The RAG backend OpenAI admits you probably don't need

What it does The ChatGPT Retrieval Plugin is a FastAPI server that turns a vector database into a retrieval backend for ChatGPT custom GPTs and the Assistants API. It chunks your documents, embeds them with OpenAI models like text-embedding-3-large, and stores them in a provider of your choice so ChatGPT can answer questions grounded in your data. In short, it is a self-hosted RAG pipeline dressed up as a plugin.

The interesting bit The README itself concedes that ChatGPT and the Assistants API already handle file retrieval natively, which makes this plugin’s existence almost apologetic. Its real value is granularity: you can pick from roughly fifteen vector databases, tune chunk length, and swap embedding dimensions—things you cannot do inside ChatGPT’s walled garden. It is less a missing feature and more an escape hatch for the paranoid or the picky.

Key highlights

• Supports a sprawling list of vector stores: Pinecone, Weaviate, Qdrant, Redis, Postgres, Elasticsearch, Chroma, and others.
• Ships with utilities for text chunking, metadata extraction, and PII detection in the services layer.
• Exposes a standard FastAPI server with an OpenAPI schema for hooking into custom GPTs or function calling.
• Defaults to text-embedding-3-large at 256 dimensions, but lets you swap models and sizes.
• Open source and self-hosted, meaning you keep the data plane under your own infrastructure.

Caveats

• OpenAI explicitly recommends using native ChatGPT or Assistants API file retrieval unless you need the extra control this backend provides.
• The original ChatGPT plugins model that this was built for is now deprecated.

Verdict Reach for this if you are an enterprise that needs to keep documents inside a specific vector database or wants to tune retrieval mechanics. Everyone else should probably just upload files to ChatGPT and call it a day.