An advanced Retrieval-Augmented Generation (RAG) system built with n8n, Supabase Vector Store, OpenAI, and NVIDIA embeddings.

Instead of relying on a single retrieval query, the system decomposes complex questions into multiple targeted sub-queries, retrieves relevant knowledge-base chunks, filters low-quality results, and performs a dedicated reasoning step before generating the final answer.

Basic RAG systems can struggle when users ask complex questions containing multiple concepts or requirements.

Common problems include:

Poor retrieval from a single search query

Irrelevant context reaching the LLM

Low-quality chunks influencing answers

Increased hallucination risk

Difficult-to-maintain retrieval logic as the system grows

I designed a multi-query RAG architecture that separates question decomposition, retrieval, relevance filtering, reasoning, and final generation into distinct stages.

The system receives the user's question.

GPT-5-mini decomposes complex questions into 1–5 targeted sub-queries.

Each sub-query searches the Supabase vector database.

Retrieved chunks are evaluated using a similarity threshold.

Results scoring below 0.4 are removed.

Remaining chunks are aggregated into a unified context.

A dedicated reasoning step synthesizes the retrieved information.

The final response is generated using the filtered knowledge-base context.

The query decomposition stage allows the system to approach complex questions from multiple retrieval angles.

For example, instead of attempting to answer a multi-part question through one semantic search, the agent can create several focused sub-queries and retrieve context independently for each one.

This improves the quality and coverage of the retrieved context before generation.

Retrieved chunks are not automatically passed to the final model.

The workflow applies a 0.4 similarity threshold, removing lower-relevance results before they become part of the final context.

This creates an additional quality-control layer between vector retrieval and answer generation.

After retrieval and filtering, the surviving context is passed through a dedicated reasoning stage.

This step synthesizes information from multiple retrieved chunks before the final answer is generated.

The architecture keeps retrieval and reasoning logically separated, making the system easier to modify, test, and scale.

The system is designed to prioritize information contained in the knowledge base.

If relevant information cannot be retrieved, the workflow can return an explicit insufficient-information response rather than encouraging the model to guess.

Conversational memory also maintains context across user queries.

90% improvement in answer quality compared with the basic RAG implementation, based on project validation

More targeted retrieval for complex questions

Improved context quality through relevance filtering

Reusable architecture for large knowledge bases

Designed to support knowledge bases containing millions of vectors

Approximately 100–500 queries/day in the demonstrated configuration

n8n — AI workflow orchestration

OpenAI GPT-5-mini — query decomposition and generation

Supabase Vector Store — semantic retrieval

NVIDIA Embeddings — vector embeddings

JavaScript — data processing and workflow logic

Build Type: Self-initiated demonstration build Role: AI Automation Engineer & Workflow Developer Platform: n8n Industry: AI / Knowledge Management