Awesome AI Agents › AI Agent Orchestration

Agent-RL/ReCall

⭐ 1437 Python added to this list on 2025-05-02 repository created 2025-03-03

ReCall is a novel framework designed to train large language models (LLMs) to reason and use tools through reinforcement learning without requiring supervised data on tool use trajectories or reasoning steps. It enables LLMs to agentically use and combine arbitrary tools, such as OpenAI o3, to create general-purpose agents capable of sophisticated tool-based reasoning. The framework introduces a new perspective by generating synthetic data with diverse environments and complex multi-step tasks, allowing LLMs to develop advanced reasoning capabilities with tools. ReCall is the successor to the ReSearch framework, extending its capabilities beyond search tools to support reasoning with any user-defined tools, and can serve as a drop-in replacement for ReSearch. The repository provides detailed installation instructions, recommending the use of conda for environment management and specifying dependencies including flash-attn and faiss-gpu for hosting a Wikipedia retrieval-augmented generation (RAG) system. It offers a sandbox environment for safely executing Python code tools remotely, with plans for more robust security in future updates. Additionally, it includes a retriever service based on FlashRAG and FastAPI for Wikipedia search, requiring pre-indexed Wikipedia data and retriever models. Training is based on the verl reinforcement learning framework, customized for ReCall's needs. The repository includes scripts for single-node and multi-node training setups, with examples for training models like Qwen2.5-7B-Instruct using multiple GPUs. Inference is simplified through a wrapper class that manages model generation and tool execution, facilitating practical use of trained ReCall models. Overall, ReCall represents an advanced approach to enhancing LLMs with tool-using reasoning capabilities via reinforcement learning, supporting flexible tool integration, synthetic data generation for training, and secure execution environments, aimed at advancing general-purpose AI agents.

https://github.com/Agent-RL/ReCall

agentfastapiflashragfunction-callinggeneral-purpose-agentsinference-wrapperlarge-language-modelsllmllmsmulti-node-trainingmulti-step-tasksopenai-o3python-code-executionqwen2.5-7b-instructreasoningrecallreinforcement-learningresearchsandbox-environmentsynthetic-datatool-useverl-frameworkwikipedia-rag