Transition to perflab

README.md

@@ -1,10 +1,10 @@
-# 🧱 Deep Research From Scratch 
+# 🧱 Deep Research From Scratch
 
 Deep research has broken out as one of the most popular agent applications. [OpenAI](https://openai.com/index/introducing-deep-research/), [Anthropic](https://www.anthropic.com/engineering/built-multi-agent-research-system), [Perplexity](https://www.perplexity.ai/hub/blog/introducing-perplexity-deep-research), and [Google](https://gemini.google/overview/deep-research/?hl=en) all have deep research products that produce comprehensive reports using [various sources](https://www.anthropic.com/news/research) of context. There are also many [open](https://huggingface.co/blog/open-deep-research) [source](https://github.com/google-gemini/gemini-fullstack-langgraph-quickstart) implementations. We built an [open deep researcher](https://github.com/langchain-ai/open_deep_research) that is simple and configurable, allowing users to bring their own models, search tools, and MCP servers. In this repo, we'll build a deep researcher from scratch! Here is a map of the major pieces that we will build:
 
 ![overview](https://github.com/user-attachments/assets/b71727bd-0094-40c4-af5e-87cdb02123b4)
 
-## 🚀 Quickstart 
+## 🚀 Quickstart
 
 ### Prerequisites
 
@@ -35,6 +35,16 @@ curl -LsSf https://astral.sh/uv/install.sh | sh
 export PATH="/Users/$USER/.local/bin:$PATH"
 ```
 
+
+- Install necessary packages.
+```bash
+pip install langgraph langchain_core langchain langchain-openai
+pip install langchain-anthropic langchain_mcp_adapters mcp tavily-python
+pip install --upgrade jupyterlab ipywidgets
+
+```
+
+
 ### Installation
 
 1. Clone the repository:
@@ -79,17 +89,17 @@ source .venv/bin/activate  # On Windows: .venv\Scripts\activate
 jupyter notebook
 ```
 
-## Background 
+## Background
 
-Research is an open‑ended task; the best strategy to answer a user request can’t be easily known in advance. Requests can require different research strategies and varying levels of search depth. Consider this request. 
+Research is an open‑ended task; the best strategy to answer a user request can’t be easily known in advance. Requests can require different research strategies and varying levels of search depth. Consider this request.
 
 [Agents](https://langchain-ai.github.io/langgraph/tutorials/workflows/#agent) are well suited to research because they can flexibly apply different strategies, using intermediate results to guide their exploration. Open deep research uses an agent to conduct research as part of a three step process:
 
 1. **Scope** – clarify research scope
 2. **Research** – perform research
 3. **Write** – produce the final report
 
-## 📝 Organization 
+## 📝 Organization
 
 This repo contains 5 tutorial notebooks that build a deep research system from scratch:
 
@@ -201,4 +211,4 @@ This repo contains 5 tutorial notebooks that build a deep research system from s
 - **State Management**: Complex state flows across subgraphs and nodes
 - **Protocol Integration**: MCP servers and tool ecosystems
 
-Each notebook builds on the previous concepts, culminating in a production-ready deep research system that can handle complex, multi-faceted research queries with intelligent scoping and coordinated execution. 
+Each notebook builds on the previous concepts, culminating in a production-ready deep research system that can handle complex, multi-faceted research queries with intelligent scoping and coordinated execution.

src/deep_research_from_scratch/multi_agent_supervisor.py

@@ -10,11 +10,12 @@
 maintaining isolated context windows for each research topic.
 """
 
-import asyncio
+import os, asyncio
 
 from typing_extensions import Literal
 
 from langchain.chat_models import init_chat_model
+from langchain_openai import AzureChatOpenAI
 from langchain_core.messages import (
     HumanMessage, 
     BaseMessage, 
@@ -36,16 +37,16 @@
 
 def get_notes_from_tool_calls(messages: list[BaseMessage]) -> list[str]:
     """Extract research notes from ToolMessage objects in supervisor message history.
-
+    
     This function retrieves the compressed research findings that sub-agents
     return as ToolMessage content. When the supervisor delegates research to
     sub-agents via ConductResearch tool calls, each sub-agent returns its
     compressed findings as the content of a ToolMessage. This function
     extracts all such ToolMessage content to compile the final research notes.
-
+    
     Args:
         messages: List of messages from supervisor's conversation history
-
+        
     Returns:
         List of research note strings extracted from ToolMessage objects
     """
@@ -68,7 +69,13 @@ def get_notes_from_tool_calls(messages: list[BaseMessage]) -> list[str]:
 # ===== CONFIGURATION =====
 
 supervisor_tools = [ConductResearch, ResearchComplete, think_tool]
-supervisor_model = init_chat_model(model="anthropic:claude-sonnet-4-20250514")
+# supervisor_model = init_chat_model(model="anthropic:claude-sonnet-4-20250514")
+model_args = dict(azure_endpoint="https://llm-proxy.perflab.nvidia.com",
+    api_key=os.getenv("PERFLAB_ONEAPI"),
+    api_version="2025-02-01-preview",)
+supervisor_model = AzureChatOpenAI(
+    deployment_name=os.getenv("RESEARCH_MODEL", "claude-sonnet-4-20250514"),
+    temperature=0.0, **model_args)
 supervisor_model_with_tools = supervisor_model.bind_tools(supervisor_tools)
 
 # System constants
@@ -84,31 +91,31 @@ def get_notes_from_tool_calls(messages: list[BaseMessage]) -> list[str]:
 
 async def supervisor(state: SupervisorState) -> Command[Literal["supervisor_tools"]]:
     """Coordinate research activities.
-
+    
     Analyzes the research brief and current progress to decide:
     - What research topics need investigation
     - Whether to conduct parallel research
     - When research is complete
-
+    
     Args:
         state: Current supervisor state with messages and research progress
-
+        
     Returns:
         Command to proceed to supervisor_tools node with updated state
     """
     supervisor_messages = state.get("supervisor_messages", [])
-
+    
     # Prepare system message with current date and constraints
     system_message = lead_researcher_prompt.format(
         date=get_today_str(), 
         max_concurrent_research_units=max_concurrent_researchers,
         max_researcher_iterations=max_researcher_iterations
     )
     messages = [SystemMessage(content=system_message)] + supervisor_messages
-
+    
     # Make decision about next research steps
     response = await supervisor_model_with_tools.ainvoke(messages)
-
+    
     return Command(
         goto="supervisor_tools",
         update={
@@ -119,41 +126,41 @@ async def supervisor(state: SupervisorState) -> Command[Literal["supervisor_tool
 
 async def supervisor_tools(state: SupervisorState) -> Command[Literal["supervisor", "__end__"]]:
     """Execute supervisor decisions - either conduct research or end the process.
-
+    
     Handles:
     - Executing think_tool calls for strategic reflection
     - Launching parallel research agents for different topics
     - Aggregating research results
     - Determining when research is complete
-
+    
     Args:
         state: Current supervisor state with messages and iteration count
-
+        
     Returns:
         Command to continue supervision, end process, or handle errors
     """
     supervisor_messages = state.get("supervisor_messages", [])
     research_iterations = state.get("research_iterations", 0)
     most_recent_message = supervisor_messages[-1]
-
+    
     # Initialize variables for single return pattern
     tool_messages = []
     all_raw_notes = []
     next_step = "supervisor"  # Default next step
     should_end = False
-
+    
     # Check exit criteria first
     exceeded_iterations = research_iterations >= max_researcher_iterations
     no_tool_calls = not most_recent_message.tool_calls
     research_complete = any(
         tool_call["name"] == "ResearchComplete" 
         for tool_call in most_recent_message.tool_calls
     )
-
+    
     if exceeded_iterations or no_tool_calls or research_complete:
         should_end = True
         next_step = END
-
+    
     else:
         # Execute ALL tool calls before deciding next step
         try:
@@ -162,7 +169,7 @@ async def supervisor_tools(state: SupervisorState) -> Command[Literal["superviso
                 tool_call for tool_call in most_recent_message.tool_calls 
                 if tool_call["name"] == "think_tool"
             ]
-
+            
             conduct_research_calls = [
                 tool_call for tool_call in most_recent_message.tool_calls 
                 if tool_call["name"] == "ConductResearch"
@@ -206,20 +213,20 @@ async def supervisor_tools(state: SupervisorState) -> Command[Literal["superviso
                         tool_call_id=tool_call["id"]
                     ) for result, tool_call in zip(tool_results, conduct_research_calls)
                 ]
-
+                
                 tool_messages.extend(research_tool_messages)
 
                 # Aggregate raw notes from all research
                 all_raw_notes = [
                     "\n".join(result.get("raw_notes", [])) 
                     for result in tool_results
                 ]
-
+                
         except Exception as e:
             print(f"Error in supervisor tools: {e}")
             should_end = True
             next_step = END
-
+    
     # Single return point with appropriate state updates
     if should_end:
         return Command(

src/deep_research_from_scratch/research_agent.py

@@ -4,13 +4,14 @@
 This module implements a research agent that can perform iterative web searches
 and synthesis to answer complex research questions.
 """
-
+import os
 from pydantic import BaseModel, Field
 from typing_extensions import Literal
 
 from langgraph.graph import StateGraph, START, END
 from langchain_core.messages import SystemMessage, HumanMessage, ToolMessage, filter_messages
 from langchain.chat_models import init_chat_model
+from langchain_openai import AzureChatOpenAI
 
 from deep_research_from_scratch.state_research import ResearcherState, ResearcherOutputState
 from deep_research_from_scratch.utils import tavily_search, get_today_str, think_tool
@@ -23,20 +24,33 @@
 tools_by_name = {tool.name: tool for tool in tools}
 
 # Initialize models
-model = init_chat_model(model="anthropic:claude-sonnet-4-20250514")
+model_args = dict(azure_endpoint="https://llm-proxy.perflab.nvidia.com",
+    api_key=os.getenv("PERFLAB_ONEAPI"),
+    api_version="2025-02-01-preview",)
+model = AzureChatOpenAI(
+    deployment_name=os.getenv("RESEARCH_MODEL", "claude-sonnet-4-20250514"),
+    temperature=0.0, **model_args)
+summarization_model = AzureChatOpenAI(
+    deployment_name=os.getenv("SUMMARISATION_MODEL", "gpt-4.1-mini-20250414"),
+    temperature=0.0, **model_args)
+compress_model = AzureChatOpenAI(
+    deployment_name=os.getenv("COMPRESS_MODEL", "gpt-4.1-20250414"),
+    temperature=0.0, max_tokens=32000, **model_args)
+
+#model = init_chat_model(model="anthropic:claude-sonnet-4-20250514")
 model_with_tools = model.bind_tools(tools)
-summarization_model = init_chat_model(model="openai:gpt-4.1-mini")
-compress_model = init_chat_model(model="openai:gpt-4.1", max_tokens=32000) # model="anthropic:claude-sonnet-4-20250514", max_tokens=64000
+#summarization_model = init_chat_model(model="openai:gpt-4.1-mini")
+#compress_model = init_chat_model(model="openai:gpt-4.1", max_tokens=32000) # model="anthropic:claude-sonnet-4-20250514", max_tokens=64000
 
 # ===== AGENT NODES =====
 
 def llm_call(state: ResearcherState):
     """Analyze current state and decide on next actions.
-
+    
     The model analyzes the current conversation state and decides whether to:
     1. Call search tools to gather more information
     2. Provide a final answer based on gathered information
-
+    
     Returns updated state with the model's response.
     """
     return {
@@ -49,18 +63,18 @@ def llm_call(state: ResearcherState):
 
 def tool_node(state: ResearcherState):
     """Execute all tool calls from the previous LLM response.
-
+    
     Executes all tool calls from the previous LLM responses.
     Returns updated state with tool execution results.
     """
     tool_calls = state["researcher_messages"][-1].tool_calls
-
+ 
     # Execute all tool calls
     observations = []
     for tool_call in tool_calls:
         tool = tools_by_name[tool_call["name"]]
         observations.append(tool.invoke(tool_call["args"]))
-
+            
     # Create tool message outputs
     tool_outputs = [
         ToolMessage(
@@ -69,28 +83,28 @@ def tool_node(state: ResearcherState):
             tool_call_id=tool_call["id"]
         ) for observation, tool_call in zip(observations, tool_calls)
     ]
-
+    
     return {"researcher_messages": tool_outputs}
 
 def compress_research(state: ResearcherState) -> dict:
     """Compress research findings into a concise summary.
-
+    
     Takes all the research messages and tool outputs and creates
     a compressed summary suitable for the supervisor's decision-making.
     """
-
+    
     system_message = compress_research_system_prompt.format(date=get_today_str())
     messages = [SystemMessage(content=system_message)] + state.get("researcher_messages", []) + [HumanMessage(content=compress_research_human_message)]
     response = compress_model.invoke(messages)
-
+    
     # Extract raw notes from tool and AI messages
     raw_notes = [
         str(m.content) for m in filter_messages(
             state["researcher_messages"], 
             include_types=["tool", "ai"]
         )
     ]
-
+    
     return {
         "compressed_research": str(response.content),
         "raw_notes": ["\n".join(raw_notes)]
@@ -100,17 +114,17 @@ def compress_research(state: ResearcherState) -> dict:
 
 def should_continue(state: ResearcherState) -> Literal["tool_node", "compress_research"]:
     """Determine whether to continue research or provide final answer.
-
+    
     Determines whether the agent should continue the research loop or provide
     a final answer based on whether the LLM made tool calls.
-
+    
     Returns:
         "tool_node": Continue to tool execution
         "compress_research": Stop and compress research
     """
     messages = state["researcher_messages"]
     last_message = messages[-1]
-
+    
     # If the LLM makes a tool call, continue to tool execution
     if last_message.tool_calls:
         return "tool_node"

src/deep_research_from_scratch/research_agent_mcp.py

@@ -19,6 +19,7 @@
 from typing_extensions import Literal
 
 from langchain.chat_models import init_chat_model
+from langchain_openai import AzureChatOpenAI
 from langchain_core.messages import SystemMessage, HumanMessage, ToolMessage, filter_messages
 from langchain_mcp_adapters.client import MultiServerMCPClient
 from langgraph.graph import StateGraph, START, END
@@ -53,8 +54,17 @@ def get_mcp_client():
     return _client
 
 # Initialize models
-compress_model = init_chat_model(model="openai:gpt-4.1", max_tokens=32000)
-model = init_chat_model(model="anthropic:claude-sonnet-4-20250514")
+# compress_model = init_chat_model(model="openai:gpt-4.1", max_tokens=32000)
+# model = init_chat_model(model="anthropic:claude-sonnet-4-20250514")
+model_args = dict(azure_endpoint="https://llm-proxy.perflab.nvidia.com",
+    api_key=os.getenv("PERFLAB_ONEAPI"),
+    api_version="2025-02-01-preview",)
+model = AzureChatOpenAI(
+    deployment_name=os.getenv("RESEARCH_MODEL", "claude-sonnet-4-20250514"),
+    temperature=0.0, **model_args)
+compress_model = AzureChatOpenAI(
+    deployment_name=os.getenv("COMPRESS_MODEL", "gpt-4.1-20250414"),
+    temperature=0.0, max_tokens=32000, **model_args)
 
 # ===== AGENT NODES =====
 
@@ -145,7 +155,7 @@ def compress_research(state: ResearcherState) -> dict:
     This function filters out think_tool calls and focuses on substantive
     file-based research content from MCP tools.
     """
-
+    
     system_message = compress_research_system_prompt.format(date=get_today_str())
     messages = [SystemMessage(content=system_message)] + state.get("researcher_messages", []) + [HumanMessage(content=compress_research_human_message)]