Agentic PageRAG: A ReAct Agent with LangGraph

pip install -U langgraph langchain-ollama langchain-core
ollama pull qwen3

from typing_extensions import TypedDict, Annotated
import operator
import os

from langgraph.graph import StateGraph, START, END
from langgraph.prebuilt import ToolNode
from langchain_ollama import ChatOllama
from langchain_core.messages import HumanMessage, SystemMessage
from IPython.display import display, Markdown
from scripts import utils, my_tools

LLM_MODEL = "qwen3"
BASE_URL = "http://localhost:11434"
llm = ChatOllama(model=LLM_MODEL, base_url=BASE_URL)

class AgentState(TypedDict):
    messages: Annotated[list, operator.add]

from scripts.my_tools import retrieve_docs

def agent_node(state: AgentState):
    messages = state['messages']

    tools = [retrieve_docs]
    llm_with_tools = llm.bind_tools(tools)

    system_prompt = """You are a financial document analysis assistant with access to a document retrieval tool.

                CRITICAL RULES:
                1. ALWAYS use the retrieve_docs tool first - NEVER answer from memory
                2. You MUST call the tool before providing any financial information
                3. Answer ONLY based on the retrieved documents
                4. If documents don't contain the answer, clearly state that

                WORKFLOW FOR COMPLEX/COMPARISON QUESTIONS:
                Step 1: Break down the question into sub-questions
                Step 2: Call retrieve_docs for EACH sub-question separately
                Step 3: Analyze all retrieved documents
                Step 4: Present comparison in TABLE format

                ANSWER FORMATTING (Use Markdown):
                - Use **headings** (##, ###) for sections
                - Use **tables** for comparisons and structured data
                - Cite sources: (Company: X, Year: Y, Quarter: Z, Page: N)

                REMEMBER:
                - ALWAYS call the tool first
                - Break complex questions into sub-questions
                - Always cite sources
                - If no relevant documents are found, try with different filters."""

    messages = [SystemMessage(system_prompt)] + messages
    response = llm_with_tools.invoke(messages)

    if hasattr(response, 'tool_calls') and response.tool_calls:
        for tc in response.tool_calls:
            print(f"[AGENT] called Tool {tc.get('name', '?')} with args {tc.get('args', '?')}")
    else:
        print(f"[AGENT] Responding...")

    return {'messages': [response]}

def should_continue(state: AgentState):
    last = state['messages'][-1]
    if hasattr(last, 'tool_calls') and last.tool_calls:
        return "tools"
    else:
        return END

def create_agent():
    builder = StateGraph(AgentState)

    builder.add_node('agent', agent_node)
    builder.add_node('tools', ToolNode([retrieve_docs]))

    builder.add_edge(START, 'agent')
    builder.add_conditional_edges('agent', should_continue, ['tools', END])
    builder.add_edge('tools', 'agent')

    return builder.compile()

agent = create_agent()