[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"similar-Agent-RL--ReCall":3,"tool-Agent-RL--ReCall":64},[4,17,27,35,43,56],{"id":5,"name":6,"github_repo":7,"description_zh":8,"stars":9,"difficulty_score":10,"last_commit_at":11,"category_tags":12,"status":16},3808,"stable-diffusion-webui","AUTOMATIC1111\u002Fstable-diffusion-webui","stable-diffusion-webui 是一个基于 Gradio 构建的网页版操作界面，旨在让用户能够轻松地在本地运行和使用强大的 Stable Diffusion 图像生成模型。它解决了原始模型依赖命令行、操作门槛高且功能分散的痛点，将复杂的 AI 绘图流程整合进一个直观易用的图形化平台。\n\n无论是希望快速上手的普通创作者、需要精细控制画面细节的设计师，还是想要深入探索模型潜力的开发者与研究人员，都能从中获益。其核心亮点在于极高的功能丰富度：不仅支持文生图、图生图、局部重绘（Inpainting）和外绘（Outpainting）等基础模式，还独创了注意力机制调整、提示词矩阵、负向提示词以及“高清修复”等高级功能。此外，它内置了 GFPGAN 和 CodeFormer 等人脸修复工具，支持多种神经网络放大算法，并允许用户通过插件系统无限扩展能力。即使是显存有限的设备，stable-diffusion-webui 也提供了相应的优化选项，让高质量的 AI 艺术创作变得触手可及。",162132,3,"2026-04-05T11:01:52",[13,14,15],"开发框架","图像","Agent","ready",{"id":18,"name":19,"github_repo":20,"description_zh":21,"stars":22,"difficulty_score":23,"last_commit_at":24,"category_tags":25,"status":16},1381,"everything-claude-code","affaan-m\u002Feverything-claude-code","everything-claude-code 是一套专为 AI 编程助手（如 Claude Code、Codex、Cursor 等）打造的高性能优化系统。它不仅仅是一组配置文件，而是一个经过长期实战打磨的完整框架，旨在解决 AI 代理在实际开发中面临的效率低下、记忆丢失、安全隐患及缺乏持续学习能力等核心痛点。\n\n通过引入技能模块化、直觉增强、记忆持久化机制以及内置的安全扫描功能，everything-claude-code 能显著提升 AI 在复杂任务中的表现，帮助开发者构建更稳定、更智能的生产级 AI 代理。其独特的“研究优先”开发理念和针对 Token 消耗的优化策略，使得模型响应更快、成本更低，同时有效防御潜在的攻击向量。\n\n这套工具特别适合软件开发者、AI 研究人员以及希望深度定制 AI 工作流的技术团队使用。无论您是在构建大型代码库，还是需要 AI 协助进行安全审计与自动化测试，everything-claude-code 都能提供强大的底层支持。作为一个曾荣获 Anthropic 黑客大奖的开源项目，它融合了多语言支持与丰富的实战钩子（hooks），让 AI 真正成长为懂上",138956,2,"2026-04-05T11:33:21",[13,15,26],"语言模型",{"id":28,"name":29,"github_repo":30,"description_zh":31,"stars":32,"difficulty_score":23,"last_commit_at":33,"category_tags":34,"status":16},2271,"ComfyUI","Comfy-Org\u002FComfyUI","ComfyUI 是一款功能强大且高度模块化的视觉 AI 引擎，专为设计和执行复杂的 Stable Diffusion 图像生成流程而打造。它摒弃了传统的代码编写模式，采用直观的节点式流程图界面，让用户通过连接不同的功能模块即可构建个性化的生成管线。\n\n这一设计巧妙解决了高级 AI 绘图工作流配置复杂、灵活性不足的痛点。用户无需具备编程背景，也能自由组合模型、调整参数并实时预览效果，轻松实现从基础文生图到多步骤高清修复等各类复杂任务。ComfyUI 拥有极佳的兼容性，不仅支持 Windows、macOS 和 Linux 全平台，还广泛适配 NVIDIA、AMD、Intel 及苹果 Silicon 等多种硬件架构，并率先支持 SDXL、Flux、SD3 等前沿模型。\n\n无论是希望深入探索算法潜力的研究人员和开发者，还是追求极致创作自由度的设计师与资深 AI 绘画爱好者，ComfyUI 都能提供强大的支持。其独特的模块化架构允许社区不断扩展新功能，使其成为当前最灵活、生态最丰富的开源扩散模型工具之一，帮助用户将创意高效转化为现实。",107662,"2026-04-03T11:11:01",[13,14,15],{"id":36,"name":37,"github_repo":38,"description_zh":39,"stars":40,"difficulty_score":23,"last_commit_at":41,"category_tags":42,"status":16},3704,"NextChat","ChatGPTNextWeb\u002FNextChat","NextChat 是一款轻量且极速的 AI 助手，旨在为用户提供流畅、跨平台的大模型交互体验。它完美解决了用户在多设备间切换时难以保持对话连续性，以及面对众多 AI 模型不知如何统一管理的痛点。无论是日常办公、学习辅助还是创意激发，NextChat 都能让用户随时随地通过网页、iOS、Android、Windows、MacOS 或 Linux 端无缝接入智能服务。\n\n这款工具非常适合普通用户、学生、职场人士以及需要私有化部署的企业团队使用。对于开发者而言，它也提供了便捷的自托管方案，支持一键部署到 Vercel 或 Zeabur 等平台。\n\nNextChat 的核心亮点在于其广泛的模型兼容性，原生支持 Claude、DeepSeek、GPT-4 及 Gemini Pro 等主流大模型，让用户在一个界面即可自由切换不同 AI 能力。此外，它还率先支持 MCP（Model Context Protocol）协议，增强了上下文处理能力。针对企业用户，NextChat 提供专业版解决方案，具备品牌定制、细粒度权限控制、内部知识库整合及安全审计等功能，满足公司对数据隐私和个性化管理的高标准要求。",87618,"2026-04-05T07:20:52",[13,26],{"id":44,"name":45,"github_repo":46,"description_zh":47,"stars":48,"difficulty_score":23,"last_commit_at":49,"category_tags":50,"status":16},2268,"ML-For-Beginners","microsoft\u002FML-For-Beginners","ML-For-Beginners 是由微软推出的一套系统化机器学习入门课程，旨在帮助零基础用户轻松掌握经典机器学习知识。这套课程将学习路径规划为 12 周，包含 26 节精炼课程和 52 道配套测验，内容涵盖从基础概念到实际应用的完整流程，有效解决了初学者面对庞大知识体系时无从下手、缺乏结构化指导的痛点。\n\n无论是希望转型的开发者、需要补充算法背景的研究人员，还是对人工智能充满好奇的普通爱好者，都能从中受益。课程不仅提供了清晰的理论讲解，还强调动手实践，让用户在循序渐进中建立扎实的技能基础。其独特的亮点在于强大的多语言支持，通过自动化机制提供了包括简体中文在内的 50 多种语言版本，极大地降低了全球不同背景用户的学习门槛。此外，项目采用开源协作模式，社区活跃且内容持续更新，确保学习者能获取前沿且准确的技术资讯。如果你正寻找一条清晰、友好且专业的机器学习入门之路，ML-For-Beginners 将是理想的起点。",84991,"2026-04-05T10:45:23",[14,51,52,53,15,54,26,13,55],"数据工具","视频","插件","其他","音频",{"id":57,"name":58,"github_repo":59,"description_zh":60,"stars":61,"difficulty_score":10,"last_commit_at":62,"category_tags":63,"status":16},3128,"ragflow","infiniflow\u002Fragflow","RAGFlow 是一款领先的开源检索增强生成（RAG）引擎，旨在为大语言模型构建更精准、可靠的上下文层。它巧妙地将前沿的 RAG 技术与智能体（Agent）能力相结合，不仅支持从各类文档中高效提取知识，还能让模型基于这些知识进行逻辑推理和任务执行。\n\n在大模型应用中，幻觉问题和知识滞后是常见痛点。RAGFlow 通过深度解析复杂文档结构（如表格、图表及混合排版），显著提升了信息检索的准确度，从而有效减少模型“胡编乱造”的现象，确保回答既有据可依又具备时效性。其内置的智能体机制更进一步，使系统不仅能回答问题，还能自主规划步骤解决复杂问题。\n\n这款工具特别适合开发者、企业技术团队以及 AI 研究人员使用。无论是希望快速搭建私有知识库问答系统，还是致力于探索大模型在垂直领域落地的创新者，都能从中受益。RAGFlow 提供了可视化的工作流编排界面和灵活的 API 接口，既降低了非算法背景用户的上手门槛，也满足了专业开发者对系统深度定制的需求。作为基于 Apache 2.0 协议开源的项目，它正成为连接通用大模型与行业专有知识之间的重要桥梁。",77062,"2026-04-04T04:44:48",[15,14,13,26,54],{"id":65,"github_repo":66,"name":67,"description_en":68,"description_zh":69,"ai_summary_zh":69,"readme_en":70,"readme_zh":71,"quickstart_zh":72,"use_case_zh":73,"hero_image_url":74,"owner_login":75,"owner_name":75,"owner_avatar_url":76,"owner_bio":77,"owner_company":77,"owner_location":77,"owner_email":77,"owner_twitter":77,"owner_website":77,"owner_url":78,"languages":79,"stars":88,"forks":89,"last_commit_at":90,"license":91,"difficulty_score":92,"env_os":93,"env_gpu":94,"env_ram":95,"env_deps":96,"category_tags":107,"github_topics":108,"view_count":23,"oss_zip_url":77,"oss_zip_packed_at":77,"status":16,"created_at":114,"updated_at":115,"faqs":116,"releases":147},1332,"Agent-RL\u002FReCall","ReCall","ReSearch: Learning to Reason with Search for LLMs via Reinforcement Learning & ReCall: Learning to Reason with Tool Call for LLMs via Reinforcement Learning","ReCall 是一套开源框架，专门用强化学习让大语言模型学会“自己调用工具”。它不需要任何人工标注的工具使用示例，就能让模型像 OpenAI o3 一样，灵活组合搜索、计算器、数据库等任意工具，完成复杂的多步推理任务。相比前身 ReSearch（仅支持搜索），ReCall 把能力扩展到所有用户自定义工具，并配套提供合成数据生成方案，帮助模型在多样化场景里练出高阶工具链思维。  \n如果你正在做 AI Agent、RAG 或需要让大模型“动手”解决实际问题，ReCall 提供了可直接运行的训练脚本、预训练模型和一键安装指南，非常适合研究人员和开发者快速上手实验。","\u003Cdiv align=\"center\">\n\n# ***ReCall***: Learning to ***Re***ason with Tool ***Call*** for LLMs via Reinforcement Learning\n\n[![Notion](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fblog-black?style=for-the-badge&logo=notion)](https:\u002F\u002Fattractive-almandine-935.notion.site\u002FReCall-Learning-to-Reason-with-Tool-Call-for-LLMs-via-Reinforcement-Learning-1d7aec91e9bb8006ad40f9edbfe2191a) [![Arxiv](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fpaper-A82F27?style=for-the-badge&logo=arxiv)](https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.19470) [![Model](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fmodel-4169E1?style=for-the-badge&logo=huggingface)](https:\u002F\u002Fhuggingface.co\u002Fcollections\u002Fagentrl\u002Fresearch-67e506a0311bea06dc54878b) \n\n\u003C\u002Fdiv>\n\nWe introduce ***ReCall***, a novel framework that trains LLMs to ***Re***ason with Tool ***Call*** via reinforcement learning—without requiring any supervised data on tool use trajectories or reasoning steps. *ReCall* empowers LLMs to agentically use and combine arbitrary tools like [OpenAI o3](https:\u002F\u002Fopenai.com\u002Findex\u002Fintroducing-o3-and-o4-mini\u002F), offering an accessible approach toward general-purpose agents. Additionally, we provide a novel perspective to generate synthetic data with diverse environments and complex multi-step tasks, enabling LLMs to develop sophisticated tool-based reasoning capabilities. This is a work in progress and we are actively working on it.\n\n> [!IMPORTANT]\n> *ReCall* is the successor to [*ReSearch*](https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.19470) and represents a more comprehensive framework that extends beyond the search tool to support reasoning with any user-defined tools. It can be a drop-in replacement of *ReSearch*. We've archived the original implementation of *ReSearch* in the branch `re-search`.\n\n\u003Cp align=\"center\">\n\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FAgent-RL_ReCall_readme_2c6974ddf912.png\" width=\"90%\" alt=\"Overview\" \u002F>\n\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FAgent-RL_ReCall_readme_bd3ed2675151.png\" width=\"90%\" alt=\"Eval\" \u002F>\n\u003C\u002Fp>\n\n## 📰 News\n- **[2025-04-24]** 🎉 We release the first version of *ReCall*, and archive the original implementation of *ReSearch*.\n  - ➡️ The name of the repository is changed from *ReSearch* to *ReCall*.\n  - 📝 We release a [blog](https:\u002F\u002Fattractive-almandine-935.notion.site\u002FReCall-Learning-to-Reason-with-Tool-Call-for-LLMs-via-Reinforcement-Learning-1d7aec91e9bb8006ad40f9edbfe2191a) to introduce the idea of *ReCall*.\n  - 📦 Current implementation of *ReCall* is based on verl 0.3.0 + vllm 0.8.4.\n- **[2025-03-27]** 🤗 We release our trained *ReSearch* models on [Hugging Face](https:\u002F\u002Fhuggingface.co\u002Fcollections\u002Fagentrl\u002Fresearch-67e506a0311bea06dc54878b), please check it out! \n- **[2025-03-26]** 🎉 We release the paper and update the code of *ReSearch*.\n  - 📝 The **paper is released** on arXiv, more details and evaluation results can be found in our [paper](https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.19470).\n  - 🛠️ The **repository is updated** with the new implementation, especially the rollout with search during RL training. This version of implementation is based on the latest release of verl.\n- **[2025-03-03]** ✅ We have released the preview version of *ReSearch* implementation.\n\n## 📦 Installation\n\nWe recommend using conda to manage the environment. First create a conda environment and activate it.\n```bash\nconda create -n re-call python==3.10\nconda activate re-call\n```\nThen install dependencies, and the packages under ```src\u002F``` will be installed in the editable mode.  Check out ```setup.py``` for details.\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FAgent-RL\u002FReCall.git\ncd ReCall\npip3 install -e .\npip3 install flash-attn --no-build-isolation\n```\nIf you want to host a Wikipedia RAG system based on FlashRAG, you need to install faiss-gpu as follow. As described in the [FlashRAG](https:\u002F\u002Fgithub.com\u002FRUC-NLPIR\u002FFlashRAG?tab=readme-ov-file#wrench-installation), due to the incompatibility when installing faiss using pip, we need to use the following conda command to install faiss-gpu.\n```bash\nconda install -c pytorch -c nvidia faiss-gpu=1.8.0\n```\n\n## 🚀 Quick Start\n\n> If you want to learn the details of current version of *ReCall*, please refer to the [blog](https:\u002F\u002Fattractive-almandine-935.notion.site\u002FReCall-Learning-to-Reason-with-Tool-Call-for-LLMs-via-Reinforcement-Learning-1d7aec91e9bb8006ad40f9edbfe2191a) first.\n\n### Data Preparation\n\n*ReCall* is trained on a mixture of our synthetic dataset `SynTool` and the training set of `MuSiQue`. You can download the preprocessed training data from [here](https:\u002F\u002Fhuggingface.co\u002Fdatasets\u002Fagentrl\u002FReCall-data), and use such data directly for training. For preparing your own data with specific tools, you can refer to the `data\u002Fprepare_musique_recall.py`, where we provide the script of preparing the data for MuSiQue with Wikipedia search tool.\n\n### Sandbox Serving\n\nSince tools are implemented in executable Python code, the tool executor is responsible for running the Python code. To ensure safety and security, we implement a sandbox for running Python code on a remote server. To launch the sandbox service, run the following command:\n```bash\ncd scripts\u002Fserving\npython sandbox.py --port {port}\n```\nNote: The current implementation is a basic sandbox environment. We plan to use a more robust and secure sandbox in future updates. We recommend hosting the sandbox on a remote server, as local hosting may expose your machine to potential security risks.\n\n### Retriever Serving\n\nFor training on MuSiQue data with a Wikipedia search tool, we provide a Wikipedia retriever service implemented using FlashRAG and FastAPI. Before starting the retriever serving, you need download the [pre-indexed wikipedia](https:\u002F\u002Fgithub.com\u002FRUC-NLPIR\u002FFlashRAG?tab=readme-ov-file#index), [wikipedia corpus and corresponding retriever models](https:\u002F\u002Fgithub.com\u002FRUC-NLPIR\u002FFlashRAG\u002Fblob\u002Fmain\u002Fdocs\u002Foriginal_docs\u002Freproduce_experiment.md#preliminary). More details can be found in the documentation of FlashRAG.\n\nFor starting the retriever serving, you need to first fill the `scripts\u002Fserving\u002Fretriever_config.yaml` with the correct path to the retrieval model, index, and corpus, and available GPU ids. Then, you can run the following command to start the retriever serving:\n```bash\ncd scripts\u002Fserving\npython retriever_serving.py \\\n    --config retriever_config.yaml \\\n    --num_retriever {num_retriever} \\  \n    --port {port}\n```\n\n### Training\n\nOur training framework is based on [verl](https:\u002F\u002Fgithub.com\u002Fvolcengine\u002Fverl), a powerful reinforcement learning framework for LLMs. We deeply customize the verl code to fit our needs, and the modified version of verl is under the `src\u002Fverl` directory. The example of training scripts are under `scripts\u002Ftrain`.\n\n#### Single-node training\nHere is an example of training Qwen2.5-7B-Instruct with 4 GPUs locally. Note that the training script below **is just an example** for single-node training, using small batch size for quick start, and do not assure the training performance.\n```bash\ncd scripts\u002Ftrain\nbash train.sh \\\n    --train_batch_size 8 \\\n    --ppo_mini_batch_size 4 \\\n    --use_re_call True \\\n    --prompt_template_name re_call_template_sys \\\n    --actor_model_path {model\u002Fpath\u002Fto\u002Fqwen2.5-7b-instruct} \\\n    --search_url {your-hosted-retriever-url} \\\n    --sandbox_url {your-hosted-sandbox-url} \\\n    --project_name {wandb-project-name} \\\n    --experiment_name {wandb-experiment-name} \\\n    --nnodes 1 \\\n    --n_gpus_per_node 4 \\\n    --save_freq 5 \\\n    --test_freq 5 \\\n    --total_epochs 2 \\\n    --wandb_api_key {your-wandb-api-key} \\\n    --save_path {path\u002Fto\u002Fsave} \\\n    --train_files \"['train1.parquet', 'train2.parquet']\" \\\n    --test_files \"['test1.parquet', 'test2.parquet']\"\n```\n\n#### Multi-node training\n\nIf you want to **fully reproduce** *ReCall*, please refer to the multi-node training script in `scripts\u002Ftrain\u002Ftrain_multi_node.sh`.\n\n### Inference\nThis section demonstrates how to perform inference using the trained *ReCall* model. We provide a standard wrapper class in `src\u002Fre_call\u002Finference\u002Fre_call.py` that simplifies the inference process. To get started, you only need to provide the model URL and sandbox URL, then use the `run` function to execute inference. The `ReCall` class handles all the orchestration between model generation and tool execution internally. For a practical example of using the `ReCall` class, please refer to our sample implementation at `scripts\u002Finference\u002Fre_call_use_case.py`.\n \nFor model serving, we recommend using [SGLang](https:\u002F\u002Fdocs.sglang.ai\u002F). You can either download our open-source models or train your own models to conduct the inference. Here is an example of how to launch the model service:\n```bash\npython3 -m sglang.launch_server \\\n        --served-model-name {trained\u002Fmodel\u002Fname} \\\n        --model-path {trained\u002Fmodel\u002Fpath} \\\n        --tp 2 \\\n        --context-length 8192 \\\n        --enable-metrics \\\n        --dtype bfloat16 \\\n        --host 0.0.0.0 \\\n        --port 80 \\\n        --trust-remote-code \\\n        --disable-overlap \\\n        --disable-radix-cache\n```\n\n### Evaluation\n\n#### Multi-hop QA\n\nFor the evaluation on multi-hop QA, we use [FlashRAG](https:\u002F\u002Fgithub.com\u002FRUC-NLPIR\u002FFlashRAG) as the standard evaluation environment. For downloading the evaluation data, please run the following command:\n```bash\ncd data\nbash download_dataset.sh\n```\nHere is an example of evaluating the performance of ReCall-Qwen-7B-Instruct on Bamboogle test set.\n```bash\ncd scripts\u002Fevaluation\npython run_eval.py \\\n    --config_path eval_config.yaml \\\n    --method_name re-call \\\n    --data_dir {root\u002Fpath\u002Fto\u002Fevaluation\u002Fdata} \\\n    --dataset_name bamboogle \\\n    --split test \\\n    --save_dir {your-save-dir} \\\n    --save_note re-call_qwen7b_ins\n    --sgl_remote_url {your-launched-sgl-url} \\\n    --remote_retriever_url {your-hosted-retriever-url} \\\n    --generator_model {your-local-model-path} \\\n    --sandbox_url {your-hosted-sandbox-url}\n```\nFor more details about the configuration, please refer to the `scripts\u002Fevaluation\u002Feval_config.yaml` file. \n\n#### BFCL\nWe will release the evaluation code on BFCL soon.\n\n## 🤝 Acknowledge\n\nThis training implementation is based on [verl](https:\u002F\u002Fgithub.com\u002Fvolcengine\u002Fverl) and the evaluation is based on [FlashRAG](https:\u002F\u002Fgithub.com\u002FRUC-NLPIR\u002FFlashRAG) and BFCL. The serving of sandbox and retriever is based on [FastAPI](https:\u002F\u002Fgithub.com\u002Ffastapi\u002Ffastapi). The model serving is based on [SGLang](https:\u002F\u002Fdocs.sglang.ai\u002F). *ReCall* models are trained based on [Qwen2.5](https:\u002F\u002Fqwenlm.github.io\u002Fblog\u002Fqwen2.5\u002F). We sincerely appreciate their contributions to the open-source community.\n\n## 📚 Citation\n\nIf you find this work useful, please cite it as follows:\n```bibtex\n@misc{chen2025research\n  title={ReSearch: Learning to Reason with Search for LLMs via Reinforcement Learning}, \n  author={Mingyang Chen and Tianpeng Li and Haoze Sun and Yijie Zhou and Chenzheng Zhu and Haofen Wang and Jeff Z. Pan and Wen Zhang and Huajun Chen and Fan Yang and Zenan Zhou and Weipeng Chen},\n  year={2025},\n  eprint={2503.19470},\n  archivePrefix={arXiv},\n  primaryClass={cs.AI},\n  url={https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.19470}, \n}\n```","\u003Cdiv align=\"center\">\n\n# ***ReCall***：通过强化学习训练大语言模型进行工具调用推理\n\n[![Notion](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fblog-black?style=for-the-badge&logo=notion)](https:\u002F\u002Fattractive-almandine-935.notion.site\u002FReCall-Learning-to-Reason-with-Tool-Call-for-LLMs-via-Reinforcement-Learning-1d7aec91e9bb8006ad40f9edbfe2191a) [![Arxiv](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fpaper-A82F27?style=for-the-badge&logo=arxiv)](https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.19470) [![Model](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002Fmodel-4169E1?style=for-the-badge&logo=huggingface)](https:\u002F\u002Fhuggingface.co\u002Fcollections\u002Fagentrl\u002Fresearch-67e506a0311bea06dc54878b) \n\n\u003C\u002Fdiv>\n\n我们提出了***ReCall***，一个新颖的框架，通过强化学习训练大语言模型进行工具调用推理——且无需任何关于工具使用轨迹或推理步骤的监督数据。*ReCall*使大语言模型能够自主地使用并组合任意工具，例如[OpenAI o3](https:\u002F\u002Fopenai.com\u002Findex\u002Fintroducing-o3-and-o4-mini\u002F)，为构建通用智能体提供了一种易于实现的方法。此外，我们还提出了一种全新的视角，用于生成包含多样化环境和复杂多步任务的合成数据，从而帮助大语言模型发展出复杂的基于工具的推理能力。这是一项正在进行中的工作，我们正积极推进其研发。\n\n> [!重要]\n> *ReCall*是[*ReSearch*](https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.19470)的后继版本，是一个更为全面的框架，不仅支持搜索工具，还能扩展到任意用户自定义工具的推理。它可直接替代*ReSearch*。我们已将*ReSearch*的原始实现存档在`re-search`分支中。\n\n\u003Cp align=\"center\">\n\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FAgent-RL_ReCall_readme_2c6974ddf912.png\" width=\"90%\" alt=\"Overview\" \u002F>\n\u003Cimg src=\"https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FAgent-RL_ReCall_readme_bd3ed2675151.png\" width=\"90%\" alt=\"Eval\" \u002F>\n\u003C\u002Fp>\n\n## 📰 新闻\n- **[2025-04-24]** 🎉 我们发布了*ReCall*的第一个版本，并存档了*ReSearch*的原始实现。\n  - ➡️ 仓库名称由*ReSearch*更改为*ReCall*。\n  - 📝 我们发布了一篇[博客](https:\u002F\u002Fattractive-almandine-935.notion.site\u002FReCall-Learning-to-Reason-with-Tool-Call-for-LLMs-via-Reinforcement-Learning-1d7aec91e9bb8006ad40f9edbfe2191a)，介绍*ReCall*的理念。\n  - 📦 当前*ReCall*的实现基于verl 0.3.0 + vllm 0.8.4。\n- **[2025-03-27]** 🤗 我们在[Hugging Face](https:\u002F\u002Fhuggingface.co\u002Fcollections\u002Fagentrl\u002Fresearch-67e506a0311bea06dc54878b)上发布了训练好的*ReSearch*模型，欢迎查看！\n- **[2025-03-26]** 🎉 我们发布了论文并更新了*ReSearch*的代码。\n  - 📝 **论文已在arXiv上发布**，更多细节和评估结果请参见我们的[论文](https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.19470)。\n  - 🛠️ **仓库已更新**，加入了新的实现，特别是在强化学习训练过程中引入了搜索功能。该版本的实现基于最新发布的verl。\n- **[2025-03-03]** ✅ 我们已发布了*ReSearch*实现的预览版。\n\n## 📦 安装\n\n我们建议使用conda来管理环境。首先创建一个conda环境并激活它。\n```bash\nconda create -n re-call python==3.10\nconda activate re-call\n```\n然后安装依赖项，其中`src\u002F`下的包将以可编辑模式安装。详细信息请参阅`setup.py`。\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FAgent-RL\u002FReCall.git\ncd ReCall\npip3 install -e .\npip3 install flash-attn --no-build-isolation\n```\n如果希望基于FlashRAG搭建一个维基百科RAG系统，则需要按如下方式安装faiss-gpu。正如[FlashRAG](https:\u002F\u002Fgithub.com\u002FRUC-NLPIR\u002FFlashRAG?tab=readme-ov-file#wrench-installation)中所述，由于使用pip安装faiss时存在兼容性问题，我们需要使用以下conda命令来安装faiss-gpu。\n```bash\nconda install -c pytorch -c nvidia faiss-gpu=1.8.0\n```\n\n## 🚀 快速入门\n\n> 如果您想了解当前版本*ReCall*的详细信息，请先参阅[博客](https:\u002F\u002Fattractive-almandine-935.notion.site\u002FReCall-Learning-to-Reason-with-Tool-Call-for-LLMs-via-Reinforcement-Learning-1d7aec91e9bb8006ad40f9edbfe2191a)。\n\n### 数据准备\n\n*ReCall*是在我们的合成数据集`SynTool`与`MuSiQue`的训练集混合的基础上训练的。您可以从[这里](https:\u002F\u002Fhuggingface.co\u002Fdatasets\u002Fagentrl\u002FReCall-data)下载预处理后的训练数据，并直接用于训练。如需准备带有特定工具的自有数据，可以参考`data\u002Fprepare_musique_recall.py`，其中提供了使用维基百科搜索工具为MuSiQue准备数据的脚本。\n\n### 沙箱服务\n\n由于工具是以可执行的Python代码实现的，因此工具执行器负责运行这些Python代码。为确保安全性和可靠性，我们实现了在远程服务器上运行Python代码的沙箱环境。要启动沙箱服务，运行以下命令：\n```bash\ncd scripts\u002Fserving\npython sandbox.py --port {port}\n```\n注意：当前实现是一个基础的沙箱环境。我们计划在未来的更新中采用更加健壮、安全的沙箱。建议将沙箱托管在远程服务器上，因为本地托管可能会使您的机器面临潜在的安全风险。\n\n### 检索器服务\n\n为了在带有维基百科搜索工具的MuSiQue数据上进行训练，我们提供了基于FlashRAG和FastAPI实现的维基百科检索器服务。在启动检索器服务之前，您需要下载[预先索引的维基百科](https:\u002F\u002Fgithub.com\u002FRUC-NLPIR\u002FFlashRAG?tab=readme-ov-file#index)、[维基百科语料库及相应的检索器模型](https:\u002F\u002Fgithub.com\u002FRUC-NLPIR\u002FFlashRAG\u002Fblob\u002Fmain\u002Fdocs\u002Foriginal_docs\u002Freproduce_experiment.md#preliminary)。更多详情请参阅FlashRAG的文档。\n\n要启动检索器服务，您需要先在`scripts\u002Fserving\u002Fretriever_config.yaml`中填写检索模型、索引和语料库的正确路径，以及可用的GPU ID。随后，您可以运行以下命令启动检索器服务：\n```bash\ncd scripts\u002Fserving\npython retriever_serving.py \\\n    --config retriever_config.yaml \\\n    --num_retriever {num_retriever} \\  \n    --port {port}\n```\n\n### 训练\n\n我们的训练框架基于 [verl](https:\u002F\u002Fgithub.com\u002Fvolcengine\u002Fverl)，这是一个用于大语言模型的强大强化学习框架。我们对 verl 的代码进行了深度定制以满足自身需求，修改后的版本位于 `src\u002Fverl` 目录下。训练脚本的示例则位于 `scripts\u002Ftrain` 目录中。\n\n#### 单节点训练\n以下是在本地使用 4 块 GPU 训练 Qwen2.5-7B-Instruct 的示例。请注意，下面的训练脚本**仅作为单节点训练的示例**，采用较小的批量大小以便快速上手，并不保证训练性能。\n```bash\ncd scripts\u002Ftrain\nbash train.sh \\\n    --train_batch_size 8 \\\n    --ppo_mini_batch_size 4 \\\n    --use_re_call True \\\n    --prompt_template_name re_call_template_sys \\\n    --actor_model_path {model\u002Fpath\u002Fto\u002Fqwen2.5-7b-instruct} \\\n    --search_url {your-hosted-retriever-url} \\\n    --sandbox_url {your-hosted-sandbox-url} \\\n    --project_name {wandb-project-name} \\\n    --experiment_name {wandb-experiment-name} \\\n    --nnodes 1 \\\n    --n_gpus_per_node 4 \\\n    --save_freq 5 \\\n    --test_freq 5 \\\n    --total_epochs 2 \\\n    --wandb_api_key {your-wandb-api-key} \\\n    --save_path {path\u002Fto\u002Fsave} \\\n    --train_files \"['train1.parquet', 'train2.parquet']\" \\\n    --test_files \"['test1.parquet', 'test2.parquet']\"\n```\n\n#### 多节点训练\n\n如果您希望**完全复现** *ReCall*，请参考 `scripts\u002Ftrain\u002Ftrain_multi_node.sh` 中的多节点训练脚本。\n\n### 推理\n本节演示如何使用训练好的 *ReCall* 模型进行推理。我们在 `src\u002Fre_call\u002Finference\u002Fre_call.py` 中提供了一个标准的封装类，以简化推理流程。开始使用时，您只需提供模型 URL 和沙盒 URL，然后调用 `run` 函数即可执行推理。`ReCall` 类会在内部完成模型生成与工具执行之间的所有协调工作。有关如何实际使用 `ReCall` 类的示例，请参阅我们的示例实现文件 `scripts\u002Finference\u002Fre_call_use_case.py`。\n\n对于模型服务，我们推荐使用 [SGLang](https:\u002F\u002Fdocs.sglang.ai\u002F)。您可以下载我们的开源模型，也可以训练自己的模型来进行推理。以下是启动模型服务的示例：\n```bash\npython3 -m sglang.launch_server \\\n        --served-model-name {trained\u002Fmodel\u002Fname} \\\n        --model-path {trained\u002Fmodel\u002Fpath} \\\n        --tp 2 \\\n        --context-length 8192 \\\n        --enable-metrics \\\n        --dtype bfloat16 \\\n        --host 0.0.0.0 \\\n        --port 80 \\\n        --trust-remote-code \\\n        --disable-overlap \\\n        --disable-radix-cache\n```\n\n### 评估\n\n#### 多跳问答\n\n对于多跳问答的评估，我们使用 [FlashRAG](https:\u002F\u002Fgithub.com\u002FRUC-NLPIR\u002FFlashRAG) 作为标准评估环境。要下载评估数据，请运行以下命令：\n```bash\ncd data\nbash download_dataset.sh\n```\n以下是评估 ReCall-Qwen-7B-Instruct 在 Bamboogle 测试集上的表现的示例。\n```bash\ncd scripts\u002Fevaluation\npython run_eval.py \\\n    --config_path eval_config.yaml \\\n    --method_name re-call \\\n    --data_dir {root\u002Fpath\u002Fto\u002Fevaluation\u002Fdata} \\\n    --dataset_name bamboogle \\\n    --split test \\\n    --save_dir {your-save-dir} \\\n    --save_note re-call_qwen7b_ins\n    --sgl_remote_url {your-launched-sgl-url} \\\n    --remote_retriever_url {your-hosted-retriever-url} \\\n    --generator_model {your-local-model-path} \\\n    --sandbox_url {your-hosted-sandbox-url}\n```\n有关配置的更多详细信息，请参阅 `scripts\u002Fevaluation\u002Feval_config.yaml` 文件。\n\n#### BFCL\n我们将很快发布 BFCL 的评估代码。\n\n## 🤝 致谢\n\n本次训练实现基于 [verl](https:\u002F\u002Fgithub.com\u002Fvolcengine\u002Fverl)，评估则基于 [FlashRAG](https:\u002F\u002Fgithub.com\u002FRUC-NLPIR\u002FFlashRAG) 和 BFCL。沙盒与检索器的服务基于 [FastAPI](https:\u002F\u002Fgithub.com\u002Ffastapi\u002Ffastapi)。模型服务则基于 [SGLang](https:\u002F\u002Fdocs.sglang.ai\u002F)。*ReCall* 模型的训练基于 [Qwen2.5](https:\u002F\u002Fqwenlm.github.io\u002Fblog\u002Fqwen2.5\u002F)。我们衷心感谢这些项目对开源社区的贡献。\n\n## 📚 引用\n\n如果您觉得这项工作有用，请按如下方式引用：\n```bibtex\n@misc{chen2025research\n  title={ReSearch: 利用强化学习让大语言模型通过搜索进行推理}, \n  author={Mingyang Chen 和 Tianpeng Li 和 Haoze Sun 和 Yijie Zhou 和 Chenzheng Zhu 和 Haofen Wang 和 Jeff Z. Pan 和 Wen Zhang 和 Huajun Chen 和 Fan Yang 和 Zenan Zhou 和 Weipeng Chen},\n  year={2025},\n  eprint={2503.19470},\n  archivePrefix={arXiv},\n  primaryClass={cs.AI},\n  url={https:\u002F\u002Farxiv.org\u002Fabs\u002F2503.19470}, \n}\n```","# ReCall 快速上手指南\n\n## 环境准备\n- Linux \u002F macOS \u002F WSL2\n- Python 3.10\n- NVIDIA GPU（建议 ≥ 4 张 A100\u002F3090）\n- CUDA ≥ 11.8\n- 网络可访问 Hugging Face（如无法直连，请配置代理或镜像）\n\n## 安装步骤\n```bash\n# 1. 创建并激活环境\nconda create -n re-call python=3.10 -y\nconda activate re-call\n\n# 2. 克隆并安装\ngit clone https:\u002F\u002Fgithub.com\u002FAgent-RL\u002FReCall.git\ncd ReCall\npip3 install -e .\npip3 install flash-attn --no-build-isolation\n\n# 3. （可选）安装 faiss-gpu，用于 Wikipedia RAG\nconda install -c pytorch -c nvidia faiss-gpu=1.8.0 -y\n```\n\n## 基本使用\n\n### 1. 下载示例数据\n```bash\n# 训练数据已预处理，直接下载即可\nhuggingface-cli download agentrl\u002FReCall-data --local-dir .\u002Fdata\n```\n\n### 2. 启动工具沙箱（本地测试）\n```bash\ncd scripts\u002Fserving\npython sandbox.py --port 5000\n```\n\n### 3. 启动 Wikipedia 检索服务（可选）\n```bash\n# 先按 FlashRAG 指引下载索引与模型\n# 编辑 retriever_config.yaml 填写路径\npython retriever_serving.py \\\n    --config retriever_config.yaml \\\n    --num_retriever 1 \\\n    --port 6000\n```\n\n### 4. 单节点 4 卡训练（示例）\n```bash\ncd scripts\u002Ftrain\nbash train.sh \\\n    --train_batch_size 8 \\\n    --ppo_mini_batch_size 4 \\\n    --use_re_call True \\\n    --prompt_template_name re_call_template_sys \\\n    --actor_model_path Qwen\u002FQwen2.5-7B-Instruct \\\n    --search_url http:\u002F\u002Flocalhost:6000 \\\n    --sandbox_url http:\u002F\u002Flocalhost:5000 \\\n    --project_name re-call-demo \\\n    --experiment_name qwen7b-4gpu \\\n    --nnodes 1 \\\n    --n_gpus_per_node 4 \\\n    --save_freq 5 \\\n    --test_freq 5 \\\n    --total_epochs 2 \\\n    --save_path .\u002Fcheckpoints\n```\n\n### 5. 推理示例\n```python\nfrom re_call.inference.re_call import ReCall\n\nrecall = ReCall(\n    model_url=\"http:\u002F\u002Flocalhost:30000\",  # SGLang 服务地址\n    sandbox_url=\"http:\u002F\u002Flocalhost:5000\"\n)\n\nanswer = recall.run(\"请用维基百科查询并总结「图灵奖」的获奖条件。\")\nprint(answer)\n```\n\n启动 SGLang 服务：\n```bash\npython3 -m sglang.launch_server \\\n    --served-model-name re-call-qwen7b \\\n    --model-path .\u002Fcheckpoints\u002Ffinal \\\n    --tp 2 \\\n    --port 30000 \\\n    --trust-remote-code\n```\n\n至此，您已完成 ReCall 的快速上手。如需多节点训练或自定义工具，请参考 `scripts\u002Ftrain\u002Ftrain_multi_node.sh` 与 `data\u002Fprepare_musique_recall.py`。","一家 30 人的跨境电商 SaaS 初创公司，正在把 AI 客服助手接入 Shopify、WooCommerce 和 TikTok Shop 的订单、库存、物流 API，让客服机器人能实时回答“我的包裹到哪了”“能不能改地址”这类复杂问题。\n\n### 没有 ReCall 时\n- 需要 3 名算法工程师花 2 周手工标注 5 000 条“工具调用轨迹”（先查订单→再查物流→再查库存），标注成本高且容易遗漏边界情况。  \n- 上线后，每当 TikTok Shop 更新 API 字段，就要重新标注、重训模型，迭代周期至少 3 天。  \n- 机器人遇到多步任务（先确认订单状态→再判断能否改地址→最后调用改地址接口）时，经常漏掉中间步骤，导致用户投诉。  \n- 不同店铺用的库存系统版本不同，传统微调模型在新系统上直接“翻车”，准确率从 92% 跌到 67%。  \n\n### 使用 ReCall 后\n- 零标注：工程师只需把 Shopify、WooCommerce、TikTok Shop 的 API 文档写成 JSON Schema，ReCall 用强化学习自动生成 20 万条合成对话，2 小时完成训练。  \n- 零重训：TikTok Shop 更新字段后，把新 Schema 替换即可，ReCall 当晚自动重跑 RL，第二天上线，无需人工干预。  \n- 零漏步：ReCall 在训练中学会“先查订单→再查物流→再查库存”的完整链路，复杂多步任务成功率从 67% 提升到 94%。  \n- 零翻车：面对不同版本的库存系统，ReCall 通过 RL 探索不同 API 组合，跨系统准确率稳定在 91% 以上。  \n\nReCall 让跨境电商客服团队用 1 名工程师、1 晚时间，就能让 AI 客服像资深运营一样熟练调用任意店铺后台，彻底告别“人工标注—微调—翻车—再标注”的循环。","https:\u002F\u002Foss.gittoolsai.com\u002Fimages\u002FAgent-RL_ReCall_bd3ed267.png","Agent-RL","https:\u002F\u002Foss.gittoolsai.com\u002Favatars\u002FAgent-RL_344e1bc0.png",null,"https:\u002F\u002Fgithub.com\u002FAgent-RL",[80,84],{"name":81,"color":82,"percentage":83},"Python","#3572A5",99.4,{"name":85,"color":86,"percentage":87},"Shell","#89e051",0.6,1359,79,"2026-04-04T21:36:54","MIT",4,"Linux","需要 NVIDIA GPU，显存 ≥24GB（示例脚本使用 4×GPU），CUDA 11.8+（与 faiss-gpu=1.8.0 兼容）","未说明",{"notes":97,"python":98,"dependencies":99},"必须使用 conda 环境；需额外部署 Sandbox 与 Retriever 两个独立服务；训练\u002F推理需先下载预训练模型及 Wikipedia 索引（数十 GB）；多机训练脚本已提供","3.10",[100,101,102,103,104,105,106],"verl 0.3.0","vllm 0.8.4","flash-attn","faiss-gpu=1.8.0","FastAPI","SGLang","FlashRAG",[13,15,26],[109,110,111,112,113],"agent","function-calling","llm","reinforcement-learning","tool-use","2026-03-27T02:49:30.150509","2026-04-06T05:15:58.665120",[117,122,127,132,137,142],{"id":118,"question_zh":119,"answer_zh":120,"source_url":121},6091,"论文中的 “number of search” 图表如何在本地复现？","该绘图脚本未放在公开仓库，但训练代码默认会把 rollout 记录保存到本地。训练完成后，可用这些记录自行统计每道题的搜索次数并绘图。数据格式与 wandb 无关，需自行写脚本分析 rollout 日志。","https:\u002F\u002Fgithub.com\u002FAgent-RL\u002FReCall\u002Fissues\u002F65",{"id":123,"question_zh":124,"answer_zh":125,"source_url":126},6086,"运行评估脚本时出现 vllm 不兼容错误，该如何解决？","该问题是由于 ReSearchPipeline 的 run_item 实现依赖 SGLang 的响应格式（包含 response['meta_info']），而 vllm 没有该字段。官方推荐两种方案：\n1. 直接使用 SGLang（最简单，无需改代码）；\n2. 如需 vllm，需手动修改代码，将 response['meta_info']['finish_reason'] 替换为 vllm 的对应字段，并确保同时兼容两种后端。\n建议优先采用方案 1，可保持与现有评估环境一致。","https:\u002F\u002Fgithub.com\u002FAgent-RL\u002FReCall\u002Fissues\u002F26",{"id":128,"question_zh":129,"answer_zh":130,"source_url":131},6087,"评估时模型输出被截断，token 长度限制如何调整？","该问题是早期代码在 active_pipeline.py 第 89 行附近的实现 bug，已在最新版本修复。直接拉取最新代码即可解决，无需手动修改配置。","https:\u002F\u002Fgithub.com\u002FAgent-RL\u002FReCall\u002Fissues\u002F13",{"id":133,"question_zh":134,"answer_zh":135,"source_url":136},6088,"项目要求的 PyTorch 与 CUDA 版本分别是什么？","README 中给出的是 CUDA 12.1 对应的 PyTorch；但 pip install -e . 会拉取 CUDA 11.8 的 torch。实测可用组合：\n- nvcc 12.4 + CUDA 12.0 驱动\n- 需额外安装 flash-attn（按 README 指引即可）\n如驱动低于 12.4，可继续使用 CUDA 11.8 的 torch，不影响功能。","https:\u002F\u002Fgithub.com\u002FAgent-RL\u002FReCall\u002Fissues\u002F69",{"id":138,"question_zh":139,"answer_zh":140,"source_url":141},6089,"训练过程中 Ray actor 意外死亡，提示 “actor is died”，如何排查？","常见原因：\n1. OOM：进程被系统 OOM killer 杀掉；\n2. 相同 search query 返回不同结果导致 watchdog 卡住（日志出现 “watchdog got stuck for 600 seconds”）。\n排查步骤：\n- 先确认内存占用，必要时减小 batch_size；\n- 若出现 watchdog 报错，参考 issue #36 和 #55，确保检索结果稳定或增加超时容错。","https:\u002F\u002Fgithub.com\u002FAgent-RL\u002FReCall\u002Fissues\u002F62",{"id":143,"question_zh":144,"answer_zh":145,"source_url":146},6090,"是否会开源训练数据的生成代码？","不会单独开源训练数据生成脚本。本项目采用强化学习范式：无需人工编写“思考”数据，仅通过 reward 信号让模型自学何时搜索、何时回答。可参考 DeepSeek-R1 的类似设定。训练所需的 HotpotQA 原始数据已提供预处理脚本 data_preprocess_hpqa.py，可直接使用。","https:\u002F\u002Fgithub.com\u002FAgent-RL\u002FReCall\u002Fissues\u002F1",[]]