Agentic RAG

​

What is Agentic RAG?

Agentic RAG (Retrieval-Augmented Generation) takes traditional RAG to the next level by incorporating intelligent agent capabilities:

• Planning: Breaking complex queries into sub-questions for comprehensive answers
• Evaluation: Verifying results against source documents for accuracy
• Multi-step reasoning: Performing multiple retrieval and reasoning steps to build deeper understanding
• Self-correction: Revising approaches based on intermediate results for better outcomes

​

How Agentset Improves RAG

Traditional RAG systems simply retrieve relevant documents and generate answers based on those documents. Agentset’s Agentic RAG approach takes this further:

1. Query Planning: The agent analyzes the user query and determines what information is needed
2. Strategic Retrieval: Instead of a single retrieval step, the agent can make multiple targeted retrievals
3. Answer Synthesis: The agent combines information from multiple sources to create a comprehensive answer
4. Citation Generation: Sources are accurately cited so users can verify information
5. Self-Verification: The agent checks its answer against source documents for accuracy

​

Benefits of Agentic RAG

​

Example: Agentic RAG vs. Traditional RAG

​

Using Agentic RAG

Agentset’s agentic capabilities are built into the chat functionality. When having a conversation with your documents:

const response = await agentset.chat({
  namespaceId: namespace.id,
  messages: [
    { role: "user", content: "How did our Q1 sales in Europe compare to Q4 last year?" }
  ],
  options: {
    // Enable detailed reasoning steps in response
    includeReasoningSteps: true,
  }
});

// Access the reasoning steps
console.log(response.reasoningSteps);

​

Customizing Agentic Behavior

You can customize the agentic RAG approach when creating a namespace or during chat:

// Configure when creating a namespace
const namespace = await agentset.namespaces.create({
  name: "Financial Reports",
  description: "Contains financial reports and analysis",
  options: {
    // Configure retrieval behavior
    retrieval: {
      // Number of documents to retrieve in each step
      docsPerRetrieval: 5,
      // Maximum number of retrieval steps
      maxRetrievalSteps: 3,
    },
    // Configure reasoning behavior
    reasoning: {
      // Whether to enable multi-step reasoning
      enableMultiStep: true,
      // Whether to verify answers against sources
      verifyAnswers: true,
    }
  }
});

// Or configure during chat
const response = await agentset.chat({
  namespaceId: namespace.id,
  messages: [{ role: "user", content: "Analyze our Q1 performance" }],
  options: {
    agentic: {
      // Whether to enable agentic RAG features
      enabled: true,
      // Maximum time to spend on deep research
      maxProcessingTime: 60, // seconds
    }
  }
});