AI-Powered RAG Chatbot for Knowledge Retrieval

To explore AI integration in product workflows, I built a no-code RAG (Retrieval-Augmented Generation) chatbot that provides context-aware answers from files stored in Google Drive. This project was inspired by Paweł Huryn's AI PM newsletter and aimed at building practical intuition for managing AI-powered products.

🌟 Objective

Create a chatbot that:

• Retrieves relevant content from uploaded files (PDF, DOCX, etc.)

• Responds using an LLM (GPT-4o-mini)

• Refuses to answer if it can't find relevant content

• Requires no code and integrates into my PM portfolio

🛠️ Tech Stack

🧩 How It Works

The system is divided into two primary workflows using n8n, with third-party integrations across OpenAI, Pinecone, Google Drive, and Lovable. Here's a breakdown of how everything is connected:

1. ✨ Document Ingestion & Vectorization Workflow

1. A user uploads files (PDFs, DOCX, TXT) to a dedicated Google Drive folder.

2. n8n workflow is triggered by a "new file" event from Google Drive.

3. The document is downloaded and split into manageable text chunks (e.g., 500-800 characters).

4. Each chunk is converted into a vector embedding using OpenAI's text-embedding-3-small model.

5. Chunks are pushed into a Pinecone index for vector storage.

   
   

   

   
   

2. ❓ User Query & Response Generation Workflow

1. A user types a query into the chatbot built in Lovable.dev.

2. The chatbot sends a request to a public n8n webhook URL with the query as a parameter.

3. The n8n flow converts the query into an embedding vector.

4. Pinecone is queried to retrieve the top matching document chunks.

5. The top chunks are passed as context to the GPT-4o-mini model along with the original query.

6. GPT-4o-mini generates a natural language response, citing relevant chunks if available.

7. The chatbot returns the answer to the user.

Additional Notes:

• If no relevant chunks meet a similarity threshold, the bot politely responds: "Sorry, I couldn't find relevant information."

• Lovable’s markdown-based input interface makes it easy to style responses.

• The webhook is production-deployed via n8n.cloud, ensuring live connectivity between UI and backend.

🧪 Result

• End-to-end system built in under 1 hour

• Successfully retrieved accurate answers from documents

• Demonstrated how to manage vector data pipelines and LLM integration

• Reusable framework for onboarding chatbots, customer support bots, and knowledge agents

📊 Skills Demonstrated

• RAG architecture and application

• No-code stack orchestration

• Practical use of LLMs in real product settings

• Product intuition around limitations of token size, latency, and fallback handling

• Collaboration across data, design, and AI tool chains

📋 Next Steps

• Expand use case for internal PM documentation retrieval

• Add source citation in response output

• Combine with fine-tuned model for specialised product reasoning