RAG-powered Conversational Agent

This endpoint offers two operating modes:

• RAG mode: Standard retrieval-augmented generation for answering questions based on knowledge base
• Research mode: Advanced capabilities for deep analysis, reasoning, and computation

RAG Mode (Default)

The RAG mode provides fast, knowledge-based responses using:

• Semantic and hybrid search capabilities
• Document-level and chunk-level content retrieval
• Optional web search integration
• Source citation and evidence-based responses

Research Mode

The Research mode builds on RAG capabilities and adds:

• A dedicated reasoning system for complex problem-solving
• Critique capabilities to identify potential biases or logical fallacies
• Python execution for computational analysis
• Multi-step reasoning for deeper exploration of topics

Available Tools

RAG Tools:

• search_file_knowledge: Semantic/hybrid search on your ingested documents
• search_file_descriptions: Search over file-level metadata
• content: Fetch entire documents or chunk structures
• web_search: Query external search APIs for up-to-date information
• web_scrape: Scrape and extract content from specific web pages

Research Tools:

• rag: Leverage the underlying RAG agent for information retrieval
• reasoning: Call a dedicated model for complex analytical thinking
• critique: Analyze conversation history to identify flaws and biases
• python_executor: Execute Python code for complex calculations and analysis

Streaming Output

When streaming is enabled, the agent produces different event types:

• thinking: Shows the model’s step-by-step reasoning (when extended_thinking=true)
• tool_cal: Shows when the agent invokes a tool
• tool_result: Shows the result of a tool call
• citation: Indicates when a citation is added to the response
• message: Streams partial tokens of the response
• final_answer: Contains the complete generated answer and structured citations

Conversations

Maintain context across multiple turns by including conversation_id in each request. After your first call, store the returned conversation_id and include it in subsequent calls. If no conversation name has already been set for the conversation, the system will automatically assign one.