Build an advanced multi-query RAG system with Supabase and GPT-5

Go beyond basic Retrieval-Augmented Generation (RAG) with this advanced template. While a simple RAG setup can answer straightforward questions, it often fails when faced with complex queries and can be polluted by irrelevant information. This workflow introduces a sophisticated architecture that empowers your AI agent to think and act like a true research assistant.

By decoupling the agent from the knowledge base with a smart sub-workflow, this template enables multi-query decomposition, relevance-based filtering, and an intermediate reasoning step. The result is an AI agent that can handle complex questions, filter out noise, and synthesize high-quality, comprehensive answers based on your data in Supabase.

Who is this for?

• AI and automation developers: Anyone building sophisticated Q&A bots, internal knowledge base assistants, or complex research agents.
• n8n power users: Users looking to push the boundaries of AI agents in n8n by implementing production-ready, robust architectural patterns.
• Anyone building a RAG system: This provides a superior architectural pattern that overcomes the common limitations of basic RAG setups, leading to dramatically better performance.

What problem does this solve?

• Handles complex questions: A standard RAG agent sends one query and gets one set of results. This agent is designed to break down a complex question like "How does natural selection work at the molecular, organismal, and population levels?" into multiple, targeted sub-queries, ensuring all facets of the question are answered.
• Prevents low-quality answers: A simple RAG agent can be fed irrelevant information if the semantic search returns low-quality matches. This workflow includes a crucial relevance filtering step, discarding any data chunks that fall below a set similarity score, ensuring the agent only reasons with high-quality context.
• Improves answer quality and coherence: By introducing a dedicated "Think" tool, the agent has a private scratchpad to synthesize the information it has gathered from multiple queries. This intermediate reasoning step allows it to connect the dots and structure a more comprehensive and logical final answer.
• Gives you more control and flexibility: By using a sub-workflow to handle data retrieval, you can add any custom logic you need (like filtering, formatting, or even calling other APIs) without complicating the main agent's design.

How it works

This template consists of a main agent workflow and a smart sub-workflow that handles knowledge retrieval.

1. Multi-query decomposition: When you ask the AI Agent a complex question, its system prompt instructs it to first break it down into an array of multiple, simpler sub-queries.
2. Decoupling with a sub-workflow: The agent doesn't have direct access to the vector store. Instead, it calls a "Query knowledge base" tool, which is a sub-workflow. It sends the entire array of sub-queries to this sub-workflow in a single tool call.
3. Iterative retrieval & filtering (in the sub-workflow): The sub-workflow loops through each sub-query. For each one, it queries your Supabase Vector Store. It then checks the similarity score of the returned data chunks and uses a Filter node to discard any that are not highly relevant (the default is a score > 0.4).
4. Intermediate reasoning step: The sub-workflow returns all the high-quality, filtered information to the main agent. The agent is then instructed to use its Think tool to review this information, synthesize the key points, and structure a plan for its final, comprehensive answer.

Setup

1. Connect your accounts:
   • Supabase: In the sub-workflow ("RAG sub-workflow"), connect your Supabase account to the Supabase Vector Store node and select your table.
   • OpenAI: Connect your OpenAI account in two places: to the Embeddings OpenAI node (in the sub-workflow) and to the OpenAI Chat Model node (in the main workflow).
2. Customize the agent's purpose: In the main workflow, edit the AI Agent's system prompt. Change the context from a "biology course" to whatever your knowledge base is about.
3. Adjust the relevance filter: In the sub-workflow, you can change the 0.4 threshold in the Filter node to be more or less strict about the quality of the information you want the agent to use.
4. Activate the workflow and start asking complex questions!

Taking it further

• Integrate different vector stores: The logic is decoupled. You can easily swap the Supabase Vector Store node in the sub-workflow with a Pinecone, Weaviate, or any other vector store node without changing the main agent's logic.
• Add more tools: Give the main agent other capabilities, like a web search a way to interact with your tech stack. The agent can then decide whether to use its internal knowledge base, search the web, or both, to answer a question.
• Better prompting: You could further work on the Agent's system prompt to increase its capacity to provide high-quality answers by being even better at leveraging the provided chunks.