Home

Awesome

ChatGPT Retrieval Plugin

Build Custom GPTs with a Retrieval Plugin backend to give ChatGPT access to personal documents. Example Custom GPT Screenshot

Introduction

The ChatGPT Retrieval Plugin repository provides a flexible solution for semantic search and retrieval of personal or organizational documents using natural language queries. It is a standalone retrieval backend, and can be used with ChatGPT custom GPTs, function calling with the chat completions or assistants APIs, or with the ChatGPT plugins model (deprecated). ChatGPT and the Assistants API both natively support retrieval from uploaded files, so you should use the Retrieval Plugin as a backend only if you want more granular control of your retrieval system (e.g. document text chunk length, embedding model / size, etc.).

The repository is organized into several directories:

DirectoryDescription
datastoreContains the core logic for storing and querying document embeddings using various vector database providers.
docsIncludes documentation for setting up and using each vector database provider, webhooks, and removing unused dependencies.
examplesProvides example configurations, authentication methods, and provider-specific examples.
local_serverContains an implementation of the Retrieval Plugin configured for localhost testing.
modelsContains the data models used by the plugin, such as document and metadata models.
scriptsOffers scripts for processing and uploading documents from different data sources.
serverHouses the main FastAPI server implementation.
servicesContains utility services for tasks like chunking, metadata extraction, and PII detection.
testsIncludes integration tests for various vector database providers.
.well-knownStores the plugin manifest file and OpenAPI schema, which define the plugin configuration and API specification.

This README provides detailed information on how to set up, develop, and deploy the ChatGPT Retrieval Plugin (stand-alone retrieval backend).

Table of Contents

Quickstart

Follow these steps to quickly set up and run the ChatGPT Retrieval Plugin:

  1. Install Python 3.10, if not already installed.

  2. Clone the repository: git clone https://github.com/openai/chatgpt-retrieval-plugin.git

  3. Navigate to the cloned repository directory: cd /path/to/chatgpt-retrieval-plugin

  4. Install poetry: pip install poetry

  5. Create a new virtual environment with Python 3.10: poetry env use python3.10

  6. Activate the virtual environment: poetry shell

  7. Install app dependencies: poetry install

  8. Create a bearer token

  9. Set the required environment variables:

    export DATASTORE=<your_datastore>
export BEARER_TOKEN=<your_bearer_token>
export OPENAI_API_KEY=<your_openai_api_key>
export EMBEDDING_DIMENSION=256 # edit this value based on the dimension of the embeddings you want to use
export EMBEDDING_MODEL=text-embedding-3-large # edit this based on your model preference, e.g. text-embedding-3-small, text-embedding-ada-002

# Optional environment variables used when running Azure OpenAI
export OPENAI_API_BASE=https://<AzureOpenAIName>.openai.azure.com/
export OPENAI_API_TYPE=azure
export OPENAI_EMBEDDINGMODEL_DEPLOYMENTID=<Name of embedding model deployment>
export OPENAI_METADATA_EXTRACTIONMODEL_DEPLOYMENTID=<Name of deployment of model for metatdata>
export OPENAI_COMPLETIONMODEL_DEPLOYMENTID=<Name of general model deployment used for completion>
export OPENAI_EMBEDDING_BATCH_SIZE=<Batch size of embedding, for AzureOAI, this value need to be set as 1>

# Add the environment variables for your chosen vector DB.
# Some of these are optional; read the provider's setup docs in /docs/providers for more information.

# Pinecone
export PINECONE_API_KEY=<your_pinecone_api_key>
export PINECONE_ENVIRONMENT=<your_pinecone_environment>
export PINECONE_INDEX=<your_pinecone_index>

# Weaviate
export WEAVIATE_URL=<your_weaviate_instance_url>
export WEAVIATE_API_KEY=<your_api_key_for_WCS>
export WEAVIATE_CLASS=<your_optional_weaviate_class>

# Zilliz
export ZILLIZ_COLLECTION=<your_zilliz_collection>
export ZILLIZ_URI=<your_zilliz_uri>
export ZILLIZ_USER=<your_zilliz_username>
export ZILLIZ_PASSWORD=<your_zilliz_password>

# Milvus
export MILVUS_COLLECTION=<your_milvus_collection>
export MILVUS_HOST=<your_milvus_host>
export MILVUS_PORT=<your_milvus_port>
export MILVUS_USER=<your_milvus_username>
export MILVUS_PASSWORD=<your_milvus_password>

# Qdrant
export QDRANT_URL=<your_qdrant_url>
export QDRANT_PORT=<your_qdrant_port>
export QDRANT_GRPC_PORT=<your_qdrant_grpc_port>
export QDRANT_API_KEY=<your_qdrant_api_key>
export QDRANT_COLLECTION=<your_qdrant_collection>

# AnalyticDB
export PG_HOST=<your_analyticdb_host>
export PG_PORT=<your_analyticdb_port>
export PG_USER=<your_analyticdb_username>
export PG_PASSWORD=<your_analyticdb_password>
export PG_DATABASE=<your_analyticdb_database>
export PG_COLLECTION=<your_analyticdb_collection>


# Redis
export REDIS_HOST=<your_redis_host>
export REDIS_PORT=<your_redis_port>
export REDIS_PASSWORD=<your_redis_password>
export REDIS_INDEX_NAME=<your_redis_index_name>
export REDIS_DOC_PREFIX=<your_redis_doc_prefix>
export REDIS_DISTANCE_METRIC=<your_redis_distance_metric>
export REDIS_INDEX_TYPE=<your_redis_index_type>

# Llama
export LLAMA_INDEX_TYPE=<gpt_vector_index_type>
export LLAMA_INDEX_JSON_PATH=<path_to_saved_index_json_file>
export LLAMA_QUERY_KWARGS_JSON_PATH=<path_to_saved_query_kwargs_json_file>
export LLAMA_RESPONSE_MODE=<response_mode_for_query>

# Chroma
export CHROMA_COLLECTION=<your_chroma_collection>
export CHROMA_IN_MEMORY=<true_or_false>
export CHROMA_PERSISTENCE_DIR=<your_chroma_persistence_directory>
export CHROMA_HOST=<your_chroma_host>
export CHROMA_PORT=<your_chroma_port>

# Azure Cognitive Search
export AZURESEARCH_SERVICE=<your_search_service_name>
export AZURESEARCH_INDEX=<your_search_index_name>
export AZURESEARCH_API_KEY=<your_api_key> (optional, uses key-free managed identity if not set)

# Azure CosmosDB Mongo vCore
export AZCOSMOS_API = <your azure cosmos db api, for now it only supports mongo>
export AZCOSMOS_CONNSTR = <your azure cosmos db mongo vcore connection string>
export AZCOSMOS_DATABASE_NAME = <your mongo database name>
export AZCOSMOS_CONTAINER_NAME = <your mongo container name>

# Supabase
export SUPABASE_URL=<supabase_project_url>
export SUPABASE_ANON_KEY=<supabase_project_api_anon_key>

# Postgres
export PG_HOST=<postgres_host>
export PG_PORT=<postgres_port>
export PG_USER=<postgres_user>
export PG_PASSWORD=<postgres_password>
export PG_DB=<postgres_database>

# Elasticsearch
export ELASTICSEARCH_URL=<elasticsearch_host_and_port> (either specify host or cloud_id)
export ELASTICSEARCH_CLOUD_ID=<elasticsearch_cloud_id>

export ELASTICSEARCH_USERNAME=<elasticsearch_username>
export ELASTICSEARCH_PASSWORD=<elasticsearch_password>
export ELASTICSEARCH_API_KEY=<elasticsearch_api_key>

export ELASTICSEARCH_INDEX=<elasticsearch_index_name>
export ELASTICSEARCH_REPLICAS=<elasticsearch_replicas>
export ELASTICSEARCH_SHARDS=<elasticsearch_shards>

# MongoDB Atlas
export MONGODB_URI=<mongodb_uri>
export MONGODB_DATABASE=<mongodb_database>
export MONGODB_COLLECTION=<mongodb_collection>
export MONGODB_INDEX=<mongodb_index>

  10. Run the API locally: poetry run start

  11. Access the API documentation at http://0.0.0.0:8000/docs and test the API endpoints (make sure to add your bearer token).

About

Retrieval Plugin

This is a standalone retrieval backend that can be used with ChatGPT custom GPTs, function calling with the chat completions or assistants APIs, or with the ChatGPT plugins model (deprecated).

It enables a model to carry out semantic search and retrieval of personal or organizational documents, and write answers informed by relevent retrieved context (sometimes referred to as "Retrieval-Augmented Generation" or "RAG"). It allows users to obtain the most relevant document snippets from their data sources, such as files, notes, or emails, by asking questions or expressing needs in natural language. Enterprises can make their internal documents available to their employees through ChatGPT using this plugin.

The plugin uses OpenAI's embeddings model (text-embedding-3-large 256 dimension embeddings by default) to generate embeddings of document chunks, and then stores and queries them using a vector database on the backend. As an open-source and self-hosted solution, developers can deploy their own Retrieval Plugin and register it with ChatGPT. The Retrieval Plugin supports several vector database providers, allowing developers to choose their preferred one from a list.

A FastAPI server exposes the plugin's endpoints for upserting, querying, and deleting documents. Users can refine their search results by using metadata filters by source, date, author, or other criteria. The plugin can be hosted on any cloud platform that supports Docker containers, such as Fly.io, Heroku, Render, or Azure Container Apps. To keep the vector database updated with the latest documents, the plugin can process and store documents from various data sources continuously, using incoming webhooks to the upsert and delete endpoints. Tools like Zapier or Make can help configure the webhooks based on events or schedules.

Retrieval Plugin with Custom GPTs

To create a custom GPT that can use your Retrieval Plugin for semantic search and retrieval of your documents, and even store new information back to the database, you first need to have deployed a Retrieval Plugin. For detailed instructions on how to do this, please refer to the Deployment section. Once you have your app URL (e.g., https://your-app-url.com), take the following steps:

  1. Navigate to the create GPT page at https://chat.openai.com/gpts/editor.
  2. Follow the standard creation flow to set up your GPT.
  3. Navigate to the "Configure" tab. Here, you can manually fill in fields such as name, description, and instructions, or use the smart creator for assistance.
  4. Under the "Actions" section, click on "Create new action".
  5. Choose an authentication method. The Retrieval Plugin supports None, API key (Basic or Bearer) and OAuth. For more information on these methods, refer to the Authentication Methods Section.
  6. Import the OpenAPI schema. You can either:
    • Import directly from the OpenAPI schema hosted in your app at https://your-app-url.com/.well-known/openapi.yaml.
    • Copy and paste the contents of this file into the Schema input area if you only want to expose the query endpoint to the GPT. Remember to change the URL under the -servers section of the OpenAPI schema you paste in.
  7. Optionally, you might want to add a fetch endpoint. This would involve editing the /server/main.py file to add an endpoint and implement this for your chosen vector database. If you make this change, please consider contributing it back to the project by opening a pull request! Adding the fetch endpoint to the OpenAPI schema would allow the model to fetch more content from a document by ID if some text is cut off in the retrieved result. It might also be useful to pass in a string with the text from the retrieved result and an option to return a fixed length of context before and after the retrieved result.
  8. If you want the GPT to be able to save information back to the vector database, you can give it access to the Retrieval Plugin's /upsert endpoint. To do this, copy the contents of this file into the schema area. This allows the GPT to store new information it generates or learns during the conversation. More details on this feature can be found at Memory Feature and in the docs here.

Remember: ChatGPT and custom GPTs natively support retrieval from uploaded files, so you should use the Retrieval Plugin as a backend only if you want more granular control of your retrieval system (e.g. self-hosting, embedding chunk length, embedding model / size, etc.).

Retrieval Plugin with Function Calling

The Retrieval Plugin can be integrated with function calling in both the Chat Completions API and the Assistants API. This allows the model to decide when to use your functions (query, fetch, upsert) based on the conversation context.

Function Calling with Chat Completions

In a call to the chat completions API, you can describe functions and have the model generate a JSON object containing arguments to call one or many functions. The latest models (gpt-3.5-turbo-0125 and gpt-4-turbo-preview) have been trained to detect when a function should be called and to respond with JSON that adheres to the function signature.

You can define the functions for the Retrieval Plugin endpoints and pass them in as tools when you use the Chat Completions API with one of the latest models. The model will then intelligently call the functions. You can use function calling to write queries to your APIs, call the endpoint on the backend, and return the response as a tool message to the model to continue the conversation. The function definitions/schemas and an example can be found here.

Function Calling with Assistants API

You can use the same function definitions with the OpenAI Assistants API, specifically the function calling in tool use. The Assistants API allows you to build AI assistants within your own applications, leveraging models, tools, and knowledge to respond to user queries. The function definitions/schemas and an example can be found here. The Assistants API natively supports retrieval from uploaded files, so you should use the Retrieval Plugin with function calling only if you want more granular control of your retrieval system (e.g. embedding chunk length, embedding model / size, etc.).

Parallel function calling is supported for both the Chat Completions API and the Assistants API. This means you can perform multiple tasks, such as querying something and saving something back to the vector database, in the same message.

Read more about function calling with the Retrieval Plugin here.

ChatGPT Plugins Model

(deprecated) We recommend using custom actions with GPTs to make use of the Retrieval Plugin through ChatGPT. Instrucitons for using retrieval with the deprecated plugins model can be found here.

API Endpoints

The Retrieval Plugin is built using FastAPI, a web framework for building APIs with Python. FastAPI allows for easy development, validation, and documentation of API endpoints. Find the FastAPI documentation here.

One of the benefits of using FastAPI is the automatic generation of interactive API documentation with Swagger UI. When the API is running locally, Swagger UI at <local_host_url i.e. http://0.0.0.0:8000>/docs can be used to interact with the API endpoints, test their functionality, and view the expected request and response models.

The plugin exposes the following endpoints for upserting, querying, and deleting documents from the vector database. All requests and responses are in JSON format, and require a valid bearer token as an authorization header.

The detailed specifications and examples of the request and response models can be found by running the app locally and navigating to http://0.0.0.0:8000/openapi.json, or in the OpenAPI schema here. Note that the OpenAPI schema only contains the /query endpoint, because that is the only function that ChatGPT needs to access. This way, ChatGPT can use the plugin only to retrieve relevant documents based on natural language queries or needs. However, if developers want to also give ChatGPT the ability to remember things for later, they can use the /upsert endpoint to save snippets from the conversation to the vector database. An example of a manifest and OpenAPI schema that gives ChatGPT access to the /upsert endpoint can be found here.

To include custom metadata fields, edit the DocumentMetadata and DocumentMetadataFilter data models here, and update the OpenAPI schema here. You can update this easily by running the app locally, copying the JSON found at http://0.0.0.0:8000/sub/openapi.json, and converting it to YAML format with Swagger Editor. Alternatively, you can replace the openapi.yaml file with an openapi.json file.

Memory Feature

A notable feature of the Retrieval Plugin is its capacity to provide ChatGPT with memory. By using the plugin's upsert endpoint, ChatGPT can save snippets from the conversation to the vector database for later reference (only when prompted to do so by the user). This functionality contributes to a more context-aware chat experience by allowing ChatGPT to remember and retrieve information from previous conversations. Learn how to configure the Retrieval Plugin with memory here.

Security

The Retrieval Plugin allows ChatGPT to search a vector database of content, and then add the best results into the ChatGPT session. This means it doesn’t have any external effects, and the main risk consideration is data authorization and privacy. Developers should only add content into their Retrieval Plugin that they have authorization for and that they are fine with appearing in users’ ChatGPT sessions. You can choose from a number of different authentication methods to secure the plugin (more information here).

Choosing an Embeddings Model

The ChatGPT Retrieval Plugin uses OpenAI's embeddings models to generate embeddings of document chunks. The default model for the Retrieval Plugin is text-embedding-3-large with 256 dimensions. OpenAI offers two latest embeddings models, text-embedding-3-small and text-embedding-3-large, as well as an older model, text-embedding-ada-002.

The new models support shortening embeddings without significant loss of retrieval accuracy, allowing you to balance retrieval accuracy, cost, and speed.

Here's a comparison of the models:

ModelEmbedding SizeAverage MTEB ScoreCost per 1k Tokens
text-embedding-3-large307264.6%$0.00013
text-embedding-3-large102464.1%$0.00013
text-embedding-3-large25662.0%$0.00013
text-embedding-3-small153662.3%$0.00002
text-embedding-3-small51261.6%$0.00002
text-embedding-ada-002153661.0%$0.0001

When choosing a model, consider:

  1. Retrieval Accuracy vs Cost: text-embedding-3-large offers the highest accuracy but at a higher cost. text-embedding-3-small is more cost-effective with competitive accuracy. The older text-embedding-ada-002 model has the lowest accuracy.

  2. Embedding Size: Larger embeddings provide better accuracy but consume more storage and could be slower to query. You can adjust the size of the embeddings to balance these factors.

For example, if your vector database supports up to 1024 dimensions, you can use text-embedding-3-large and set the dimensions API parameter to 1024. This shortens the embedding from 3072 dimensions, trading off some accuracy for lower storage and query costs.

To change your chosen embeddings model and size, edit the following environment variables:

EMBEDDING_DIMENSION=256 # edit this value based on the dimension of the embeddings you want to use
EMBEDDING_MODEL="text-embedding-3-large" # edit this value based on the model you want to use e.g. text-embedding-3-small, text-embedding-ada-002

Development

Setup

This app uses Python 3.10, and poetry for dependency management.

Install Python 3.10 on your machine if it isn't already installed. It can be downloaded from the official Python website or with a package manager like brew or apt, depending on your system.

Clone the repository from GitHub:

git clone https://github.com/openai/chatgpt-retrieval-plugin.git

Navigate to the cloned repository directory:

cd /path/to/chatgpt-retrieval-plugin

Install poetry:

pip install poetry

Create a new virtual environment that uses Python 3.10:

poetry env use python3.10
poetry shell

Install app dependencies using poetry:

poetry install

Note: If adding dependencies in the pyproject.toml, make sure to run poetry lock and poetry install.

General Environment Variables

The API requires the following environment variables to work:

NameRequiredDescription
DATASTOREYesThis specifies the vector database provider you want to use to store and query embeddings. You can choose from elasticsearch, chroma, pinecone, weaviate, zilliz, milvus, qdrant, redis, azuresearch, supabase, postgres, analyticdb, mongodb-atlas.
BEARER_TOKENYesThis is a secret token that you need to authenticate your requests to the API. You can generate one using any tool or method you prefer, such as jwt.io.
OPENAI_API_KEYYesThis is your OpenAI API key that you need to generate embeddings using the one of the OpenAI embeddings model. You can get an API key by creating an account on OpenAI.

Using the plugin with Azure OpenAI

The Azure Open AI uses URLs that are specific to your resource and references models not by model name but by the deployment id. As a result, you need to set additional environment variables for this case.

In addition to the OPENAI_API_BASE (your specific URL) and OPENAI_API_TYPE (azure), you should also set OPENAI_EMBEDDINGMODEL_DEPLOYMENTID which specifies the model to use for getting embeddings on upsert and query. For this, we recommend deploying text-embedding-ada-002 model and using the deployment name here.

If you wish to use the data preparation scripts, you will also need to set OPENAI_METADATA_EXTRACTIONMODEL_DEPLOYMENTID, used for metadata extraction and OPENAI_COMPLETIONMODEL_DEPLOYMENTID, used for PII handling.

Choosing a Vector Database

The plugin supports several vector database providers, each with different features, performance, and pricing. Depending on which one you choose, you will need to use a different Dockerfile and set different environment variables. The following sections provide brief introductions to each vector database provider.

For more detailed instructions on setting up and using each vector database provider, please refer to the respective documentation in the /docs/providers/<datastore_name>/setup.md file (folders here).

Pinecone

Pinecone is a managed vector database designed for speed, scale, and rapid deployment to production. It supports hybrid search and is currently the only datastore to natively support SPLADE sparse vectors. For detailed setup instructions, refer to /docs/providers/pinecone/setup.md.

Weaviate

Weaviate is an open-source vector search engine built to scale seamlessly into billions of data objects. It supports hybrid search out-of-the-box, making it suitable for users who require efficient keyword searches. Weaviate can be self-hosted or managed, offering flexibility in deployment. For detailed setup instructions, refer to /docs/providers/weaviate/setup.md.

Zilliz

Zilliz is a managed cloud-native vector database designed for billion-scale data. It offers a wide range of features, including multiple indexing algorithms, distance metrics, scalar filtering, time travel searches, rollback with snapshots, full RBAC, 99.9% uptime, separated storage and compute, and multi-language SDKs. For detailed setup instructions, refer to /docs/providers/zilliz/setup.md.

Milvus

Milvus is an open-source, cloud-native vector database that scales to billions of vectors. It is the open-source version of Zilliz and shares many of its features, such as various indexing algorithms, distance metrics, scalar filtering, time travel searches, rollback with snapshots, multi-language SDKs, storage and compute separation, and cloud scalability. For detailed setup instructions, refer to /docs/providers/milvus/setup.md.

Qdrant

Qdrant is a vector database capable of storing documents and vector embeddings. It offers both self-hosted and managed Qdrant Cloud deployment options, providing flexibility for users with different requirements. For detailed setup instructions, refer to /docs/providers/qdrant/setup.md.

Redis

Redis is a real-time data platform suitable for a variety of use cases, including everyday applications and AI/ML workloads. It can be used as a low-latency vector engine by creating a Redis database with the Redis Stack docker container. For a hosted/managed solution, Redis Cloud is available. For detailed setup instructions, refer to /docs/providers/redis/setup.md.

LlamaIndex

LlamaIndex is a central interface to connect your LLM's with external data. It provides a suite of in-memory indices over your unstructured and structured data for use with ChatGPT. Unlike standard vector databases, LlamaIndex supports a wide range of indexing strategies (e.g. tree, keyword table, knowledge graph) optimized for different use-cases. It is light-weight, easy-to-use, and requires no additional deployment. All you need to do is specifying a few environment variables (optionally point to an existing saved Index json file). Note that metadata filters in queries are not yet supported. For detailed setup instructions, refer to /docs/providers/llama/setup.md.

Chroma

Chroma is an AI-native open-source embedding database designed to make getting started as easy as possible. Chroma runs in-memory, or in a client-server setup. It supports metadata and keyword filtering out of the box. For detailed instructions, refer to /docs/providers/chroma/setup.md.

Azure Cognitive Search

Azure Cognitive Search is a complete retrieval cloud service that supports vector search, text search, and hybrid (vectors + text combined to yield the best of the two approaches). It also offers an optional L2 re-ranking step to further improve results quality. For detailed setup instructions, refer to /docs/providers/azuresearch/setup.md

Azure CosmosDB Mongo vCore

Azure CosmosDB Mongo vCore supports vector search on embeddings, and it could be used to seamlessly integrate your AI-based applications with your data stored in the Azure CosmosDB. For detailed instructions, refer to /docs/providers/azurecosmosdb/setup.md

Supabase

Supabase offers an easy and efficient way to store vectors via pgvector extension for Postgres Database. You can use Supabase CLI to set up a whole Supabase stack locally or in the cloud or you can also use docker-compose, k8s and other options available. For a hosted/managed solution, try Supabase.com and unlock the full power of Postgres with built-in authentication, storage, auto APIs, and Realtime features. For detailed setup instructions, refer to /docs/providers/supabase/setup.md.

Postgres

Postgres offers an easy and efficient way to store vectors via pgvector extension. To use pgvector, you will need to set up a PostgreSQL database with the pgvector extension enabled. For example, you can use docker to run locally. For a hosted/managed solution, you can use any of the cloud vendors which support pgvector. For detailed setup instructions, refer to /docs/providers/postgres/setup.md.

AnalyticDB

AnalyticDB is a distributed cloud-native vector database designed for storing documents and vector embeddings. It is fully compatible with PostgreSQL syntax and managed by Alibaba Cloud. AnalyticDB offers a powerful vector compute engine, processing billions of data vectors and providing features such as indexing algorithms, structured and unstructured data capabilities, real-time updates, distance metrics, scalar filtering, and time travel searches. For detailed setup instructions, refer to /docs/providers/analyticdb/setup.md.

Elasticsearch

Elasticsearch currently supports storing vectors through the dense_vector field type and uses them to calculate document scores. Elasticsearch 8.0 builds on this functionality to support fast, approximate nearest neighbor search (ANN). This represents a much more scalable approach, allowing vector search to run efficiently on large datasets. For detailed setup instructions, refer to /docs/providers/elasticsearch/setup.md.

Mongodb-Atlas

MongoDB Atlas Currently, the procedure involves generating an Atlas Vector Search index for all collections featuring vector embeddings of 2048 dimensions or fewer in width. This applies to diverse data types coexisting with additional data on your Atlas cluster, and the process is executed through the Atlas UI and Atlas Administration AP, refer to /docs/providers/mongodb_atlas/setup.md.

Running the API locally

To run the API locally, you first need to set the requisite environment variables with the export command:

export DATASTORE=<your_datastore>
export BEARER_TOKEN=<your_bearer_token>
export OPENAI_API_KEY=<your_openai_api_key>
<Add the environment variables for your chosen vector DB here>

Start the API with:

poetry run start

Append docs to the URL shown in the terminal and open it in a browser to access the API documentation and try out the endpoints (i.e. http://0.0.0.0:8000/docs). Make sure to enter your bearer token and test the API endpoints.

Note: If you add new dependencies to the pyproject.toml file, you need to run poetry lock and poetry install to update the lock file and install the new dependencies.

Personalization

You can personalize the Retrieval Plugin for your own use case by doing the following:

Authentication Methods

You can choose from four options for authenticating requests to your plugin:

  1. No Authentication: Anyone can add your plugin and use its API without any credentials. This option is suitable if you are only exposing documents that are not sensitive or already public. It provides no security for your data. If using this method, copy the contents of this main.py into the actual main.py file. Example manifest here.

  2. HTTP Bearer: You can use a secret token as a header to authorize requests to your plugin. There are two variants of this option:

    • User Level (default for this implementation): Each user who adds your plugin to ChatGPT must provide the bearer token when adding the plugin. You can generate and distribute these tokens using any tool or method you prefer, such as jwt.io. This method provides better security as each user has to enter the shared access token. If you require a unique access token for each user, you will need to implement this yourself in the main.py file. Example manifest here.

    • Service Level: Anyone can add your plugin and use its API without credentials, but you must add a bearer token when registering the plugin. When you install your plugin, you need to add your bearer token, and will then receive a token from ChatGPT that you must include in your hosted manifest file. Your token will be used by ChatGPT to authorize requests to your plugin on behalf of all users who add it. This method is more convenient for users, but it may be less secure as all users share the same token and do not need to add a token to install the plugin. Example manifest here.

  3. OAuth: Users must go through an OAuth flow to add your plugin. You can use an OAuth provider to authenticate users who add your plugin and grant them access to your API. This method offers the highest level of security and control, as users authenticate through a trusted third-party provider. However, you will need to implement the OAuth flow yourself in the main.py file and provide the necessary parameters in your manifest file. Example manifest here.

Consider the benefits and drawbacks of each authentication method before choosing the one that best suits your use case and security requirements. If you choose to use a method different to the default (User Level HTTP), make sure to update the manifest file here.

Deployment

You can deploy your app to different cloud providers, depending on your preferences and requirements. However, regardless of the provider you choose, you will need to update two files in your app: openapi.yaml and ai-plugin.json. As outlined above, these files define the API specification and the AI plugin configuration for your app, respectively. You need to change the url field in both files to match the address of your deployed app.

Render has a 1-click deploy option that automatically updates the url field in both files:

<img src="https://render.com/images/deploy-to-render-button.svg" alt="Deploy to Render" />

Before deploying your app, you might want to remove unused dependencies from your pyproject.toml file to reduce the size of your app and improve its performance. Depending on the vector database provider you choose, you can remove the packages that are not needed for your specific provider. Refer to the respective documentation in the /docs/deployment/removing-unused-dependencies.md file for information on removing unused dependencies for each provider.

Instructions:

Once you have deployed your app, consider uploading an initial batch of documents using one of these scripts or by calling the /upsert endpoint.

Webhooks

To keep the documents stored in the vector database up-to-date, consider using tools like Zapier or Make to configure incoming webhooks to your plugin's API based on events or schedules. For example, this could allow you to sync new information as you update your notes or receive emails. You can also use a Zapier Transfer to batch process a collection of existing documents and upload them to the vector database.

If you need to pass custom fields from these tools to your plugin, you might want to create an additional Retrieval Plugin API endpoint that calls the datastore's upsert function, such as upsert-email. This custom endpoint can be designed to accept specific fields from the webhook and process them accordingly.

To set up an incoming webhook, follow these general steps:

After setting up the webhook, you may want to run a backfill to ensure that any previously missed data is included in the vector database.

Remember that if you want to use incoming webhooks to continuously sync data, you should consider running a backfill after setting these up to avoid missing any data.

In addition to using tools like Zapier and Make, you can also build your own custom integrations to sync data with your Retrieval Plugin. This allows you to have more control over the data flow and tailor the integration to your specific needs and requirements.

Scripts

The scripts folder contains scripts to batch upsert or process text documents from different data sources, such as a zip file, JSON file, or JSONL file. These scripts use the plugin's upsert utility functions to upload the documents and their metadata to the vector database, after converting them to plain text and splitting them into chunks. Each script folder has a README file that explains how to use it and what parameters it requires. You can also optionally screen the documents for personally identifiable information (PII) using a language model and skip them if detected, with the services.pii_detection module. This can be helpful if you want to avoid uploading sensitive or private documents to the vector database unintentionally. Additionally, you can optionally extract metadata from the document text using a language model, with the services.extract_metadata module. This can be useful if you want to enrich the document metadata. Note: if using incoming webhooks to continuously sync data, consider running a backfill after setting these up to avoid missing any data.

The scripts are:

Pull Request (PR) Checklist

If you'd like to contribute, please follow the checklist below when submitting a PR. This will help us review and merge your changes faster! Thank you for contributing!

  1. Type of PR: Indicate the type of PR by adding a label in square brackets at the beginning of the title, such as [Bugfix], [Feature], [Enhancement], [Refactor], or [Documentation].

  2. Short Description: Provide a brief, informative description of the PR that explains the changes made.

  3. Issue(s) Linked: Mention any related issue(s) by using the keyword Fixes or Closes followed by the respective issue number(s) (e.g., Fixes #123, Closes #456).

  4. Branch: Ensure that you have created a new branch for the changes, and it is based on the latest version of the main branch.

  5. Code Changes: Make sure the code changes are minimal, focused, and relevant to the issue or feature being addressed.

  6. Commit Messages: Write clear and concise commit messages that explain the purpose of each commit.

  7. Tests: Include unit tests and/or integration tests for any new code or changes to existing code. Make sure all tests pass before submitting the PR.

  8. Documentation: Update relevant documentation (e.g., README, inline comments, or external documentation) to reflect any changes made.

  9. Review Requested: Request a review from at least one other contributor or maintainer of the repository.

  10. Video Submission (For Complex/Large PRs): If your PR introduces significant changes, complexities, or a large number of lines of code, submit a brief video walkthrough along with the PR. The video should explain the purpose of the changes, the logic behind them, and how they address the issue or add the proposed feature. This will help reviewers to better understand your contribution and expedite the review process.

Pull Request Naming Convention

Use the following naming convention for your PR branches:

<type>/<short-description>-<issue-number>

Example:

feature/advanced-chunking-strategy-123

Limitations

While the ChatGPT Retrieval Plugin is designed to provide a flexible solution for semantic search and retrieval, it does have some limitations:

Future Directions

The ChatGPT Retrieval Plugin provides a flexible solution for semantic search and retrieval, but there is always potential for further development. We encourage users to contribute to the project by submitting pull requests for new features or enhancements. Notable contributions may be acknowledged with OpenAI credits.

Some ideas for future directions include:

We welcome contributions from the community to help improve the ChatGPT Retrieval Plugin and expand its capabilities. If you have an idea or feature you'd like to contribute, please submit a pull request to the repository.

Contributors

We would like to extend our gratitude to the following contributors for their code / documentation contributions, and support in integrating various vector database providers with the ChatGPT Retrieval Plugin: