Cost-Optimized Retrieval Augmented Generation (RAG) Setup - n8n Workflow

AI developers looking for advanced n8n templates for RAG.
Users who need to implement custom knowledge bases using vector stores.
Anyone seeking to reduce operational costs associated with powerful LLMs like GPT-4 in production RAG systems.
Technical users deploying custom language models and AI agents via an n8n workflow.

This powerful n8n workflow demonstrates a crucial technique for cost optimization in Retrieval-Augmented Generation (RAG) applications. It is split into two logical flows: a Data Loading flow and a Retriever flow. The core value of this n8n template lies in its intelligent use of two separate OpenAI models. The initial AI Agent reasoning and tool selection utilize a more powerful, 'Expensive model' (e.g., GPT-4) for high-quality decision-making. However, the heavy lifting of answering the query based on retrieved documents—which is called frequently—is delegated to a 'Cheap Model' (e.g., GPT-4-mini or similar low-cost model). This strategic separation significantly reduces token consumption costs while maintaining high quality results for the entire n8n workflow. The flow starts with an n8n trigger for data upload and uses a separate n8n trigger for the chat interface.

This n8n workflow operates in two distinct phases:

1. Data Loading (RAG Setup):

1. Retriever Flow (Q&A and Cost Optimization):

To deploy this comprehensive n8n workflow, follow these steps:

1. Import the n8n template: Copy the provided JSON and import it into your n8n instance via the 'Workflows' section, using the 'New' button and selecting 'Import from JSON'.


2. Set up OpenAI Credentials: You must configure credentials for the OpenAI Service. Both the Embeddings OpenAI, Expensive model, and Cheap Model n8n nodes require an active OpenAI API key credential (ID '14' in the provided JSON).


3. Run the Load Data Flow: Click 'Execute Workflow' to activate the Upload your file here n8n trigger and provide your document(s) (PDF or CSV).


4. Run the Retriever Flow: Once data is loaded, open the chat interface associated with the When chat message received n8n trigger and start asking questions about the document content. The overall n8n workflow is now active and ready to use.

Upload your file here (n8n trigger): A Form Trigger that initiates the data loading pipeline. It is configured to accept .pdf and .csv files.
Embeddings OpenAI: Generates high-dimensional vector embeddings for the documents using the OpenAI service, essential for RAG.
Insert Data to Store: A LangChain vectorStoreInMemory n8n node. This stores the document chunks and their embeddings in a temporary, in-memory location for quick retrieval using the key vectorstorekey.
When chat message received (n8n trigger): A Chat Trigger that listens for incoming messages, kicking off the Q&A process. This is the primary n8n trigger for the retrieval flow.
Expensive model (LangChain LM): Uses a powerful model (gpt-4.1 placeholder) to provide high-quality reasoning to the AI Agent for complex tool use and planning.
AI Agent: The central decision-making n8n node. It determines when and how to use the available tools, driven by the Expensive model.
Answer questions with a vector store (LangChain Tool): A custom tool exposed to the AI Agent. Crucially, this tool's internal retrieval and response generation uses the designated Cheap Model (gpt-4.1-mini placeholder), enabling the cost-saving strategy of this specific n8n workflow.
Query Data Tool: Retrieves relevant document snippets from the in-memory vector store based on the user query embedding.