🤖 AI Onboarding Assistant for Communities

This project implements a stateful AI agent designed to assist in managing and scaling community onboarding. The agent can answer questions using RAG (Retrieval-Augmented Generation), evaluate new applications using LLM-based scoring, and provide an interactive chat interface via Gradio — all built and tested in a Kaggle notebook environment.

I started this project after completing a 5 days intensive GenAI course: Course link

🧠 What This Agent Does

• ✅ Answers Community Questions
  Uses a vector database (ChromaDB) and semantic search to retrieve relevant information from internal documentation (PDFs), powered by Gemini embeddings and LangChain’s RAG integration.

• 📝 Scores New Member Applications
  Takes answers from a Google Form stored in a Google Sheet, and evaluates each application based on predefined criteria using Gemini LLM scoring with structured output (JSON schema).

• 🔍 Tool-Using Agent Architecture
  Built using LangGraph to define the agent as a state machine with tools and routing logic.

• 💬 Gradio Chat Interface
  Provides a styled, interactive chat experience where users can ask questions, invoke tools implicitly, and view agent logs and scoring results in real-time.

---

🧱 Technologies Used

Component	Description
LangGraph	Defines the AI agent as a graph of nodes (chat, tool handling, scoring).
LangChain + Gemini	LLM interface for chat and scoring tasks (via ChatGoogleGenerativeAI).
ChromaDB	Vector store to enable fast semantic search for RAG.
Python requests	Pulls new applications submitted through a Google Form by reading a public spreadsheet
Gradio	Provides a fully interactive frontend in the notebook.
PyMuPDF & unstructured	Used for chunking and extracting content from community PDF files.

---

🔄 How It Works

1. RAG (Retrieval-Augmented Generation)

• PDFs about the community (mission, goals, values, policies) are processed into semantic chunks using three strategies:
  • Fixed-size + overlap
  • One chunk per document
  • Semantic chunking via unstructured
• Chunks are embedded using Gemini’s text-embedding-004 model and stored in ChromaDB.
• get_info(query: str) is a tool that performs semantic search and returns a formatted markdown snippet.

2. Application Scoring

• The score_application() and check_new_applications() tools allow the agent to:
  • Evaluate individual or batch applications.
  • Parse responses into a JSON schema with score, verdict (approve, review, reject), and detailed reasoning.
  • Save the evaluations to a CSV and display results in a styled table.

3. Agent Logic with LangGraph

• Nodes:

  • chatbot: LLM chat interaction
  • score_node: Handles structured scoring
  • tools: Automatically resolves tool calls
  • human: Manages user input from Gradio

• Routing:

  • Routes messages from chatbot to tools if tool calls are present
  • Routes back to chat node after tool execution
  • Ends when user types “q” or “quit”

---

💻 Running in Kaggle Notebook

All components (agent logic, UI, RAG setup, and evaluation) are implemented in a Kaggle Notebook. The notebook includes:

• Inline markdown explanations
• Tool-calling logic
• Cell-level breakdown for easy learning
• Integration with Google Sheets via API key

You can duplicate the notebook and run everything end-to-end without extra setup.

---

🧪 Example Prompts

• “What is the mission of this community?”
• “Can you check and score new applications?”
• “How can I contribute as a new member?”
• “Can you score this application: ...”

---

📦 Folder Structure

.
├── ai-agent-community-assistant.ipynb   # Main development notebook
├── community-docs/                          # Directory with PDF docs
├── scored_applications.csv                  # Output file with scored apps
└── README.md                                # This file

---

✅ Results

• Built and deployed a fully functional GenAI agent from scratch.
• Solidified understanding of agents, embeddings, RAG, and prompt engineering.
• Learned LangGraph by implementing custom nodes and conditional routing.
• Wrapped everything into an interactive, usable UI using Gradio.

---

📚 Background

This project was created as a follow-up to the Google x Kaggle Generative AI Intensive (Q1 2025) — a 5-day, hands-on course on LLMs, embeddings, agent tools, and MLOps.

---

✨ Try It or Extend It

Fork this project or run the notebook on Kaggle to:

• Add more tools or community documents
• Score applicants using different criteria
• Extend the UI with buttons, file uploads, or user feedback

---

🙌 Credits

Built by [Riad Bensalem] — feel free to reach out or connect on LinkedIn.