Answer Questions from Documents with RAG using Supabase, OpenAI, and Cohere Reranker

This n8n workflow demonstrates a sophisticated Retrieve-and-Generate (RAG) system for answering user questions based on a corpus of documents. It leverages Supabase as a vector store, OpenAI for embeddings and chat models, and Cohere for reranking search results, providing highly relevant and accurate answers.

What it does

This workflow automates the process of ingesting documents, vectorizing their content, storing them in a Supabase vector database, and then using this knowledge base to answer user questions with enhanced relevance.

1. Triggers on Chat Message: The workflow is activated when a new chat message (user question) is received.
2. Ingests Documents from Google Drive (Manual Trigger):
   • A manual trigger allows for on-demand execution to process documents.
   • It retrieves files from a specified Google Drive folder.
   • Files are extracted and loaded using a default data loader.
   • OpenAI Embeddings are generated for the document content.
   • These embeddings and the document content are stored in a Supabase Vector Store.
3. Processes User Questions (Chat Trigger):
   • When a user asks a question via the "Chat Trigger", the AI Agent is activated.
   • The AI Agent uses an OpenAI Chat Model (via OpenRouter) to understand the query.
   • It performs a similarity search against the Supabase Vector Store to retrieve relevant document chunks.
   • A Cohere Reranker is applied to the search results to improve the relevance of the retrieved documents.
   • The reranked documents are then used as context by the OpenAI Chat Model to generate a comprehensive answer to the user's question.

Prerequisites/Requirements

To use this workflow, you will need:

• n8n Account: A running n8n instance (cloud or self-hosted).
• Google Drive Account: To store the documents you want to use as a knowledge base.
• OpenAI API Key: For generating embeddings and powering the chat model.
• Supabase Account: To host the vector database for document embeddings.
• Cohere API Key: For the reranking functionality to improve search relevance.
• OpenRouter API Key (Optional but recommended): To access OpenAI models through OpenRouter, potentially offering more flexibility or cost-effectiveness.

Setup/Usage

1. Import the workflow: Download the JSON definition and import it into your n8n instance.
2. Configure Credentials:
   • Set up your Google Drive credentials.
   • Set up your OpenAI credentials for both the Embeddings OpenAI node and the OpenRouter Chat Model (if using OpenRouter, otherwise configure OpenAI directly in the AI Agent).
   • Set up your Supabase credentials (API URL, API Key, and table name for the vector store).
   • Set up your Cohere credentials for the Reranker Cohere node.
3. Ingest Documents:
   • Open the workflow.
   • Locate the Google Drive node and configure it to point to the folder containing your source documents.
   • Execute the workflow manually by clicking "Execute workflow" on the Manual Trigger node. This will process your documents and populate the Supabase vector store.
   • Note: This part of the workflow is designed for initial document ingestion. You might want to disable the Google Drive branch after initial setup or integrate it with a separate trigger for continuous updates.
4. Start Chatting:
   • Activate the workflow.
   • Use the Chat Trigger (e.g., via the n8n chat interface or by sending a message to the configured chat platform) to ask questions. The AI Agent will retrieve relevant information from your Supabase knowledge base and provide an answer.