github assafelovic/gpt-researcher v3.7.0
Jev update and many bug fixes and improvements
on GitHub

4 hours ago

Smarter context, no embeddings required

GPT Researcher now picks the context it sends to the LLM by usefulness, not similarity.
Every passage scraped for a research question is scored by Jev
on how useful it is for answering that question, and only the passages that help are kept.

On a benchmark of 28 research tasks replayed on identical sources:

image-1790437831580

Nothing is required. With TYPESAFE_API_KEY set, Jev is used. Without one, or if a Jev call
fails, GPT Researcher falls back to local keyword (BM25) ranking, so you no longer need an
embeddings provider for standard research. Report chat uses the same filter.

⚠️ Changes to note

  • Python 3.12+ is now required.
  • Default context filter: without a TypeSafe key, the default is now keyword ranking instead of
    embeddings. It matched or beat embeddings on every measure in the benchmark. To keep the old
    behaviour, set CONTEXT_FILTER=embeddings.

Also in this release

  • Retriever plugins: ship a search provider as its own package via the gpt_researcher.retrievers entry point (#2154)
  • Fixed the frontend WebSocket opening a second connection on a new research run (#2159)
  • Multi-agent sections no longer repeat each other's content (#2158)
  • Self-hosted frontend assets, no third-party CDNs (#2082)
  • 25+ community fixes across retrievers, scrapers and deep research
  • Docs: redesigned homepage, current model hallucination leaderboard, fixed builds on fresh installs

Install: pip install -U gpt-researcher (0.16.0)

What's Changed

  • fix: define stream output for multi-agent execution by @alaningtrump in #2092
  • fix(deps): pin mcp below 2.0 to keep MCPRetriever importable by @DAI-BZ in #2147
  • fix(bing): bound request and handle transport failures by @anupamking01 in #2133
  • fix(semantic-scholar): bound request and guard response parsing by @anupamking01 in #2130
  • fix(openalex): handle malformed JSON responses by @anupamking01 in #2128
  • fix(multi-agents): load .env before tracing setup by @anupamking01 in #2126
  • fix: preserve caller-provided empty visited_urls set by @anupamking01 in #2122
  • fix: handle zero generated queries in deep research by @anupamking01 in #2111
  • fix: reject whitespace-only deep research query fields by @L4XB in #2117
  • fix(google): filter YouTube hosts case-insensitively by @anupamking01 in #2113
  • fix(retrievers): fall back after blank normalization by @remote-controlled-man in #2105
  • fix(research): preserve complementary web context by @remote-controlled-man in #2104
  • fix(research): record prefetched URLs in visited_urls by @feiiiiii5 in #2123
  • fix(research): preserve initial planning sources in report context by @guybass in #2145
  • fix(mcp): detect URL schemes case-insensitively by @anupamking01 in #2114
  • fix(chat): preserve sources from the actual search tool invocation by @beemines in #2112
  • fix(scraper): keep inline text in its line when extracting page text by @L4XB in #2149
  • fix(llm): let llm_kwargs / LLM_KWARGS override the computed temperature by @rbales79 in #2107
  • fix: correct typos in user-facing warnings, subtopic prompt, and docstring by @feizhuzheng in #2080
  • Make Serper retriever base URL configurable via SERPER_API_BASE by @starnikovoleg in #2103
  • fix(frontend): self-host browser assets by @ousamabenyounes in #2082
  • docs: fix dead retrievers link in hybrid research example by @yetuge in #2136
  • docs: add Russian README translation by @MonteNegroX in #2083
  • chore(deps): update numpy requirement from <2.3.0,>=2.0.0 to >=2.0.0,<2.5.0 by @dependabot[bot] in #2087
  • chore(deps): update deepagents requirement from >=0.6 to >=0.7.8 by @dependabot[bot] in #2085
  • chore: untrack maintainer triage notes; stop dependabot raising floors by @assafelovic in #2151
  • feat(retrievers): load third-party retrievers from entry points by @assafelovic in #2154
  • chore(deps): update numpy requirement from <2.3.0,>=2.0.0 to >=2.0.0,<2.5.0 by @dependabot[bot] in #2155
  • Bump python from 3.12-slim-bookworm to 3.14-slim-bookworm by @dependabot[bot] in #1612
  • fix(document): tolerate missing metadata source on loaded pages by @Bartok9 in #2152
  • Add quality_eval: ground-truth-free report quality evaluation by @SeanHe727 in #1821
  • fix: carry over the remaining fixes from #2140 and #2121 by @assafelovic in #2157
  • fix(multi_agents): tell each parallel section what its siblings cover by @assafelovic in #2158
  • fix(websocket): stop a new research run from opening a second socket by @assafelovic in #2159
  • ci: don't fail fork PRs on the LLM cost comment by @assafelovic in #2160
  • feat(context): filter scraped content with Jev; embeddings become optional by @assafelovic in #2161
  • feat(context): keyword fallback so no filter needs a key or embeddings by @assafelovic in #2162
  • docs: current model leaderboard; fix the docs build on fresh installs by @assafelovic in #2170
  • docs: redesign the documentation homepage by @assafelovic in #2171
  • docs(homepage): drop the code card and stats strip; centre the hero by @assafelovic in #2172
  • docs(homepage): restore the two-column hero by @assafelovic in #2173

New Contributors

Full Changelog: v3.6.1...v3.7.0

Check out latest releases or
releases around assafelovic/gpt-researcher v3.7.0

Don't miss a new gpt-researcher release

NewReleases is sending notifications on new releases.

Get notifications