🤖 Multi-Agent AI System A high-performance, LangGraph-powered AI system with a Gradio front-end for evaluating multi-agent interactions across documents such as PDFs, Emails, and JSON. It uses intelligent input classification, specialized LLM agents, and a shared memory store—designed for scalable enterprise deployment.

🚀 Project Highlights ⚙️ LangGraph Multi-Agent Orchestration

🤝 Tool-Equipped Agents with shared context

📎 Input-Aware Routing: Email, PDF, JSON-specific logic

🧠 Cohere LLM for semantic extraction and intent recognition

💾 SQLite Context Memory with async and TTL support

🌐 Gradio Interface for testing and live interaction

🐳 Dockerized for consistent deployment

🧩 Agent Architecture mermaid Copy Edit graph TD A[User Input] --> B[Classifier Agent] B --> |PDF| C[PDF Agent] B --> |Email| D[Email Agent] B --> |JSON| E[JSON Agent] C --> F[Shared SQLite Memory] D --> F E --> F G --> H[Gradio Interface] ✨ Features Feature Description 🧭 LangGraph Routing Dynamically controls agent execution paths 📄 PDF Agent Extracts text with pdfplumber or OCR fallback ✉️ Email Agent Parses MIME emails, extracts metadata, analyzes urgency 📊 JSON Agent Validates schema, standardizes data 🔍 Evaluator Agent Automatically scores responses using Cohere 🧠 Redis Memory Shared context store with TTL 📈 Gradio UI Live, interactive testing 🛠️ Tool Integration Agents invoke external tools as needed

🛠️ Tech Stack 🧠 LLM: Cohere

🧭 Framework: LangGraph

⚡ Optional Backend: FastAPI

💾 Memory: SQLite

🐳 Deployment: Docker + Docker Compose

🧪 Testing: Pytest

⚡ Quickstart 🔋 Requirements Python 3.10+

Docker + Docker Compose

SQLite

Redis

Cohere API Key

🧪 Run Locally bash Copy Edit git clone https://github.com/S-n-e-h/Multi-Agent cd Multi-Agent

Set environment

cp .env.example .env

Fill in COHERE_API_KEY and REDIS_URL in .env

Install dependencies

pip install -r requirements.txt

Start the app

python app.py 🐳 Docker Setup (Recommended) bash Copy Edit docker-compose up --build 🧠 Core Agents 🧭 Classifier Agent Detects input type (PDF, Email, JSON)

Uses Cohere to extract business intent (e.g., Invoice, Complaint)

Routes to appropriate agent

Example: Input: email.eml Output: Routed to Email Agent with intent: Complaint

📄 PDF Agent Text extraction using pdfplumber or OCR

Extracts fields like invoice number, date, amount

Calls LLM for contract analysis

Example: PDF: Invoice-2024.pdf Output: { "InvoiceNo": "INV-123", "Total": "$400", "Vendor": "Acme Corp" }

✉️ Email Agent Parses .eml files

Extracts sender, urgency, tone, and subject

Converts to structured CRM-ready format

Example: Email: "Your service was delayed!" Output: { "Urgency": "High", "Sentiment": "Negative", "Sender": "john@example.com" }

⚙️ JSON Agent Validates schema

Detects missing or malformed fields

Extracts actionable entities

Example: Input: { "order_id": 123, "amount": null } Output: "amount" is missing or invalid.

📊 Evaluator Node Feature Details 🎯 Purpose Validate agent-generated responses 🔄 Inputs User query, Ground truth, Agent response 📉 Outputs Score (0–1), Comment, Reason 🤖 Powered By Cohere or a fine-tuned evaluator model

Example:

Ground Truth: Order total is $400

Model Output: Invoice shows $450

Score: 0.6 – mismatch in value

📁 Folder Structure css Copy Edit Multi-Agent/ ├── agents/ │ ├── classifier_agent.py │ ├── pdf_agent.py │ ├── email_agent.py │ └── json_agent.py ├── memory/ │ └── memory_store.py ├── Screenchots/ ├── inputs/ │ └── Demo-input-files ├── utils/ │ └── pdf_loader.py ├── router/ │ └── action_router.py ├── app.py ├── docker-compose.yml ├── Dockerfile ├── *.db (SQLite files) └── .env.example 🔍 API Overview Endpoint Method Description /process POST Accepts file input for processing /evaluate POST Evaluate agent response /health GET Basic health check /contexts GET Lists saved memory contexts

🧠 Future Roadmap 🔄 Webhook-based Input Ingestion

📚 Vector DB Integration (Pinecone/Weaviate)

💬 Chat-style Conversational Chains

📊 LangSmith / OpenInference Support

📈 Feedback Loop for Continual Agent Improvement

📌 Example Use Case 🎯 Scenario: Email Complaint → CRM-ready Output Input: .eml file with subject: "Billing issue – incorrect charges"

Classifier Agent: → Type: Email | Intent: Complaint

Email Agent: → { "Urgency": "High", "Sentiment": "Negative", "CustomerName": "Jane Doe" }

Memory Store: → Stores complaint ID and context for follow-up

🙌 Contributions Pull requests and suggestions are welcome! For major changes, open an issue first to discuss what you’d like to change.