n8n Workflow: Document-Based RAG System with Mistral, OCR, Qdrant, and Gemini AI

This n8n workflow demonstrates a robust Retrieval Augmented Generation (RAG) system designed to process PDF documents, extract text using OCR, embed the content, store it in a vector database (Qdrant), and then use a Large Language Model (Google Gemini) to answer questions based on the document's content. It allows for both initial document processing and subsequent question-answering.

What it does

This workflow orchestrates the following steps:

1. Trigger: The workflow can be initiated manually or by an external sub-workflow call.
2. Document Retrieval: It fetches a PDF document from Google Drive.
3. OCR Processing (Placeholder): An HTTP Request node is included, likely intended to call an OCR service (e.g., Mistral AI, as hinted by the directory name) to extract text from the PDF.
4. Text Preparation: The extracted text is then processed:
   • Edit Fields (Set): Prepares the data for further processing.
   • Token Splitter: Divides the document text into smaller, manageable chunks (tokens) suitable for embedding.
5. Embedding Generation:
   • Embeddings OpenAI: Generates vector embeddings for each text chunk using the OpenAI embeddings model.
6. Vector Store Management (Qdrant):
   • Qdrant Vector Store: Stores the generated embeddings in a Qdrant vector database, making the document content searchable.
7. Question Answering (RAG):
   • Vector Store Retriever: When a chat message (question) is received, this node retrieves relevant document chunks from Qdrant based on the question's embedding.
   • Google Gemini Chat Model: Utilizes the Google Gemini LLM to generate an answer, augmented by the retrieved document chunks.
   • Question and Answer Chain: Orchestrates the retrieval and generation process to provide a coherent answer.
8. Summarization (Optional): A Summarization Chain node is present, suggesting the capability to summarize documents or retrieved content.
9. Sub-workflow Execution: The Execute Sub-workflow node enables modularity, allowing parts of this process to be called from other workflows.
10. Wait: A Wait node is included, potentially for rate limiting or pausing between operations.
11. Code Node: A Code node is present, offering flexibility for custom JavaScript logic.

Prerequisites/Requirements

To use this workflow, you will need:

• n8n Instance: A running n8n instance.
• Google Drive Account: Configured credentials for Google Drive to access PDF documents.
• OCR Service: Access to an OCR service (e.g., Mistral AI, not directly configured in this JSON but implied). The HTTP Request node would need to be configured for this.
• OpenAI API Key: Credentials for OpenAI to generate text embeddings.
• Qdrant Instance: A running Qdrant vector database instance and its connection details.
• Google Gemini API Key: Credentials for Google Gemini to power the chat model.
• Langchain Nodes: Ensure the @n8n/n8n-nodes-langchain package is installed in your n8n instance.

Setup/Usage

1. Import the Workflow: Import the provided JSON into your n8n instance.
2. Configure Credentials:
   • Set up your Google Drive credentials.
   • Configure your OpenAI credentials for the Embeddings OpenAI node.
   • Set up your Qdrant credentials for the Qdrant Vector Store node.
   • Configure your Google Gemini credentials for the Google Gemini Chat Model node.
3. OCR Service Configuration: Update the HTTP Request node (ID 19) with the endpoint and authentication details for your chosen OCR service (e.g., Mistral AI).
4. Activate the Workflow: Once all credentials and configurations are set, activate the workflow.
5. Triggering:
   • Manual: Click "Execute Workflow" on the Manual Trigger node to test the document processing.
   • Sub-workflow: Call the Execute Workflow Trigger from another n8n workflow.
   • Chat: Use the Chat Trigger to initiate a question-answering session once documents are processed.

This workflow provides a powerful foundation for building intelligent document-based RAG applications.