The ChatGPT Retrieval Plugin is a standalone retrieval backend written in Python that enables semantic search and retrieval of personal or organizational documents through natural language queries. It serves as a flexible solution for integrating document retrieval capabilities with ChatGPT custom GPTs, function calling via the chat completions and assistants APIs, or the deprecated ChatGPT plugins model. The plugin leverages OpenAI's embeddings model, specifically text-embedding-3-large with 256 dimension embeddings by default, to generate embeddings of document chunks which are then stored and queried using a vector database backend.
The repository is comprehensively organized into multiple functional directories. The datastore directory contains core logic for storing and querying document embeddings across various vector database providers. The docs directory provides setup and usage documentation for each supported vector database provider, webhooks, and dependency management. Examples directory offers configuration samples, authentication methods, and provider-specific implementations. The local_server directory contains a localhost testing implementation, while models houses the data structures for documents and metadata. Scripts are provided for processing and uploading documents from different data sources, and the server directory implements the main FastAPI server. Services handle utility functions like document chunking, metadata extraction, and PII detection. Tests include integration tests for various vector database providers, and the .well-known directory stores the plugin manifest and OpenAPI schema defining the plugin configuration and API specification.
The plugin supports an extensive array of vector database providers including Pinecone, Elasticsearch, MongoDB Atlas, Weaviate, Zilliz, Milvus, Qdrant, Redis, Llama Index, Chroma, Azure Cognitive Search, Azure CosmosDB Mongo vCore, Supabase, Postgres, and AnalyticDB, allowing developers to select their preferred backend. A FastAPI server exposes endpoints for upserting, querying, and deleting documents, with users able to refine search results through metadata filters by source, date, author, or other criteria. The plugin implements retrieval-augmented generation to enable models to write answers informed by relevant retrieved context.
For deployment, the plugin can be hosted on cloud platforms supporting Docker containers such as Fly.io, Heroku, Render, or Azure Container Apps. It supports continuous document updates through incoming webhooks to upsert and delete endpoints, with tools like Zapier or Make helping configure webhooks based on events or schedules. The plugin offers multiple authentication methods including API key authentication with Basic or Bearer tokens and OAuth support.
According to GitGenius activity tracking, the repository shows a median issue and pull request response latency of 0.0 hours across 26 tracked items, with a mean latency of 2857.3 hours. The most active contributors tracked include jymmyjr with 4 events, caseyclements with 2 events, and Coder-Krishna-189251 with 1 event. The repository shares overlapping contributors with github/gh-aw, meteor/meteor, and langchain-ai/langchain, indicating cross-project collaboration within the broader development ecosystem.
