Adding an example of a scientific deep research - #23
Adding an example of a scientific deep research#23Ivan Reznikov (IvanReznikov) wants to merge 2 commits into
Conversation
|
Output: |
Harrison Chase (hwchase17)
left a comment
There was a problem hiding this comment.
left some minor comments
one major comment though: why does this deserve to be its own example? It seems pretty similar to deep-research. is there something novel about this? a novel set of tools or prompting technique or subagents? is there a common benchmark to benchmark it on? I dont see any of those at the moment, so questioning whether this deserves to be its own thing a bit
|
|
||
| See [examples/scientific_research/scientific_agent.py](examples/scientific_research/scientific_agent.py) for the full implementation. | ||
|
|
||
| ## Examples |
There was a problem hiding this comment.
i like this section, but lets remove line 69 then
The overall idea is to widen the audience of deepagents. |
|
I've also added ANTI-HALLUCINATION sections in the prompt. Quite a useful thing, when APIs don't return proper results (package not installed, api is down, too many responses, overload, etc.) the LLM starts to hallucinate. |
|
hey Ivan Reznikov (@IvanReznikov) thanks for the example, we won't include this one right now, quite similar to our existing deep research one. a great one to still share with the community on socials! |
… exFAT resource forks 排除 Item A: ADR v5 langchain-ai#23 sync 主档案 - pmagent 接受副本 commit b246404 (2026-05-04) 已落地 langchain-ai#23 - 主档案 ADR-0002 v3 §0 v5 增补区同步增补 langchain-ai#23 完整 entry - langchain-ai#23: 设计文档(spec / SOP / 装配支援包 / 决策记录)必须含 verification 章节 - 4 项要素: (1) 客观判定标准 (2) 验证方法(含对照实验设计)(3) 失败可观察信号 (4) 谁负责跑 verification - 与 langchain-ai#22 区别: langchain-ai#22 防实验设计 design flaw(实验维度),langchain-ai#23 防文档设计 design flaw(spec 维度) - 沉淀来源: CTO 自 audit 3 findings (D2-A1/fork-diff-A1/H3-A1) 共同模式 + 三方专家 v1.6 review APPROVE Item B: .gitignore 加 macOS exFAT resource forks 排除 - 新增 **/._* 防御 docs/architecture/ 等目录出现 ?? docs/.../._*.md untracked noise - 与 CLAUDE.md memory "外部 volume /Volumes/0-/ creates ._* resource fork files" 一致 - 也是 CTO 第 8 次失误的 process 改进(之前 git status 含大量 ._* 噪音影响审视清晰度) 应用 ADR v5 langchain-ai#16/langchain-ai#17/langchain-ai#22/langchain-ai#23(避免重复 CTO 第 8 次失误): - langchain-ai#16: 跑 git status / git diff / git log 实证(不依赖记忆) - langchain-ai#17: fidelity 量化(git ls-files 实证 vs 内部"我已 sync" claim) - langchain-ai#22: 实证而非记忆 - langchain-ai#23: 本 commit message 本身含 verification 路径(实证命令 + 对照标准) 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
… v5 langchain-ai#22 + langchain-ai#23) CTO 自 audit 3 findings (D2-A1/fork-diff-A1/H3-A1) 修订完整闭环。本 commit 是 §6.1 A 组 3 项 CTO domain spec 修订(A1+, A4, A5)。 A1+ H-3 §3.2.5 反向测试要求(应用 langchain-ai#22) - Phase 1 装配支援包 H-3 §3.2 增"反向测试要求"section - 每项 invariant 必须配反向验证(破坏 invariant 看是否 FAIL) - 含 T-3 反向验证模板供 pmagent 复制到其他 12 项 - 触发: pmagent v1.6 audit MEDIUM langchain-ai#2 (仅 T-3 反向,T-1~T-13 单向 assertion 风险) A4 D-2 v2 §6 R-3 对照测试覆盖(应用 langchain-ai#22) - 单元测试 spec 增 3 项对照测试要求: ① 嵌套 SubAgent (subagent 调用 subagent) hook 触发对照 ② async dispatch vs sync dispatch 双路径对照 ③ 多 hook 订阅者 firing 顺序对照 - 触发: CTO 自 audit D2-A1 finding (设计未要求实施时做对照测试) A5 fork-vs-pypi-diff-evaluation.md spec(应用 langchain-ai#23) - 创建独立 spec 文档(之前仅 message 形式,不在 git tracking) - §3.3 测试驱动评估方法(替代主观"是否仍依赖"判断) - §3.2 失败 → V2 子类化决策树 - §5 完整 verification 章节(应用 ADR v5 langchain-ai#23 SOP) - 估时 2-3 h(绿灯)/ +1-3 d V2 子类化(黄灯)/ 翻盘 ADR(红灯) - 触发: CTO 自 audit fork-diff-A1 finding (缺可执行评估方法) 应用 ADR v5 全套 (langchain-ai#16-langchain-ai#23): - langchain-ai#16: 所有 finding 配具体行号引用 + commit hash - langchain-ai#17: fidelity 量化 (fork diff 数据来自 1c3a85a 实证 +3,798 行) - langchain-ai#22: 三个 spec 修订全部含对照实验要求 - langchain-ai#23: A5 创建独立 spec 文档含 verification 章节(CTO 第 8 次失误教训:spec 必须 git tracked,不依赖记忆 sync) 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…sistence Implements a pure-Python Telegram Bot API channel adapter for the talon host, following the same patterns as the existing WhatsApp channel. **Telegram channel adapter** (): - with env-var mapping ( prefix) and alias - using — no bridge subprocess needed - Long-polling via with configurable timeout and interval - MarkdownV2 message formatting with automatic plain-text fallback - for images, for other media types - typing indicator - for inbound media downloads - Operator auto-capture from first inbound message - Configurable exposure modes (default, allowlist, open) **Offset persistence** (ticket 23): - in the session directory - Atomic writes via - Loads persisted offset on startup, saves after each polling cycle - Graceful handling of missing/corrupt/invalid offset files **CLI wiring** (): - flag to enable the Telegram channel - Updated factory to accept both and - Both channels can coexist simultaneously **Config** (): - Added to - Added to for alias capture **Tests**: - 36 unit tests covering config, MarkdownV2 escaping, polling, exposure, outbound text/media, edit, typing, inbound media, status, offset persistence - 7 integration tests for factory and simultaneous channel coexistence Closes #21, #22, #23
Input:
Design a multi-step synthesis route for the complex natural product Taxol (Paclitaxel),
including retrosynthetic analysis, literature review of key reactions, and detailed
experimental procedures. Consider the challenges of stereochemistry, protecting group
strategies, and scalability. Also analyze the current state of research on Taxol
synthesis and identify potential improvements or alternative approaches.