Building a RAG AI Assistant with n8n, Google Drive, and Supabase

Building the Knowledge Base

A useful AI assistant needs more than an LLM. If the assistant has to answer questions about a specific set of documents, it needs access to reliable information and a way to retrieve the relevant context before generating a response.

This workflow was built around n8n, Google Drive, Supabase, and an LLM. The goal was to turn documents from a Google Drive folder into a searchable knowledge base that an AI assistant could use when answering questions.

The first part of the workflow handles document ingestion. Files are downloaded from Google Drive and converted when necessary. Their contents are extracted and cleaned before being prepared for further processing. The workflow also includes a text-cleaning stage to make the extracted content more consistent.

From Documents to Vector Search

Sending complete documents to an LLM for every question would quickly become inefficient. Instead, the workflow converts document content into embeddings and stores those vectors in a Supabase vector database.

When a document is processed, its text is prepared for embedding and inserted into the vector store. At query time, the user's question can be converted into a vector and matched against the stored document content. The most relevant information can then be supplied to the language model as context.

This is the core idea behind retrieval-augmented generation, or RAG: the model does not need to memorize the entire knowledge base. It retrieves relevant information at the time of the request.

The AI Assistant

The second part of the workflow connects the knowledge base to an AI agent.

The agent receives the user's question and can query the company knowledge base before generating an answer. The workflow also includes conversation memory, allowing the assistant to maintain context across interactions.

The language model is responsible for interpreting the question and producing the final response, while the knowledge base provides the information the response should be based on.

This separation is important. The LLM handles reasoning and language generation, while the retrieval layer handles access to the actual source material.

Keeping the Knowledge Base Up to Date

The workflow also supports updating the knowledge base when source documents change. For example, a GitHub webhook can trigger processing when relevant documentation is modified, allowing updated content to be retrieved and processed instead of relying on an outdated copy.

This makes the system closer to a maintainable knowledge pipeline rather than a one-time document import.

Why n8n Fits This Architecture

n8n acts as the orchestration layer between the different components.

Instead of building separate services for document ingestion, vector storage, model calls, and automation logic, the workflow connects these components in one visual pipeline. This makes it easier to inspect individual steps, add validation, handle errors, and change individual integrations without rebuilding the entire system.

For example, the workflow includes database health checks and validation before continuing with the AI processing pipeline.

Trade-offs

RAG does not automatically make an AI assistant accurate. Retrieval quality depends on document extraction, chunking, embeddings, and the quality of the search results.

There is also additional infrastructure compared with a simple LLM chatbot. Documents need to be processed, embeddings need to be generated and stored, and every query adds a retrieval step before the model can produce its response.

For a small amount of static information, a simpler approach may be enough. RAG becomes more useful when the knowledge base grows and the assistant needs to work with information that changes independently of the model.

Practical Takeaway

The important part of this architecture is not any individual tool. It is the separation between knowledge, retrieval, orchestration, and language generation.

Google Drive provides the source documents, n8n orchestrates the workflow, Supabase provides storage and vector search, and the LLM turns retrieved information into a useful response.

That architecture provides a practical foundation for building AI assistants that need to work with private or constantly changing information without putting the entire knowledge base directly into every model request.