n8n PDF-based RAG System with OpenAI, Pinecone, and Cohere Reranking

This n8n workflow demonstrates a robust Retrieval-Augmented Generation (RAG) system designed to answer user queries based on a corpus of PDF documents. It leverages OpenAI for embeddings and chat, Pinecone as a vector store for document retrieval, and Cohere for reranking search results to improve relevance.

What it does

This workflow orchestrates a sophisticated RAG process:

1. Triggers on Chat Message: The workflow starts when a new chat message is received, likely containing a user's question or query.
2. Loads Documents: It uses a "Default Data Loader" to process and load documents. While the JSON doesn't specify the source (e.g., PDF files), this node is typically configured to ingest data from various sources.
3. Splits Text: The loaded documents are then broken down into smaller, manageable chunks using a "Recursive Character Text Splitter." This is crucial for efficient embedding and retrieval.
4. Generates Embeddings: OpenAI's Embeddings node converts these text chunks into numerical vector representations, capturing their semantic meaning.
5. Stores and Retrieves Vectors: The "Pinecone Vector Store" is used to store these embeddings. When a query comes in, it's also embedded, and Pinecone is queried to find the most semantically similar document chunks.
6. Reranks Results: The "Reranker Cohere" node takes the initial search results from Pinecone and re-evaluates their relevance to the original query, using Cohere's advanced reranking capabilities to provide a more accurate set of top results.
7. Manages Chat Memory: A "Simple Memory" node (Buffer Window Memory) maintains the context of the conversation, allowing the AI agent to remember previous turns in the chat.
8. Orchestrates AI Agent: An "AI Agent" node (LangChain Agent) acts as the central intelligence, using the OpenAI Chat Model, the Pinecone Vector Store (for retrieval), and the Cohere Reranker to formulate a comprehensive answer based on the retrieved and reranked document chunks and the ongoing conversation.

Prerequisites/Requirements

To use this workflow, you will need:

• n8n Instance: A running n8n instance.
• OpenAI API Key: For generating embeddings and powering the chat model.
• Pinecone Account & API Key: For setting up and interacting with your vector database.
• Cohere API Key: For utilizing the Cohere reranking service.
• Documents: A collection of documents (e.g., PDFs, text files) that you want your RAG system to query. The "Default Data Loader" will need to be configured to access these.

Setup/Usage

1. Import the Workflow:
   • Download the provided JSON file.
   • In your n8n instance, go to "Workflows" and click "New Workflow".
   • Click the three dots menu (...) in the top right and select "Import from JSON".
   • Paste the JSON content or upload the file.
2. Configure Credentials:
   • Locate the "Embeddings OpenAI" and "OpenAI Chat Model" nodes and configure your OpenAI API credentials.
   • Locate the "Pinecone Vector Store" node and configure your Pinecone API credentials (API Key, Environment, and Index Name).
   • Locate the "Reranker Cohere" node and configure your Cohere API credentials.
3. Configure Document Loading:
   • Adjust the "Default Data Loader" node to point to your specific document source (e.g., local files, S3 bucket, Google Drive). You will need to ensure n8n has access to these documents.
   • Ensure the "Recursive Character Text Splitter" is configured appropriately for your document structure.
4. Activate the Workflow:
   • Save the workflow.
   • Activate the workflow by toggling the "Active" switch in the top right corner.
5. Interact:
   • Use the "When chat message received" trigger to send queries to your RAG system. This typically involves integrating with a chat platform (e.g., Slack, Discord, custom chat UI) that can send messages to n8n's webhook URL for this trigger.