Skip to main content
Traditional RAG is passive: User asks -> System searches -> LLM answers. Agentic RAG is active: Agent decides what to search, verify results, and iterate.

​
How StateBase Powers Agentic RAG

StateBase serves as the Store and Trace layer for this loop.

​
The Loop

  1. Reasoning: Agent thinks “I need to check the pricing page.”
  2. Tool Call: Agent calls search_tool("pricing").
  3. State Update: StateBase records the tool call and the raw result.
  4. Reflection: Agent analyzes result. “This is outdated 2024 pricing.”
  5. Refinement: Agent calls search_tool("pricing 2025").

​
Implementation

# 1. Decision Step
state = sb.sessions.get_state(session_id)
action = agent.decide(state)

if action == "search":
    # 2. Tool Execution
    result = search_tool.run(query)
    
    # 3. State Preservation
    sb.sessions.add_turn(
        input={"type": "tool_result", "content": result},
        output={"type": "thought", "content": "Analyzing pricing data..."}
    )
By storing the intermediate thought process (Reasoning traces) in StateBase, your agent can “learn” from failed retrieval attempts in future sessions.