Skip to content

v0.26.1

Choose a tag to compare

View all tags
@Verlintas Verlintas released this 09 Sep 15:56
· 32 commits to main since this release
v0.26.1
3f15326

v0.26.1 — 搜索工具全面升级

将联网搜索从「标题+链接列表」升级为一次调用即可完成研究的核心能力：

一次搜索闭环（read_top）

  • 搜索后自动抓取正文（默认前 1 条，可 0-3 条），多数问题一次搜索即可回答，无需逐条打开网页
  • 无损提取：语义元素级抓取（正文段落/标题/列表），自动过滤作者信息栏等噪音（码龄/阅读量/收藏/红包等 20+ 模式），消除源页面重复段落；单篇预算最高 5000 字符，随 read_top 动态分配
  • 反爬容错：自动跳过被拒网站（如百度百科 403），最多尝试 8 个来源直到抓满

六引擎并发 + 结果治理

  • 新增 360 搜索，Bing/百度/Brave/DuckDuckGo/Mojeek/360 六引擎并发合并
  • 域名去重（同站最多 2 条）让结果覆盖更多来源，官方站更易浮现
  • URL 清洗：剥离 utm/spm/from 等追踪参数
  • 引擎诊断：全部失败提示网络受限；部分可用提示结果可能不完整

更聪明、不卡死

  • 工具轮次上限 8 → 12；用尽时明确指示模型基于已有信息回答
  • 工具结果预算 1500 → 6000 字符（正文完整到达模型）
  • 空结果返回可执行建议（加限定词/换表述/试英文），不再鼓励盲目重试
  • 查询结果 5 分钟缓存（按参数区分），refresh=true 可强制重搜时效内容
  • Max 模式内置搜索策略指导；web_read 失败提示换源而非重试同一 URL

SHA-256:

  • full: 24d02daa001aeeae15db270a1bb3f460f0502a39dad0b4d5166da216979804f1
  • lite: 20d8d7a8c0b221a05d5dab3299e4978d471e07e430f1d5a8d3c64987381ddc87

v0.26.1 — Search overhaul

Web search is upgraded from "title + link list" to a one-call research loop:

One-call loop (read_top)

  • Search now auto-fetches article bodies (default top 1, 0-3 configurable); most questions are answered in a single call — no per-URL web_read round-trips
  • Lossless extraction: walks semantic elements (paragraphs/headings/lists), strips author-bar boilerplate (views/favorites/follow prompts etc., 20+ noise patterns), removes duplicated paragraphs from source pages; per-article budget up to 5000 chars scaled by read_top
  • Anti-bot resilience: skips blocked sites (e.g. baike 403), tries up to 8 sources until the requested number of bodies succeed

Six engines + result governance

  • New 360 engine: Bing/Baidu/Brave/DuckDuckGo/Mojeek/360 merged concurrently
  • Domain dedup (max 2 per site) diversifies coverage, surfacing official sources
  • URL hygiene: utm/spm/from tracking params stripped
  • Engine diagnostics: all-fail = network blocked note; partial coverage warning

Smarter, no more deadlocks

  • Tool rounds 8 → 12; exhaustion message tells the model to answer from what it has
  • Tool-result budget 1500 → 6000 chars (bodies survive to the model)
  • Empty results return actionable guidance instead of encouraging blind retries
  • 5-min per-parameter cache with refresh=true bypass for time-sensitive follow-ups
  • Max-mode prompt teaches search strategy; web_read suggests switching sources instead of retrying the same URL

SHA-256:

  • full: 24d02daa001aeeae15db270a1bb3f460f0502a39dad0b4d5166da216979804f1
  • lite: 20d8d7a8c0b221a05d5dab3299e4978d471e07e430f1d5a8d3c64987381ddc87
Assets 4
Loading