n8n Assignment Report: Chatmate

Course: GenAI Assignment: 1 Date: 2025-11-25

1. Workflow Overview

This project implements a “Chatmate” system using n8n, which allows users to interactively query a PDF document (“Attention Is All You Need”). The system is built on a Retrieval-Augmented Generation (RAG) architecture, utilizing Supabase for vector storage and Groq for LLM inference.

The workflow consists of two distinct flows:

1. Document Ingestion Flow: Handles the loading, text extraction, embedding generation, and storage of the PDF content.
2. Conversational Retrieval Flow: Handles user queries, retrieves relevant context from the vector store, and generates answers using the LLM.

2. Node Configuration & Implementation Details

Flow 1: Document Ingestion

This flow is designed to be run once to populate the vector database.

• Trigger: Manual Trigger (“When clicking Execute Workflow”) - Initiates the ingestion process manually.
• Read File: Read/Write Files from Disk - Reads the PDF file from the local file system.
  • Path: /media/2tbhdd/LinuxFiles/SanctuaryObs/Sanctuary/University/WiSe25/GenAI/Ass/n8n/Attention_Is_All_You_Need.pdf
• Text Extraction: Extract Text (PDF Parser) - Extracts raw text content from the binary PDF data.
• Embeddings: HuggingFace Inference - Generates vector embeddings for the extracted text chunks.
  • Model: sentence-transformers/distilbert-base-nli-mean-token
  • Output Dimension: 768
• Vector Storage: Supabase Vector Store - Inserts the text chunks and their corresponding embeddings into the Supabase documents table.

Flow 2: Conversational Retrieval

This flow runs for every user interaction in the chat interface.

• Trigger: Chat Trigger (“When chat message received”) - Captures the user’s input from the chat UI.
• Orchestrator: AI Agent - The central node that manages the conversation flow, calls tools, and invokes the LLM.
  • System Prompt: Configured to act as a helpful assistant answering based on the provided context.
• LLM: Groq Chat Model - Provides the reasoning and generation capabilities.
  • Model: llama3-8b-8192 (Selected for speed and performance).
• Retrieval Tool: Supabase Vector Store - Connected as a tool to the AI Agent. It allows the agent to perform semantic searches against the stored PDF content to find relevant information.
• Memory: Simple Memory (Window Buffer) - Maintains the context of the conversation, allowing for follow-up questions.

3. Challenges & Solutions

• Challenge: Ensuring the PDF was accessible to the n8n instance (Docker container).
  • Solution: I ensured the file path was correctly mounted and accessible. In a real Docker setup, I would use a bind mount (e.g., -v /local/path:/n8n/path) to map the local directory to the container’s file system.
• Challenge: Selecting the right embedding model.
  • Solution: I used sentence-transformers/distilbert-base-nli-mean-token as recommended, which offers a good balance between performance and embedding quality for general English text.
• Challenge: Configuring the Supabase connection.
  • Solution: I ensured the table schema in Supabase matched the expected format (id, content, metadata, embedding) and that the vector extension was enabled (create extension vector;).

4. Conclusion

The implemented workflow successfully demonstrates a RAG pipeline. The separation of ingestion and retrieval allows for efficient querying, as the heavy lifting of processing the document is done only once. The use of Groq ensures low-latency responses, making the chat experience smooth and interactive.