ysymyth/ReAct
ReAct
ReAct este un șablon de flux de lucru agentic și un framework de prompting pentru modele de limbaj mari (LLM). Acesta implementează un tipar logic care integrează raționamentul de tip „chain-of-thought” cu execuția de instrumente externe pentru a rezolva sarcini complexe, în mai mulți pași.
Framework-ul utilizează o logică intercalată de raționament și acțiune, forțând modelul să își documenteze procesul intern de gândire înainte de a executa o acțiune. Acest ciclu de planificare și acțiune permite sistemului să interacționeze cu API-uri sau baze de date externe și să injecteze date din lumea reală înapoi în contextul modelului pentru a rafina căile de raționament.
Proiectul acoperă execuția autonomă a sarcinilor și orchestrarea agenților prin combinarea integrării observație-feedback cu logica de execuție bazată pe prompt-uri. Acest lucru asigură un ciclu continuu de gândire, acțiune și observare pentru a atinge obiectivele fără intervenție umană constantă.
Features
- Reasoning-Action Loops - Implements a core loop that alternates internal reasoning steps with external tool execution to iteratively solve tasks.
- Agent Task Orchestrators - Orchestrates the interaction between reasoning cycles and tool use to manage stateful agent workflows.
- Agentic LLM Frameworks - Provides a framework for building agents that combine chain-of-thought reasoning with external tool execution.
- LLM Tooling Integrations - Implements connectors that allow language models to access external data and execute software tools.
- Chain of Thought Implementations - Implements a logic structure that requires the model to document its reasoning before executing an action.
- Autonomous Task Execution - Enables language models to plan and execute sequences of actions to achieve goals without human intervention.
- Chain-of-Thought Prompting - Implements a structured chain-of-thought prompting methodology to guide models through step-by-step logical reasoning.
- Interleaved Logic Execution - Utilizes a prompting pattern that alternates between internal thoughts and external tool-based actions.
- Conversation Context Management - Maintains conversation coherence by appending tool observations and reasoning traces to the interaction state.
- External Tool Integrations - Connects language models to external APIs and databases to retrieve real-time information and perform actions.
- Reasoning-Action Implementations - Implements the specific logic pattern of alternating planning and acting to solve complex multi-step tasks.
- Reasoning Task Execution - Combines chain-of-thought reasoning with tool execution to solve complex, multi-step problem sets.
- Prompt State Updates - Updates the internal model context by appending the outputs of external API calls to the conversation history.
- Prompt-Driven Execution Loops - Uses structured natural language templates to orchestrate a repeating cycle of thinking, acting, and observing.
- Prompt Feedback Loops - Injects real-world tool outputs back into the model context to correct reasoning and update the internal state.
- Prompt Chaining Frameworks - Orchestrates a sequence of prompts where reasoning steps and tool outputs inform subsequent model inputs.
- Reasoning-Action Templates - Provides a structured prompting template for guiding models through an iterative cycle of planning and observation.
- Agent Action Frameworks - Synergizing reasoning and acting in language models.
- Embodied Agents - Synergizing reasoning and acting capabilities in language models.
- Task and Motion Planning - Synergizing reasoning and acting in language models.
