n8n Workflow: Agentic RAG with Qdrant, Mistral, and OCR for Community Contest

This n8n workflow demonstrates an advanced Retrieval-Augmented Generation (RAG) system, designed for the Agentic Arena Community Contest. It leverages Qdrant for vector storage, Mistral (via OpenAI Chat Model) for language generation, and OCR capabilities to process documents, providing a robust solution for querying information from various sources.

What it does

This workflow orchestrates a sophisticated RAG pipeline, enabling intelligent querying of document content.

1. Trigger: The workflow can be initiated manually or as a sub-workflow call, making it flexible for integration into larger systems or for standalone testing.
2. Document Loading: It's designed to process documents, likely from Google Drive, using a "Default Data Loader" which implies OCR capabilities to extract text from various file types (e.g., PDFs, images).
3. Text Splitting: The extracted document text is then split into manageable chunks using a "Character Text Splitter" to prepare it for embedding.
4. Embedding Generation: "Embeddings OpenAI" is used to convert these text chunks into vector embeddings, which capture the semantic meaning of the text.
5. Vector Storage: These embeddings are stored and managed in a "Qdrant Vector Store," a high-performance vector database, enabling efficient similarity searches.
6. Query Processing (AI Agent): An "AI Agent" (likely configured with a ReAct or Plan-and-Execute approach) handles incoming queries. It uses a "Simple Memory" to maintain conversational context.
7. Language Model: The "OpenAI Chat Model" (acting as the Mistral model through an OpenAI-compatible API) is used by the AI Agent for generating responses based on retrieved information.
8. Reranking (Optional but present): A "Reranker Cohere" node suggests the capability to re-rank retrieved documents, improving the relevance of information passed to the language model.
9. Evaluation: The workflow includes "Evaluation Trigger" and "Evaluation" nodes, indicating it's set up for testing and measuring the performance of the RAG system.
10. HTTP Request: An "HTTP Request" node is present, suggesting external API calls are made, possibly to an OCR service, a custom Mistral API endpoint, or other external tools.
11. Data Manipulation: "Edit Fields (Set)" and "Code" nodes are used for data transformation and custom logic within the workflow.
12. Looping: A "Loop Over Items (Split in Batches)" node allows for processing multiple documents or queries in batches.
13. Conditional Logic: A "Filter" node enables conditional branching based on specific criteria.
14. Delay: A "Wait" node can introduce pauses in the workflow, useful for rate limiting or waiting for external processes.

Prerequisites/Requirements

To run this workflow, you will need:

• n8n Instance: A running n8n instance.
• OpenAI API Key: For the "Embeddings OpenAI" and "OpenAI Chat Model" nodes. This is likely used to access Mistral via an OpenAI-compatible API endpoint.
• Qdrant Instance: Access to a Qdrant vector database instance.
• Google Drive Account: If documents are sourced from Google Drive.
• Cohere API Key: If the "Reranker Cohere" node is actively used.
• OCR Service: The "Default Data Loader" implies an underlying OCR capability, which might require an API key or configuration for a specific OCR service (e.g., Google Cloud Vision, Mindee, etc., though not explicitly configured in this JSON).
• Custom API Endpoints: Depending on the "HTTP Request" node's configuration, additional API keys or access to custom services might be required.

Setup/Usage

1. Import the Workflow: Import the provided JSON into your n8n instance.
2. Configure Credentials:
   • Set up your OpenAI API key credentials for the "Embeddings OpenAI" and "OpenAI Chat Model" nodes.
   • Configure your Qdrant credentials for the "Qdrant Vector Store" node.
   • If using Google Drive, set up your Google Drive credentials.
   • If using Cohere, configure your Cohere API key.
   • Ensure any necessary OCR service credentials are set up if the "Default Data Loader" relies on an external service.
3. Customize Nodes:
   • Adjust the "Google Drive" node to point to your specific document folders or files.
   • Modify the "Character Text Splitter" parameters (e.g., chunk size, overlap) as needed for your documents.
   • Configure the "Qdrant Vector Store" with your desired collection name and other settings.
   • Review and adjust the "AI Agent" and "OpenAI Chat Model" settings, including the model name (e.g., for Mistral), temperature, and system prompts.
   • If using, configure the "Reranker Cohere" according to your needs.
   • Examine the "HTTP Request" and "Code" nodes for any custom logic or external API calls that require specific configuration.
4. Activate and Test:
   • Activate the workflow.
   • Use the "Manual Trigger" to test the workflow with sample inputs, or integrate it as a sub-workflow into your main application.
   • Utilize the "Evaluation Trigger" and "Evaluation" nodes to assess the RAG system's performance.