GPT Researcher now runs on Jev by default
Assaf Elovic says the retrieval path no longer needs embeddings at all, on an eval his own team ran.
GPT Researcher has made Jev its default retrieval path and dropped embeddings from the pipeline entirely. Assaf Elovic posted the numbers behind the switch, having tested both paths on 28 research tasks drawn from SimpleQA and open ended research. He reports 73% relevant context with Jev against 46% with embeddings, reports preferred 15 to 3 in blind comparisons, and the same cost per report.
GPT Researcher now runs on Jev by default, and no longer needs embeddings at all.
This is the follow through on the swap he earlier described as an experiment. The change now is that the embeddings path is gone rather than optional, and that Elovic is describing the combination of LangChain, Tavily and Jev as the whole retrieval stack. The repo and the write up are both linked from his post.
What the number actually is
The 73% is one team's measurement of their own pipeline, scored on their own definition of relevant context, on 28 tasks. Nobody outside GPT Researcher has reproduced it in public. The 15 to 3 preference figure is a blind comparison of generated reports, which is a softer signal than it looks, since it comes from the same evaluation exercise rather than an independent panel. Cost per report being unchanged is the more load bearing claim for anyone deciding whether to try this, because it means the trade is quality against engineering effort rather than quality against spend.
Harrison Chase of LangChain picked the result up and used it to make a broader argument about where decision models belong in an agent harness. Diogo Almeida amplified it with two words.
For people building on it, note what the comparison does not cover. A vector index is cheap to query at scale and the eval here is 28 tasks, not a production corpus under load, so the per document decision call pattern that works at this size is the exact pattern that has drawn criticism elsewhere for batch sized workloads. If you are retrieving over millions of chunks rather than a research session's worth, this result does not tell you what happens.
retrieval is a decision problem, not just a similarity problem
Know what Jev is good at before you bet on it
One email a day. Under three minutes. What builders found out about Jev yesterday.

