sebasmos's picture
Update demo with latest codebase changes
420bcec
"""
AI Agent for project management using LangGraph.
"""
from typing import TypedDict, Annotated, Sequence, List, Dict, Any
import operator
import os
from langchain_core.messages import BaseMessage, HumanMessage, AIMessage, SystemMessage
from langchain_huggingface import ChatHuggingFace, HuggingFaceEndpoint
from langchain_google_genai import ChatGoogleGenerativeAI
from langgraph.graph import StateGraph, END
from src.rag import ProjectRAG
class AgentState(TypedDict):
"""State for the agent."""
messages: Annotated[Sequence[BaseMessage], operator.add]
query: str
retrieved_context: List[Dict[str, Any]]
action_items: List[Dict[str, Any]]
blockers: List[Dict[str, Any]]
next_step: str
final_answer: str
class ProjectAgent:
"""AI Agent for project management queries."""
def __init__(self, rag: ProjectRAG, provider: str = "huggingface", model_name: str = None):
"""Initialize the agent.
Args:
rag: ProjectRAG instance for retrieval
provider: "huggingface" (free) or "google" (paid)
model_name: Optional model name override
"""
self.rag = rag
self.provider = provider
if provider == "google":
# Use Google Gemini API (paid)
google_api_key = os.getenv("GOOGLE_API_KEY")
if not google_api_key:
raise ValueError("GOOGLE_API_KEY environment variable not set")
self.llm = ChatGoogleGenerativeAI(
model=model_name or "gemini-2.5-flash-lite",
temperature=0.1,
google_api_key=google_api_key,
timeout=60, # 60 second timeout
convert_system_message_to_human=True # Better compatibility
)
else:
# Use HF Inference API (free tier)
hf_token = os.getenv("HF_TOKEN") or os.getenv("HUGGING_FACE_HUB_TOKEN")
if not hf_token:
raise ValueError("HF_TOKEN environment variable not set")
llm = HuggingFaceEndpoint(
repo_id=model_name or "meta-llama/Llama-3.2-3B-Instruct",
temperature=0.1,
max_new_tokens=512,
huggingfacehub_api_token=hf_token,
timeout=60 # 60 second timeout to prevent hanging
)
self.llm = ChatHuggingFace(llm=llm)
self.graph = self._build_graph()
def _build_graph(self) -> StateGraph:
"""Build the agent's state graph."""
workflow = StateGraph(AgentState)
# Add nodes
workflow.add_node("analyze_query", self.analyze_query)
workflow.add_node("retrieve_context", self.retrieve_context)
workflow.add_node("get_action_items", self.get_action_items)
workflow.add_node("get_blockers", self.get_blockers)
workflow.add_node("generate_answer", self.generate_answer)
# Add edges
workflow.set_entry_point("analyze_query")
workflow.add_edge("analyze_query", "retrieve_context")
workflow.add_conditional_edges(
"retrieve_context",
self.route_after_retrieval,
{
"action_items": "get_action_items",
"blockers": "get_blockers",
"answer": "generate_answer"
}
)
workflow.add_edge("get_action_items", "generate_answer")
workflow.add_edge("get_blockers", "generate_answer")
workflow.add_edge("generate_answer", END)
return workflow.compile()
def analyze_query(self, state: AgentState) -> AgentState:
"""Analyze the user's query to understand intent."""
query = state["query"]
system_prompt = """You are a query analyzer for a project management assistant.
Analyze queries and determine what information is being requested."""
analysis_prompt = f"""Analyze this query and determine:
1. What information is being requested?
2. Which project (if specified)?
3. What type of query is this (action items, blockers, status, decisions, general)?
Query: {query}
Respond in this format:
Type: [action_items|blockers|status|decisions|general]
Project: [project name or "all"]
Intent: [brief description]
"""
messages = [
SystemMessage(content=system_prompt),
HumanMessage(content=analysis_prompt)
]
response = self.llm.invoke(messages)
state["messages"] = state.get("messages", []) + [
HumanMessage(content=query),
AIMessage(content=f"Analysis: {response.content}")
]
return state
def retrieve_context(self, state: AgentState) -> AgentState:
"""Retrieve relevant context from the RAG system."""
query = state["query"]
# Extract project name if mentioned
project_filter = None
projects = self.rag.get_all_projects()
for project in projects:
if project.lower() in query.lower():
project_filter = project
break
# Search for relevant context
results = self.rag.search(query, n_results=5, project_filter=project_filter)
state["retrieved_context"] = results
return state
def route_after_retrieval(self, state: AgentState) -> str:
"""Route to appropriate node based on query type."""
query = state["query"].lower()
if any(term in query for term in ["action item", "todo", "task", "what's next", "what should"]):
return "action_items"
elif any(term in query for term in ["blocker", "issue", "problem", "stuck"]):
return "blockers"
else:
return "answer"
def get_action_items(self, state: AgentState) -> AgentState:
"""Get action items from the RAG system."""
query = state["query"].lower()
# Extract project name if mentioned
project_filter = None
projects = self.rag.get_all_projects()
for project in projects:
if project.lower() in query:
project_filter = project
break
action_items = self.rag.get_open_action_items(project=project_filter)
state["action_items"] = action_items
return state
def get_blockers(self, state: AgentState) -> AgentState:
"""Get blockers from the RAG system."""
query = state["query"].lower()
# Extract project name if mentioned
project_filter = None
projects = self.rag.get_all_projects()
for project in projects:
if project.lower() in query:
project_filter = project
break
blockers = self.rag.get_blockers(project=project_filter)
state["blockers"] = blockers
return state
def generate_answer(self, state: AgentState) -> AgentState:
"""Generate the final answer using retrieved context."""
query = state["query"]
context = state.get("retrieved_context", [])
action_items = state.get("action_items", [])
blockers = state.get("blockers", [])
# Build context string
context_parts = []
if context:
context_parts.append("Relevant meeting context:")
for i, result in enumerate(context[:3], 1):
context_parts.append(f"\n[Context {i}]")
context_parts.append(result['content'])
if 'metadata' in result:
meta = result['metadata']
context_parts.append(f"(From: {meta.get('project', 'Unknown')} - {meta.get('title', 'Unknown')})")
if action_items:
context_parts.append("\nOpen Action Items:")
for item in action_items:
assignee = f" ({item['assignee']})" if item.get('assignee') else ""
deadline = f" by {item['deadline']}" if item.get('deadline') else ""
context_parts.append(f"- {item['task']}{assignee}{deadline}")
if blockers:
context_parts.append("\nCurrent Blockers:")
for blocker in blockers:
context_parts.append(f"- {blocker['blocker']}")
context_str = "\n".join(context_parts)
# Generate answer
system_prompt = """You are a helpful AI assistant that helps users manage their projects.
Use the provided context to answer the user's question accurately and concisely.
Format your response using bullet points for clarity.
For action items, list the task with the assignee in parentheses at the end.
For blockers and risks, list them directly without project names.
Keep responses brief and to the point. Avoid lengthy explanations.
Example format:
## Next Actions
- Task description (Assignee) by deadline
- Another task (Assignee)
## Blockers/Risks
- Blocker description
- Another blocker"""
messages = [
SystemMessage(content=system_prompt),
HumanMessage(content=f"Context:\n{context_str}\n\nQuestion: {query}\n\nAnswer:")
]
response = self.llm.invoke(messages)
state["final_answer"] = response.content
return state
def query(self, user_query: str) -> str:
"""Process a user query and return an answer."""
initial_state = {
"messages": [],
"query": user_query,
"retrieved_context": [],
"action_items": [],
"blockers": [],
"next_step": "",
"final_answer": ""
}
result = self.graph.invoke(initial_state)
return result["final_answer"]
def stream_query(self, user_query: str):
"""Process a user query and stream the answer token by token."""
# First run analysis and retrieval (non-streaming)
initial_state = {
"messages": [],
"query": user_query,
"retrieved_context": [],
"action_items": [],
"blockers": [],
"next_step": "",
"final_answer": ""
}
# Run through analysis and retrieval nodes
state = self.analyze_query(initial_state)
state = self.retrieve_context(state)
# Determine route and get additional data
route = self.route_after_retrieval(state)
if route == "action_items":
state = self.get_action_items(state)
elif route == "blockers":
state = self.get_blockers(state)
# Now stream the final answer generation
query = state["query"]
context = state.get("retrieved_context", [])
action_items = state.get("action_items", [])
blockers = state.get("blockers", [])
# Build context string
context_parts = []
if context:
context_parts.append("Relevant meeting context:")
for i, result in enumerate(context[:3], 1):
context_parts.append(f"\n[Context {i}]")
context_parts.append(result['content'])
if 'metadata' in result:
meta = result['metadata']
context_parts.append(f"(From: {meta.get('project', 'Unknown')} - {meta.get('title', 'Unknown')})")
if action_items:
context_parts.append("\nOpen Action Items:")
for item in action_items:
assignee = f" ({item['assignee']})" if item.get('assignee') else ""
deadline = f" by {item['deadline']}" if item.get('deadline') else ""
context_parts.append(f"- {item['task']}{assignee}{deadline}")
if blockers:
context_parts.append("\nCurrent Blockers:")
for blocker in blockers:
context_parts.append(f"- {blocker['blocker']}")
context_str = "\n".join(context_parts)
# Generate streaming answer
system_prompt = """You are a helpful AI assistant that helps users manage their projects.
Use the provided context to answer the user's question accurately and concisely.
Format your response using bullet points for clarity.
For action items, list the task with the assignee in parentheses at the end.
For blockers and risks, list them directly without project names.
Keep responses brief and to the point. Avoid lengthy explanations.
Example format:
## Next Actions
- Task description (Assignee) by deadline
- Another task (Assignee)
## Blockers/Risks
- Blocker description
- Another blocker"""
messages = [
SystemMessage(content=system_prompt),
HumanMessage(content=f"Context:\n{context_str}\n\nQuestion: {query}\n\nAnswer:")
]
# Stream tokens
full_response = ""
try:
for chunk in self.llm.stream(messages):
if hasattr(chunk, 'content') and chunk.content:
full_response += chunk.content
yield full_response
except Exception:
# Fallback to non-streaming if streaming not supported
response = self.llm.invoke(messages)
yield response.content