Building ChatGPT plugins
Use Supabase as a Retrieval Store for your ChatGPT plugin.
ChatGPT recently released Plugins which help ChatGPT access up-to-date information, run computations, or use third-party services. If you're building a plugin for ChatGPT, you'll probably want to answer questions from a specific source. We can solve this with “retrieval plugins”, which allow ChatGPT to access information from a database.
What is ChatGPT Retrieval Plugin?#
A Retrieval Plugin is a Python project designed to inject external data into a ChatGPT conversation. It does a few things:
- Turn documents into smaller chunks.
- Converts chunks into embeddings using OpenAI's
- Stores the embeddings into a vector database.
- Queries the vector database for relevant documents when a question is asked.
It allows ChatGPT to dynamically pull relevant information into conversations from your data sources. This could be PDF documents, Confluence, or Notion knowledge bases.
Example: Chat with Postgres Docs#
Let’s build an example where we can “ask ChatGPT questions” about the Postgres documentation. Although ChatGPT already knows about the Postgres documentation because it is publicly available, this is a simple example which demonstrates how to work with PDF files.
This plugin requires several steps:
- Download all the Postgres docs as a PDF
- Convert the docs into chunks of embedded text and store them in Supabase
- Run our plugin locally so that we can ask questions about the Postgres docs.
We'll be saving the Postgres documentation in Postgres, and ChatGPT will be retrieving the documentation whenever a user asks a question:
Step 1: Fork the ChatGPT Retrieval Plugin repository#
Fork the ChatGPT Retrieval Plugin repository to your GitHub account and clone it to your local machine. Read through the
README.md file to understand the project structure.
Step 2: Install dependencies#
Choose your desired datastore provider and remove unused dependencies from
pyproject.toml. For this example, we'll use Supabase. And install dependencies with Poetry:
Step 3: Create a Supabase project#
_10export OPENAI_API_KEY=<open_ai_api_key>_10export DATASTORE=supabase_10export SUPABASE_URL=<supabase_url>_10export SUPABASE_SERVICE_ROLE_KEY=<supabase_key>
For Postgres datastore, you'll need to export these environment variables instead:
_10export OPENAI_API_KEY=<open_ai_api_key>_10export DATASTORE=postgres_10export PG_HOST=<postgres_host_url>_10export PG_PASSWORD=<postgres_password>
Step 4: Run Postgres Locally#
To start quicker you may use Supabase CLI to spin everything up locally as it already includes pgvector from the start. Install
supabase-cli, go to the
examples/providers folder in the repo and run:
This will pull all docker images and run supabase stack in docker on your local machine. It will also apply all the necessary migrations to set the whole thing up. You can then use your local setup the same way, just export the environment variables and follow to the next steps.
supabase-cli is not required and you can use any other docker image or hosted version of PostgresDB that includes
pgvector. Just make sure you run migrations from
Step 5: Obtain OpenAI API key#
To create embeddings Plugin uses OpenAI API and
text-embedding-ada-002 model. Each time we add some data to our datastore, or try to query relevant information from it, embedding will be created either for inserted data chunk, or for the query itself. To make it work we need to export
OPENAI_API_KEY. If you already have an account in OpenAI, you just need to go to User Settings - API keys and Create new secret key.
Step 6: Run the plugin#
Execute the following command to run the plugin:
_10poetry run dev_10# output_10INFO: Will watch for changes in these directories: ['./chatgpt-retrieval-plugin']_10INFO: Uvicorn running on http://localhost:3333 (Press CTRL+C to quit)_10INFO: Started reloader process  using WatchFiles_10INFO: Started server process _10INFO: Waiting for application startup._10INFO: Application startup complete.
The plugin will start on your localhost - port
:3333 by default.
Step 6: Populating data in the datastore#
For this example, we'll upload Postgres documentation to the datastore. Download the Postgres documentation and use the
/upsert-file endpoint to upload it:
_10curl -X POST -F \\"file=@./postgresql-15-US.pdf\\" <http://localhost:3333/upsert-file>
The plugin will split your data and documents into smaller chunks automatically. You can view the chunks using the Supabase dashboard or any other SQL client you prefer. For the whole Postgres Documentation I got 7,904 records in my documents table, which is not a lot, but we can try to add index for
embedding column to speed things up by a little. To do so, you should run the following SQL command:
_10create index on documents_10using ivfflat (embedding vector_ip_ops)_10with (lists = 10);
This will create an index for the inner product distance function. Important to note that it is an approximate index. It will change the logic from performing the exact nearest neighbor search to the approximate nearest neighbor search.
We are using
lists = 10, because as a general guideline, you should start looking for optimal lists constant value with the formula:
rows / 1000 when you have less than 1 million records in your table.
Step 7: Using our plugin within ChatGPT#
To integrate our plugin with ChatGPT, register it in the ChatGPT dashboard. Assuming you have access to ChatGPT Plugins and plugin development, select the Plugins model in a new chat, then choose "Plugin store" and "Develop your own plugin." Enter
localhost:3333 into the domain input, and your plugin is now part of ChatGPT.
You can now ask questions about Postgres and receive answers derived from the documentation.
Let's try it out: ask ChatGPT to find out when to use
check and when to use
using. You will be able to see what queries were sent to our plugin and what it responded to.
And after ChatGPT receives a response from the plugin it will answer your question with the data from the documentation.