github assafelovic/gpt-researcher v3.8.0
New paper UI, detailed reports in 3:27, and 53 community fixes
on GitHub

32 minutes ago

A new UI, and detailed reports in under 4 minutes

GPT Researcher turning a question into a cited report

The Next.js app has a new paper design that matches gptr.dev. Settings live in one modal with tabs for Research, MCP, Documents and Backend; a run shows a plain-language feed, and the sources margin slides in once writing starts. The previous UI is still available at /classic.

Detailed reports now finish in 3:27 instead of 9:11. Sections are researched in parallel, and the default writer is now gpt-6-sol, which starts writing within seconds instead of reasoning for close to a minute per section first. Measurements and the model comparison are in #2272.

⚠️ Changes to note

  • Default writer model: SMART_LLM is now openai:gpt-6-sol (was openai:gpt-5.4). Set SMART_LLM to keep the previous model.
  • LLM idle timeout: an LLM call that sends nothing for 300 seconds now fails with a clear error instead of hanging the run. Tune it with LLM_TIMEOUT (0 disables it). (#2220)
  • Docker: the root .env is no longer copied into images. docker compose loads it at runtime (needs Compose 2.24+); with docker run, pass --env-file .env. If you ever pushed an image you built locally, rotate the keys that were in your .env.
  • generate_report now raises when every LLM attempt fails, instead of returning an empty report. (#2219)

Also in this release

  • Prebuilt Docker images for linux/amd64 and linux/arm64 on GHCR: ghcr.io/assafelovic/gpt-researcher and ghcr.io/assafelovic/gpt-researcher-frontend-nextjs, tagged with the version and latest (#1517)
  • Research observer: an optional lifecycle observer for integrations that want structured progress events (#2232)
  • MiniMax regional providers and Flash defaults (#1924), a configurable TypeSafe base URL (#2243), and Deploy on Render with a render.yaml blueprint (#2184)
  • Multi-agent plan review feedback from the web UI works again, and the multi-agent writer uses SMART_TOKEN_LIMIT and REASONING_EFFORT (#2272)
  • Deep research honors reasoning budgets, counts retries and nested researchers in costs, and keeps cited learnings at the word limit (#2260, #2209, #2208)
  • MCP search no longer blocks the event loop, and the max_tools cap is enforced (#2262, #2263)
  • Configuration: reasoning_effort from JSON config files, DOC_PATH / REPORT_SOURCE env overrides, UTF-8 config files, and image generation settings are all honored again (#2230, #2248, #2249, #2199, #2246)
  • Retriever, scraper and document loader fixes across Exa, Google, Bing, Bocha, PubMed Central, arXiv, online documents and Windows text files
  • Retriever plugins: a community list with Keenable and Sciverse (#2176, #2225)

53 community pull requests went into this release. Thank you to everyone who contributed.

Install: pip install -U gpt-researcher (0.17.0)

What's Changed

  • fix(readme): point star history chart back to official star-history.com by @assafelovic in #2204
  • docs: drop STORM references and describe current research approach by @assafelovic in #2205
  • fix(config): preserve loaded MCP settings by @imharjot in #2193
  • fix(config): stop sharing mutable defaults between Config instances by @assafelovic in #2206
  • fix(ci): cap pytest below 9.2 so --forked tests stop crashing by @assafelovic in #2269
  • docs: use a single black tile logo in the READMEs by @assafelovic in #2270
  • New paper UI, and detailed reports in 3:27 instead of 9:11 by @assafelovic in #2272
  • docs(readme): show only the demo GIF by @assafelovic in #2273
  • Release 0.17.0: 53 community fixes, Docker secrets fix, GHCR multi-arch images by @assafelovic in #2274
  • fix(report): raise error and log failure when generate_report LLM calls fail by @pei711 in #2219
  • docs: add Sciverse to community retriever plugins by @gary-Shen in #2225
  • fix(config): honor reasoning effort from JSON configuration by @yunyv in #2230
  • feat: add an optional research lifecycle observer by @EvanZhao119 in #2232
  • docs: fix links broken by the deep_agents branch split and two upstream migrations by @SoulmatelynchVFX in #2236
  • fix(deep-research): validate follow-up question shapes before iteration by @WitBlue6 in #2238
  • (feat) Custom typesafe base url by @eriktews in #2243
  • fix: image generation config read with wrong (uppercase) attribute names - feature unreachable by @Akaliup in #2246
  • fix: DOC_PATH and REPORT_SOURCE env overrides ignored by doc path setup by @Akaliup in #2248
  • fix(config): read JSON configuration files as UTF-8 by @DZH166 in #2249
  • fix(document): detect online document type from URL path by @Elon-Jiang in #2252
  • fix(scraper): select the arXiv backend by URL hostname by @DZH166 in #2254
  • fix(markdown): decode HTML entities in extracted section text by @DZH166 in #2255
  • Preserve configuration and document inputs in detailed subtopic reports by @yunyv in #2256
  • docs(mcp): fix broken links in Claude integration documentation by @pei711 in #2258
  • fix(deep-research): honor reasoning budgets and track retry costs by @SJ-Xia-Q in #2260
  • Clear stale sources after empty deep research runs by @Quinnod345 in #2261
  • fix(research): run the blocking MCP retriever search off the event loop by @DZH166 in #2262
  • fix(mcp): enforce the max_tools cap when the LLM picks MCP tools by @DZH166 in #2263
  • fix(cost): update_cost uses raw unsanitized token counts inconsistent with calculate_cost by @Cmz9111 in #2264
  • fix(scraper): get_image_hash returns None on error despite str return type annotation by @Cmz9111 in #2265
  • fix(costs): guard estimate_embedding_cost against None and non-str inputs by @pei711 in #2266
  • fix(bocha): implement get_api_key error handling and add unit tests by @pei711 in #2267
  • fix(bing): apply query_domains and harden payload parsing in BingSearch by @pei711 in #2268
  • feat: update MiniMax regional providers and Flash defaults by @octo-patch in #1924
  • fix(multi_agents): treat an empty console plan reply as approval, as the web form does by @hesam-shams in #2164
  • fix: revalidate PyMuPDFScraper retry output against content-quality checks by @MrSampson in #2167
  • fix: ReportGenerator.write_report() must read researcher.websocket live by @MrSampson in #2169
  • docs: add a community plugins list to Retriever Plugins by @ilya-bogin-keenable in #2176
  • fix(exa): harden find_similar and get_contents against client failures by @RerankerGuo in #2181
  • fix(document): harden LangChainDocumentLoader.load against malformed rows by @RerankerGuo in #2183
  • feat: add Deploy on Render page and render.yaml by @Ho1yShif in #2184
  • docs: document generic MCP connection headers and fix Quick Start syntax by @ct-jaryn in #2185
  • chore(deps): update numpy requirement from <2.5.0,>=2.0.0 to >=2.0.0,<2.6.0 by @dependabot[bot] in #2187
  • test: stop tests leaking sys.modules stubs; run CI suite without --forked by @Saadanjum0 in #2189
  • fix: resolve PubMed IDs before fetching PMC full text by @GoldArowana in #2195
  • fix(images): read image_generation_* config keys in the case Config stores them by @Winoren in #2199
  • fix(document): keep OnlineDocumentLoader url a string when metadata omits source by @Winoren in #2200
  • docs: fix broken README language and evals links by @pratikgx in #2201
  • fix(deep-research): keep cited learnings when the context reaches the word limit by @porkedchop in #2208
  • fix(deep-research): count strategic LLM calls and nested researchers in costs by @porkedchop in #2209
  • fix(scraper): treat the "Client Challenge" interstitial as a block page by @yail259 in #2213
  • fix(google): match YouTube exclusions by hostname by @sanjayrohith in #2215
  • fix: replace retired Google and Anthropic model defaults by @pcbeingused333 in #2216
  • fix: don't send temperature to Claude models that reject it by @pcbeingused333 in #2217
  • fix(llm): bound stalled responses with configurable idle timeouts by @Xiaoshaoxinxing in #2220
  • docs: sync Chinese README (README-zh_CN.md) with the English README by @yuee3 in #2222
  • fix(pubmed_central): keep words whole at inline markup and put each block on its own line by @L4XB in #2224
  • fix(scraper): retry HTTP 403 with a transparent user agent by @EvanZhao119 in #2231
  • fix(document): load UTF-8 .txt and .html documents on Windows by @tanishqbaweja in #2237
  • fix(prompts): preserve config in invalid prompt family fallback by @jixuantang in #2245
  • fix(vector-store): skip empty chunk batches by @Elon-Jiang in #2250

New Contributors

Full Changelog: v3.7.0...v3.8.0

Check out latest releases or
releases around assafelovic/gpt-researcher v3.8.0

Don't miss a new gpt-researcher release

NewReleases is sending notifications on new releases.

Get notifications