 to $0.122 increases retrieval quality to 100% and reduces context tokens by 90%. Other engineers have reported similar gains, including a 12% performance increase in internal RAG evaluations when utilizing Jev as a reranker.