Multi-Agent LangGraph News QA Assistant

Overview

This project is a multi-agent AI assistant built using LangGraph framework. It demonstrates how multiple agents can collaborate to extract topics from user queries, fetch relevant news, and generate concise summaries. The system is implemented in Python and leverages HuggingFace Transformers, LangGraph, and custom tools for a robust workflow.

The assistant currently runs as a console-based application, demonstrating multi-agent orchestration without requiring a UI.

---

Key Features

Intent Detection

• Agent: intent_agent
• Extracts a clear topic keyword from the user’s query using a local LLM (FLAN-T5-small).
• Ensures structured input for downstream agents.

News Retrieval

• Agent: news_agent
• Fetches the top relevant news articles for the detected topic.
• Uses a custom tool get_news to retrieve articles dynamically.

Summarization

• Agent: summarizer_agent
• Summarizes each article clearly and concisely.
• Outputs structured summaries including title, brief summary, and URL.

Multi-Agent Orchestration

• Implements a StateGraph to connect agents:

  • Intent → News → Summarizer

• Agents communicate via a typed state dictionary (NewsState) to maintain structured data flow.

• Console outputs track workflow progress, providing feedback for each step.

Human-in-the-Loop (HITL)

• After the summarization step, the system pauses to request user feedback.
• Users can approve the generated summaries or request refinement.
• If refinement is requested, the summarization agent is re-invoked with adjusted instructions before the
  workflow completes.
• This mechanism improves reliability, user trust, and output quality.

Robustness

• Automatically detects GPU availability for faster model inference.
• Uses HuggingFace pipelines for efficient text-to-text generation.

---

Technical Stack

• Python 3.x

• Transformers & HuggingFace Pipeline: FLAN-T5-small for local LLM inference

• LangGraph: Prebuilt REACT agents & StateGraph orchestration

• Custom Tools:

  • detect_topic_llm: Topic extraction
  • get_news: Fetch relevant news
  • summarize_articles: Article summarization

---

Installation & Usage

1. Clone the repository:
   git clone https://github.com/nirmalakuslekar/multi-agent-langgraph-news-qa-assistant.git

2. Navigate to the project directory:
   cd multi-agent-langgraph-news-qa-assistant

3. Create and activate a virtual environment:
   python -m venv venv
   source venv/bin/activate # Windows: venv\Scripts\activate

4. Install dependencies:
   pip install -r requirements.txt

5. Set the NewsAPI key as an environment variable:
   export NEWS_API_KEY="your_api_key_here" # Linux/macOS
   setx NEWS_API_KEY "your_api_key_here" # Windows

6. Run the application:
   python main.py

---

How It Works

1️⃣ Model Initialization

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, pipeline
from langchain_huggingface import ChatHuggingFace
from langchain_huggingface.llms import HuggingFacePipeline

device = 0 if torch.cuda.is_available() else -1
model_name = "google/flan-t5-small"
tokenizer = AutoTokenizer.from_pretrained(model_name)
base_model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
pipe = pipeline("text2text-generation", model=base_model, tokenizer=tokenizer, device=device)
llm_pipeline = HuggingFacePipeline(pipeline=pipe)
local_llm = ChatHuggingFace(llm=llm_pipeline)

2️⃣ Define State Structure

from typing import TypedDict, List

class NewsState(TypedDict):
    user_input: str
    topic: str
    articles: List[dict]
    summaries: List[dict]

3️⃣ Create REACT Agents

• Intent Agent: extracts topic keywords
• News Agent: retrieves news for the topic
• Summarizer Agent: generates concise summaries

intent_agent = create_react_agent(model=local_llm, tools=[detect_topic_llm], prompt="Extract topic keyword")
news_agent = create_react_agent(model=local_llm, tools=[get_news], prompt="Fetch relevant news")
summarizer_agent = create_react_agent(model=local_llm, tools=[summarize_articles], prompt="Summarize articles")

4️⃣ Build StateGraph Workflow

graph = StateGraph(NewsState)

graph.add_node(intent_node)
graph.add_node(news_node)
graph.add_node(summarizer_node)

graph.add_edge("intent_node", "news_node")
graph.add_edge("news_node", "summarizer_node")

graph.set_entry_point("intent_node")
graph.set_finish_point("summarizer_node")

5️⃣ Run Workflow

initial_state = NewsState(user_input="Get me the latest news about Netflix", topic="", articles=[], summaries=[])
compiled_graph = graph.compile()
final_state = compiled_graph.invoke(initial_state)

---

System Performance Evaluation

The system performance was evaluated qualitatively using real-world news queries
across domains such as entertainment, technology, and business.

Evaluation Metrics

• Topic Extraction Accuracy: Correctness of the topic keyword extracted by the Intent Agent.
• News Relevance: Relevance of fetched articles to the detected topic.
• Summary Quality: Coherence, clarity, and factual consistency of generated summaries.
• Human-in-the-Loop Approval Rate: Percentage of summaries accepted without refinement.
• End-to-End Latency: Approximate time taken to complete the full workflow.

Testing Methodology
Each user query was processed through the complete multi-agent pipeline:
Intent Detection → News Retrieval → Summarization → Human Review.

The system was tested manually with multiple queries. Human feedback was used
to validate summary quality and trigger refinement when necessary.

Baseline Comparison
Compared to a single-agent summarization pipeline, the multi-agent LangGraph
approach provides:

• Better modularity and maintainability
• Clear separation of responsibilities
• Improved interpretability and debugging
• Higher user trust through human-in-the-loop validation

Multi-Agent LangGraph News QA Assistant

This diagram illustrates the LangGraph-based multi-agent workflow, showing how the Intent Agent, News Agent, and Summarizer Agent communicate through a shared state.

This figure shows the runtime execution flow of the system, including human-in-the-loop interaction.

Project Repository / Source Code

The complete working code for this Multi-Agent LangGraph News QA Assistant can be accessed on GitHub:

GitHub Repository