Agentic Knowledge System

An intelligent, multi-agent knowledge system that combines LLMs with web research and RAG. Built with LangGraph — queries flow through a Supervisor → Classifier → Research → Knowledge Base → Answer pipeline.

Agent Pipeline

User Query
    │
    ▼
┌────────────────┐
│  SUPERVISOR    │  Decides: direct answer or research needed
└──────┬─────────┘
       │
       ▼ (if research)
┌────────────────┐
│  CLASSIFIER    │  Routes query to domain: software, school, job_prep, general, math, non_tech
└──────┬─────────┘
       │
       ▼
┌────────────────┐
│  RESEARCH      │  Web search (Tavily), Arxiv, coding sources, finance news
└──────┬─────────┘
       │
       ▼
┌────────────────┐
│  KNOWLEDGE BASE│  Stores research → Qdrant vector DB → retrieves relevant docs
└──────┬─────────┘
       │
       ▼
┌────────────────┐
│  SUPERVISOR    │  Final answer with RAG context
└──────┬─────────┘
       │
       ▼
   Final Answer

Features

• Multi-agent graph pipeline — LangGraph orchestrates 4 specialized nodes
• Structured LLM routing — Classifier uses structured output for domain prediction
• Web research tools — Tavily search, Arxiv papers, coding sources, finance news
• RAG with Qdrant — Research data is chunked, embedded (AWS Bedrock Titan), and stored for retrieval
• Gradio chat UI — Step-by-step visualization of the pipeline
• CLI interface — Simple interactive REPL for terminal use

Requirements

• Python 3.14+
• Groq API key (LLM )
• Tavily API key (web search)
• News API key (finance/news tool)
• AWS credentials (Bedrock Titan embeddings)

Installation

git clone https://github.com/arpanchakraborty23/Agentic-Knowledge-System.git
cd Agentic-Knowledge-System

cp .env.example .env
# Edit .env with your API keys

uv sync

Usage

CLI

uv run python main.py

Web UI (Gradio)

uv run python ui.py

Opens a browser at http://localhost:7860. Each pipeline step streams in as it completes.

Programmatic

from src.agents.agent import GraphAgent

agent = GraphAgent()
response = agent.invoke(user_id="user_001", query="What is quantum computing?")
print(response.get("supervisor_answer"))

Project Structure

├── src/
│   ├── agents/agent.py      # GraphAgent, GraphBuilder
│   ├── nodes/               # Supervisor, Classifier, Research, Knowledge
│   ├── tools/               # Research tools (web, arxiv, code, finance) + RAG tool
│   ├── prompts/             # LangChain prompt templates
│   ├── rag/                 # DataParser, TextChunker, VectorStoreManager
│   ├── constants/           # GraphState, config, schemas
│   └── utils/               # Logger, LLM chain helper, URL fetcher
├── main.py                  # CLI entry point
├── ui.py                    # Gradio web interface
├── pyproject.toml
└── qdrant_data/             # Persistent vector store

Configuration

Variable	Description
GROQ_API_KEY	LLM provider (llama-3.3-70b-versatile)
TAVILY_API_KEY	Web search
NEWS_API_KEY	Finance/market news
AWS_ACCESS_KEY_ID	Bedrock embeddings
AWS_SECRET_ACCESS_KEY	Bedrock embeddings
AWS_REGION	AWS region (e.g. us-east-1)
RAG_CHUNK_SIZE	Doc chunk size (default: 1000)
RAG_TTL_DAYS	Vector store TTL (default: 10)

Tech Stack

• LangChain + LangGraph — agent graph orchestration
• Groq — LLM inference (llama-3.3-70b-versatile)
• AWS Bedrock — embeddings (titan-embed-text-v2)
• Qdrant — vector database
• Tavily — web search API
• Gradio — web UI
• Playwright — URL content extraction