deep-research-agent / research_agent.py
pythonpyro's picture
Update research_agent.py
dc729fc verified
Raw History Blame Contribute Delete
20.7 kB
# ============================================================================
# research_agent.py - BACKEND
# ============================================================================
from typing import TypedDict, List, Literal, Annotated
import os
from dotenv import load_dotenv
from langgraph.graph import StateGraph, START, END
from langgraph.graph.message import add_messages
from langgraph.prebuilt import ToolNode
from langchain.tools import tool
from langchain_openai import ChatOpenAI
from langchain_core.messages import HumanMessage, SystemMessage
from openai import OpenAI
from tavily import TavilyClient
# ============================================================================
# CONFIGURATION & INITIALIZATION
# ============================================================================
# Load environment variables from .env file
load_dotenv()
# Initialize API clients with environment variables as fallback
# These will be overridden by user-provided keys in process_message
OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
TAVILY_API_KEY = os.getenv("TAVILY_API_KEY")
# Initialize API clients with environment variables
# These will be re-initialized with user-provided keys when needed
openai_client = None
llm = None
light_llm = None
def initialize_clients(openai_key: str, tavily_key: str):
"""
Initialize API clients with provided keys.
This function is called from process_message with user-provided API keys.
"""
global openai_client, llm, light_llm, tool_node, llm_with_tools
if not openai_key or not tavily_key:
raise ValueError("Both OpenAI and Tavily API keys must be provided.")
openai_client = OpenAI(api_key=openai_key)
llm = ChatOpenAI(model="gpt-5.1", api_key=openai_key, temperature=0) # Main LLM for research and synthesis
light_llm = ChatOpenAI(model="gpt-4o-mini", api_key=openai_key, temperature=0) # Lighter LLM for clarification
# Initialize tool node and bind tools to LLM
tool_node = ToolNode(tools)
llm_with_tools = llm.bind_tools([rag_search, web_search], tool_choice="auto")
# Create and compile the research graph now that tool_node is ready
create_research_graph()
# Set maximum number of research iterations to prevent infinite loops
# This is a default value and will be overridden by user input
DEFAULT_MAX_ITERATIONS = 3
# Global variable to store max iterations for research
MAX_ITERATIONS = DEFAULT_MAX_ITERATIONS
# Global variable to store tavily API key for use in web_search tool
_tavily_api_key = None
# ============================================================================
# VECTOR STORE SETUP
# ============================================================================
# Connect to pre-existing vector store containing embedded PDF documents
VECTOR_STORE_NAME = "yt-store"
store_id = None
def get_vector_store_id():
"""
Retrieve the vector store ID. This needs to be called after clients are initialized.
"""
global store_id
if not openai_client:
raise ValueError("OpenAI client not initialized. Call initialize_clients first.")
# Search for the vector store by name in the OpenAI account
for s in openai_client.vector_stores.list().data:
if s.name == VECTOR_STORE_NAME:
store_id = s.id
break
# Raise error if vector store is not found
if not store_id:
raise ValueError("❌ Vector store not found!")
print(f"β™» Using vector store ID: {store_id}")
return store_id
# ============================================================================
# TOOL DEFINITIONS
# ============================================================================
@tool
def rag_search(query: str) -> str:
"""
Search INTERNAL knowledge base containing curated documents,
architecture guides, tutorials, and domain expertise.
Use for conceptual understanding, internal best practices,
and previously authored materials.
"""
print(f"πŸ” RAG Search called with query: {query}")
# Query the vector store with semantic search
res = openai_client.vector_stores.search(
vector_store_id=store_id,
query=query,
max_num_results=3, # Return top 3 most relevant chunks
ranking_options={"score_threshold": 0.5, "ranker": "auto"} # Filter by relevance score
)
# Handle case where no results are found
if not res.data:
return "No relevant information found in internal documents."
# Extract text content from search results
out = []
for item in res.data:
for block in item.content or []:
if block.type == "text" and block.text:
out.append(block.text)
# Combine all text chunks with separator
result = "\n\n".join(out) if out else "No relevant text extracted from internal documents."
print(f"βœ… RAG Search returned {len(out)} chunks")
return result
@tool
def web_search(query: str) -> str:
"""
Search the public internet for recent developments,
benchmarks, and up-to-date information.
"""
global _tavily_api_key
print(f"🌐 Web Search called with query: {query}")
try:
# Use Tavily API for web search
tavily_client = TavilyClient(api_key=_tavily_api_key)
response = tavily_client.search(query, max_results=3)
# Format search results into readable text
results = []
for result in response.get('results', []):
results.append(f"Title: {result.get('title', 'N/A')}\nContent: {result.get('content', 'N/A')}\nURL: {result.get('url', 'N/A')}")
final_result = "\n\n---\n\n".join(results) if results else "No web search results found."
print(f"βœ… Web Search returned {len(results)} results")
return final_result
except Exception as e:
print(f"❌ Web search error: {str(e)}")
return f"Web search error: {str(e)}"
# ============================================================================
# TOOL NODE SETUP
# ============================================================================
# Combine tools into a list for the agent
tools = [rag_search, web_search]
# These will be initialized when initialize_clients() is called
tool_node = None
llm_with_tools = None
# ============================================================================
# STATE DEFINITION
# ============================================================================
# Define the state structure that will be passed between nodes
class AgentState(TypedDict):
messages: Annotated[List, add_messages] # Conversation history with automatic message aggregation
research_steps: List[str] # Track the research plan steps
final_answer: str # Store the synthesized final answer
iterations: int # Count how many research iterations have occurred
# ============================================================================
# CLARIFICATION NODE
# ============================================================================
# System prompt for the clarification phase
CLARIFICATION_PROMPT = """
You are a research assistant helping refine a user's research request.
Your goal is NOT to answer yet.
Instead:
1. Acknowledge the topic positively
2. Ask smart follow-up questions to narrow scope
3. Offer categories to choose from
4. Ask about:
- target audience
- depth level
- specific domains
- time range
- technical vs business focus
Example style:
"Great topic! Before I begin deep research, could you help me refine the scope?"
Then ask 3-5 concise questions.
DO NOT start researching yet.
"""
def clarification_node(state: AgentState):
"""
Node that asks clarifying questions to refine the research scope.
This runs BEFORE research begins.
"""
print("\nπŸ’¬ Clarification phase...\n")
messages = state["messages"]
# If we already asked clarification, just wait for user input
if state.get("phase") == "awaiting_user":
return state
# Use lighter LLM to generate clarification questions
response = light_llm.invoke([
SystemMessage(content=CLARIFICATION_PROMPT),
messages[-1]
])
return {
"messages": [response]
}
# ============================================================================
# PLANNER NODE
# ============================================================================
# System prompt for research planning
PLANNER_PROMPT = """
You are a deep research planning assistant.
Your job is to create a DUAL-SOURCE research plan.
For each major aspect of the user question:
Generate:
- ONE query specifically designed for INTERNAL knowledge (rag_search)
- ONE complementary query for EXTERNAL knowledge (web_search)
Return the plan in this EXACT format:
TOPIC 1:
RAG_QUERY: <query text optimized for internal documents>
WEB_QUERY: <query text optimized for current internet info>
TOPIC 2:
RAG_QUERY: ...
WEB_QUERY: ...
TOPIC 3:
RAG_QUERY: ...
WEB_QUERY: ...
At least 3 topics must be generated.
The goal is to ensure BOTH tools are used for comprehensive coverage.
"""
def planner_node(state: AgentState):
"""
Node that creates a structured research plan.
Generates paired RAG and Web queries for each research topic.
"""
print("\n🧠 Planning research approach...\n")
# Extract the user's refined question
user_question = state["messages"][-1].content
messages = [
SystemMessage(content=PLANNER_PROMPT),
HumanMessage(content=user_question)
]
# Generate the research plan
plan = llm.invoke(messages)
return {
"messages": [plan],
"research_steps": [plan.content] # Store plan for tracking
}
# ============================================================================
# RESEARCH AGENT NODE
# ============================================================================
# System prompt for the research agent
RESEARCH_PROMPT = """
You are a balanced deep research assistant.
You MUST use BOTH tools for every major topic:
- rag_search β†’ to gather internal curated knowledge
- web_search β†’ to gather current external information
Do NOT rely on only one tool.
Process:
1. Look at the research plan
2. For each topic:
- First call rag_search with the RAG_QUERY
- Then call web_search with the WEB_QUERY
You may call tools multiple times in one response.
Continue until all planned topics have been covered.
Only stop when:
- every RAG query has been executed
- every WEB query has been executed
"""
def research_agent_node(state: AgentState):
"""
Node that executes the research plan by calling tools.
This is the core agentic loop that gathers information.
"""
print("\nπŸ”Ž Research agent analyzing next step...\n")
messages = state["messages"]
# Add system prompt if not already present
if not any(isinstance(msg, SystemMessage) for msg in messages):
messages = [SystemMessage(content=RESEARCH_PROMPT)] + messages
# LLM decides which tools to call based on the plan
response = llm_with_tools.invoke(messages)
return {
"messages": [response],
"iterations": state.get("iterations", 0) + 1 # Increment iteration counter
}
# ============================================================================
# SYNTHESIZER NODE
# ============================================================================
# System prompt for final synthesis
SYNTHESIS_PROMPT = """
You are a research report writer.
Using ALL the previous research results:
- Internal RAG findings
- Web findings
- Multiple iterations of searches
Create a detailed, well-structured final answer.
Format:
OVERVIEW
-------
INTERNAL KNOWLEDGE FINDINGS
--------------------------
WEB RESEARCH FINDINGS
---------------------
FINAL SYNTHESIS
--------------
CONCLUSION
---------
"""
def synthesizer_node(state: AgentState):
"""
Node that synthesizes all research findings into a structured report.
This is the final step that produces the deliverable.
"""
print("\n✍️ Synthesizing final answer...\n")
# Clean messages to remove tool calls and tool responses
# (only keep actual content for cleaner synthesis)
clean_messages = []
for msg in state["messages"]:
# Skip any message containing tool calls
if hasattr(msg, "tool_calls") and msg.tool_calls:
continue
# Skip any raw tool response messages
if msg.type == "tool":
continue
clean_messages.append(msg)
# Add synthesis prompt and generate final report
messages = clean_messages + [
SystemMessage(content=SYNTHESIS_PROMPT)
]
final = llm.invoke(messages)
return {
"messages": [final],
"final_answer": final.content # Store the final answer
}
# ============================================================================
# ROUTING LOGIC
# ============================================================================
def research_router(state: AgentState) -> Literal["tools", "synthesize"]:
"""
Router that decides whether to continue research or move to synthesis.
Decision logic:
1. Ensure both RAG and Web tools have been used at least once
2. Stop if max iterations reached (safety limit)
3. Continue if agent is still calling tools
4. Otherwise, move to synthesis
"""
last_message = state["messages"][-1]
current_iterations = state.get("iterations", 0)
print(f"πŸ”„ Iteration count: {current_iterations}")
# Check conversation history to see which tools were used
history = str(state["messages"])
used_rag = "rag_search" in history
used_web = "web_search" in history
# Force at least one call to each tool for comprehensive research
if not (used_rag and used_web):
print("πŸ” Ensuring both RAG and Web sources are used...")
return "tools"
# HARD STOP: Prevent infinite loops by limiting iterations
if current_iterations >= MAX_ITERATIONS:
print("πŸ›‘ Max iterations reached – forcing synthesis")
return "synthesize"
# If agent is still requesting tool calls, continue research
if hasattr(last_message, "tool_calls") and last_message.tool_calls:
print("πŸ” Continuing research loop...")
return "tools"
# Otherwise, research is complete
print("βœ… Research sufficient. Moving to synthesis.")
return "synthesize"
# ============================================================================
# GRAPH CONSTRUCTION
# ============================================================================
# ---------- CLARIFICATION GRAPH ----------
# Simple linear graph: just asks clarifying questions
clarify_graph = StateGraph(AgentState)
# Add nodes
clarify_graph.add_node("clarify", clarification_node)
# Add edges: START β†’ clarify β†’ END
clarify_graph.add_edge(START, "clarify")
clarify_graph.add_edge("clarify", END)
# Compile the graph into an executable app
clarify_app = clarify_graph.compile()
# ---------- RESEARCH WORKFLOW GRAPH ----------
# This will be compiled on first use after clients are initialized
research_app = None
def create_research_graph():
"""
Create and compile the research workflow graph.
Must be called after initialize_clients() so that tool_node is ready.
"""
global research_app
if tool_node is None:
raise ValueError("Tool node not initialized. Call initialize_clients first.")
research_graph = StateGraph(AgentState)
# Add nodes
research_graph.add_node("planner", planner_node)
research_graph.add_node("researcher", research_agent_node)
research_graph.add_node("tools", tool_node)
research_graph.add_node("synthesizer", synthesizer_node)
# Add edges
# Flow: START β†’ planner β†’ researcher β†’ (conditional loop) β†’ synthesizer β†’ END
research_graph.add_edge(START, "planner")
research_graph.add_edge("planner", "researcher")
# Conditional edge: researcher decides to continue or finish
research_graph.add_conditional_edges(
"researcher",
research_router, # Router function decides next step
{
"tools": "tools", # Continue research
"synthesize": "synthesizer" # Finish and synthesize
}
)
# After tools execute, loop back to researcher
research_graph.add_edge("tools", "researcher")
# After synthesis, end the workflow
research_graph.add_edge("synthesizer", END)
# Compile the research graph
research_app = research_graph.compile()
return research_app
# ============================================================================
# BACKEND API FUNCTIONS
# ============================================================================
def create_initial_state():
"""
Initialize the session state for a new conversation.
Tracks: messages, iterations, answers, and current stage.
"""
return {
"messages": [],
"iterations": 0,
"final_answer": "",
"research_steps": [],
"stage": "awaiting_topic" # UI flow stage
}
def process_message(user_input, history, state, openai_api_key, tavily_api_key, max_iterations=2):
"""
Main message processing function that handles the multi-turn conversation.
Flow:
1. User enters topic β†’ Clarification questions
2. User answers β†’ Deep research begins
3. Research completes β†’ Final report shown
Args:
max_iterations: Number of research loops to perform (2 means one loop of the research node)
"""
global _tavily_api_key, MAX_ITERATIONS
# Set the global MAX_ITERATIONS with user-provided value
MAX_ITERATIONS = max_iterations
# Initialize clients with provided API keys
try:
initialize_clients(openai_api_key, tavily_api_key)
_tavily_api_key = tavily_api_key
# Retrieve vector store ID with initialized client
get_vector_store_id()
except ValueError as e:
error_msg = f"❌ Configuration Error: {str(e)}"
if history is None:
history = []
history.append({"role": "assistant", "content": error_msg})
return "", history, state
# Initialize state if first message
if state is None:
state = create_initial_state()
if history is None:
history = []
stage = state.get("stage", "awaiting_topic")
# ---- FIRST TURN: Initial topic entry ----
if stage == "awaiting_topic":
# Store user's initial question
state["messages"] = [HumanMessage(content=user_input)]
# Run clarification graph to ask follow-up questions
state = clarify_app.invoke(state)
clarification = state["messages"][-1].content
# Update chat history
history.append({"role": "user", "content": user_input})
history.append({"role": "assistant", "content": clarification})
# Move to next stage
state["stage"] = "awaiting_clarification"
return "", history, state
# ---- SECOND TURN: User provides clarification ----
elif stage == "awaiting_clarification":
# Add user's clarification to conversation
state["messages"].append(HumanMessage(content=user_input))
# Clean artifacts from previous interactions
clean_messages = []
for msg in state["messages"]:
if hasattr(msg, "tool_calls") and msg.tool_calls:
continue
if msg.type == "tool":
continue
clean_messages.append(msg)
# Reset state for fresh research
state["messages"] = clean_messages
state["iterations"] = 0
# Update UI
history.append({"role": "user", "content": user_input})
history.append({"role": "assistant", "content": "πŸ” Starting deep research... please wait..."})
# Run the full research workflow graph
final_state = research_app.invoke(state)
# Extract final answer
final_answer = final_state["messages"][-1].content
# Show result in chat
history.append({"role": "assistant", "content": final_answer})
# Mark research as completed
final_state["stage"] = "completed"
return "", history, final_state
# ---- AFTER COMPLETION: No further research ----
else:
history.append({"role": "assistant", "content": "Research already completed. Refresh page to start a new topic."})
return "", history, state