> ## Documentation Index > Fetch the complete documentation index at: https://docs-v1.latitude.so/llms.txt > Use this file to discover all available pages before exploring further. # ReAct (Reasoning and Acting) Prompting > Learn how to combine reasoning and acting in a thought-action loop to solve complex tasks using external tools ## What is ReAct Prompting? ReAct (Reasoning and Acting) prompting is a paradigm that enables AI models to solve complex tasks by combining natural language reasoning with external tool interactions. It mimics human problem-solving by creating a thought-action loop where the model reasons about the problem, takes actions to gather information, observes results, and iteratively refines its approach until reaching a solution. ## Why Use ReAct Prompting? * **Complex Problem Solving**: Handles multi-step tasks requiring external information * **Dynamic Information Access**: Retrieves real-time data through tool interactions * **Human-like Reasoning**: Mirrors how humans think and act to solve problems * **Iterative Improvement**: Learns from action results to refine strategies * **Tool Integration**: Seamlessly combines reasoning with external capabilities * **Transparent Process**: Shows the thinking and action steps for explainability * **Agent-like Behavior**: First step towards autonomous agent modeling ## How ReAct Prompting Works ReAct operates through a continuous thought-action loop: 1. **Thought**: The model reasons about the current state and plans next actions 2. **Action**: The model executes tools or queries to gather information 3. **Observation**: The model processes the results from actions 4. **Iteration**: The cycle repeats with updated understanding until goal completion This process requires careful prompt management, including maintaining conversation history and trimming excessive content to stay within context limits. ## Basic Implementation in Latitude Here's a simple ReAct example using Latitude's built-in tools: ```markdown Basic ReAct Research theme={null} --- provider: OpenAI model: gpt-4o temperature: 0.1 tools: - latitude/search - latitude/extract --- # Research Assistant with ReAct I will help you research {{ research_topic }} using a systematic thought-action approach. ## Instructions: Follow this ReAct pattern: 1. **Thought**: Reason about what information you need 2. **Action**: Use tools to gather that information 3. **Observation**: Analyze the results 4. **Thought**: Plan your next step based on what you learned 5. Repeat until you have comprehensive information Let me start researching {{ research_topic }}: **Thought**: I need to understand the current state and recent developments in {{ research_topic }}. Let me start with a broad search to get an overview. **Action**: I'll search for recent information about {{ research_topic }}. ``` ## Advanced ReAct Implementation For more complex tasks, create structured ReAct workflows: ```markdown Advanced ReAct Analysis theme={null} --- provider: OpenAI model: gpt-4o temperature: 0.2 tools: - latitude/search - latitude/extract - latitude/code --- # ReAct Problem Solving Framework I'm tasked with: {{ complex_task }} ## Initial Assessment: **Thought**: Let me break down this complex task and identify what information and actions I need to complete it successfully. 1. **Problem Analysis**: What are the key components of this task? 2. **Information Requirements**: What data do I need to gather? 3. **Tool Strategy**: Which tools will be most effective? 4. **Success Criteria**: How will I know when the task is complete? Let me begin the ReAct process: # ReAct Execution Cycle Previous context: {{ initial_assessment }} ## Thought-Action Loop: **Thought 1**: Based on my analysis, I need to start by {{ first_reasoning_step }} **Action 1**: [Execute first action using appropriate tools] **Observation 1**: [Process and analyze results] **Thought 2**: Given these results, my next step should be {{ next_reasoning_step }} **Action 2**: [Execute second action] **Observation 2**: [Analyze new information] Continue this pattern until task completion... # Synthesis and Conclusion Previous ReAct cycles: {{ execution_cycles }} ## Final Synthesis: **Thought**: Now I need to synthesize all the information I've gathered and provide a comprehensive response. **Final Analysis**: [Combine all observations and reasoning] **Conclusion**: [Present final results and recommendations] **Reflection**: [Evaluate the effectiveness of the ReAct process] ``` ## Domain-Specific ReAct Applications ### Market Research ReAct ```markdown Market Research ReAct theme={null} --- provider: OpenAI model: gpt-4o temperature: 0.3 tools: - latitude/search - latitude/extract - latitude/code --- # Market Research Agent Research market opportunity for: {{ product_concept }} ## ReAct Research Process: **Thought**: To assess this market opportunity, I need to gather data on market size, competition, trends, and customer needs. Let me start systematically. **Action**: Search for market size and growth data for {{ product_concept }} [Tool will execute search] **Observation**: [Analyze market size data] **Thought**: Now I need competitive intelligence. Who are the key players and what gaps exist? **Action**: Search for competitors and competitive analysis in {{ market_segment }} [Continue ReAct cycle through:] - Market trends analysis - Customer pain points research - Regulatory considerations - Technology landscape assessment **Final Market Assessment**: [Synthesized conclusion] ``` ### Technical Problem Solving ReAct ```markdown Technical ReAct Debugging theme={null} --- provider: OpenAI model: gpt-4o temperature: 0.1 tools: - latitude/search - latitude/code - latitude/extract --- # Technical Debugging Assistant Debug and solve: {{ technical_problem }} ## ReAct Debugging Process: **Thought**: I need to understand this technical issue systematically. Let me gather information about the error, check documentation, and test potential solutions. **Action**: Search for common causes and solutions for {{ error_type }} **Observation**: [Analyze search results for patterns] **Thought**: Based on these patterns, let me examine the specific technical details and run some diagnostic code. **Action**: Execute diagnostic code to analyze {{ system_component }} **Observation**: [Review diagnostic results] **Thought**: The diagnostics suggest {{ hypothesis }}. Let me verify this with additional research and testing. [Continue ReAct cycle through:] - Documentation research - Code analysis - Solution testing - Validation steps **Solution**: [Present debugged solution with reasoning] ``` ### Content Creation ReAct ```markdown Content ReAct Strategy theme={null} --- provider: OpenAI model: gpt-4o temperature: 0.4 tools: - latitude/search - latitude/extract --- # Strategic Content Creator Create content strategy for: {{ content_topic }} ## ReAct Content Development: **Thought**: To create effective content, I need to understand the audience, analyze successful content in this space, and identify unique angles. **Action**: Research trending content and successful approaches for {{ content_topic }} **Observation**: [Analyze content trends and engagement patterns] **Thought**: Now I need to understand the target audience better and identify content gaps. **Action**: Search for audience demographics and content preferences in {{ target_market }} **Observation**: [Review audience insights] **Thought**: With this audience data, let me identify unique angles and content opportunities. [Continue ReAct cycle through:] - Competitive content analysis - SEO and keyword research - Platform-specific optimization - Content format testing **Content Strategy**: [Present comprehensive strategy] ``` ## Best Practices for ReAct Prompting **Clear Pattern Establishment**: * Always label thoughts, actions, and observations explicitly * Maintain consistent format throughout the conversation * Ensure each thought logically leads to the next action * Make observations comprehensive and actionable **Reasoning Quality**: * Encourage detailed reasoning in thought phases * Connect new information to previous observations * Show how each action builds toward the goal * Maintain logical flow between iterations **Effective Tool Usage**: * Choose appropriate tools for each information need * Combine multiple tools when necessary * Use tool results to inform subsequent actions * Validate information across multiple sources **Tool Strategy**: * Start with broad searches, then narrow focus * Use extraction tools for detailed analysis * Employ code tools for calculations and data processing * Chain tool calls for complex workflows **Conversation History**: * Maintain relevant context from previous cycles * Trim excessive detail while preserving key insights * Reference previous observations in new reasoning * Build cumulative understanding over iterations **Memory Optimization**: * Summarize key findings periodically * Remove redundant information to save tokens * Prioritize recent and relevant context * Use step-based approaches for complex tasks **Validation Techniques**: * Cross-verify information from multiple sources * Test hypotheses through targeted actions * Evaluate solution effectiveness before concluding * Maintain skeptical reasoning throughout **Error Handling**: * Acknowledge when tools return unexpected results * Adjust strategy based on failed actions * Seek alternative information sources * Document limitations and assumptions ## Advanced ReAct Techniques ### Multi-Agent ReAct Coordinate multiple specialized agents in ReAct loops: ```markdown Multi-Agent ReAct theme={null} --- provider: OpenAI model: gpt-4o temperature: 0.3 tools: - latitude/search - latitude/extract - latitude/code type: agent agents: - agents/researcher - agents/analyst - agents/strategist --- # Multi-Agent ReAct Coordination Task: {{ complex_multi_faceted_task }} ## Coordinated ReAct Process: **Coordination Thought**: This task requires multiple specialized perspectives. Let me coordinate researcher, analyst, and strategist agents in a ReAct workflow. **Action**: Initiate research phase with specialist agents [Agents execute their ReAct cycles] **Observation**: Synthesize findings from all agent perspectives **Coordination Thought**: Based on multi-agent insights, determine next coordinated actions [Continue coordinated ReAct cycles] **Final Synthesis**: Integrate all agent findings into comprehensive solution ``` ### Hierarchical ReAct Structure ReAct with multiple levels of planning: ```markdown Hierarchical ReAct theme={null} --- provider: OpenAI model: gpt-4o temperature: 0.2 tools: - latitude/search - latitude/extract - latitude/code --- # Strategic ReAct Planning High-level task: {{ strategic_objective }} **Strategic Thought**: I need to break this into manageable sub-goals and plan a hierarchical approach. ## Strategic Planning: 1. **Goal Decomposition**: Break into sub-objectives 2. **Priority Setting**: Determine order of operations 3. **Resource Planning**: Identify tool and information needs 4. **Success Metrics**: Define completion criteria **Strategic Action**: Define detailed execution plan # Tactical ReAct Execution Strategic plan: {{ strategic_plan }} ## Tactical ReAct Cycles: For each sub-objective: **Tactical Thought**: [Specific reasoning for this sub-goal] **Tactical Action**: [Focused tool usage] **Tactical Observation**: [Sub-goal specific analysis] [Repeat for each tactical objective] # Operational ReAct Implementation Tactical progress: {{ tactical_results }} ## Operational Actions: **Operational Thought**: Now execute specific implementation steps **Operational Action**: [Detailed implementation] **Operational Observation**: [Immediate results] [Rapid operational cycles] **Integration**: Combine operational results with tactical and strategic levels ``` ### Self-Correcting ReAct Implement error detection and correction in ReAct loops: ```markdown Self-Correcting ReAct theme={null} --- provider: OpenAI model: gpt-4o temperature: 0.3 tools: - latitude/search - latitude/extract - latitude/code --- # Self-Correcting ReAct System ## Enhanced ReAct with Error Detection: **Thought**: [Standard reasoning] **Action**: [Tool execution] **Observation**: [Standard result analysis] **Validation Thought**: Let me check if this information seems accurate and complete **Validation Action**: Cross-reference with additional sources **Validation Observation**: [Quality assessment] **Correction Thought**: [If errors detected] I need to correct my approach because [reasoning] **Correction Action**: [Alternative approach] **Correction Observation**: [Verified results] **Meta-Thought**: Evaluate the effectiveness of my ReAct process and adjust if needed [Continue with improved approach] ``` ## Integration with Other Techniques ReAct prompting combines effectively with other approaches: * **Chain-of-Thought + ReAct**: Detailed reasoning within each thought phase * **Self-Consistency + ReAct**: Multiple ReAct cycles to verify solutions * **Step-Back + ReAct**: Establish principles before action planning * **Few-Shot + ReAct**: Provide examples of effective thought-action patterns ## Common Patterns and Templates ### The "Investigation" Pattern * Thought: Identify information gaps * Action: Gather specific data * Observation: Analyze findings * Repeat: Drill deeper or pivot based on results ### The "Problem-Solving" Pattern * Thought: Hypothesize solutions * Action: Test hypotheses with tools * Observation: Evaluate effectiveness * Iterate: Refine approach based on results ### The "Synthesis" Pattern * Thought: Plan comprehensive analysis * Action: Gather diverse information sources * Observation: Compare and contrast findings * Conclude: Synthesize insights into coherent solution ReAct prompting transforms AI from passive responders to active problem-solvers, enabling complex task completion through iterative reasoning and tool-assisted action cycles.