Semantic Embeddings with ChatGPT

This project uses Zep, a semantic embeddings database, to power a ChatGPT conversational platform that is able to entertain Q&A style conversation with chat context recall and semantic embedding search for better contextual responses. See this blog post for more information on the theory and implementation.

Deployment

First, clone the repo with:

git clone https://github.com/TheMatrixMaster/gpt-embeddings

Then you will need to pull the zep submodule using the following command from the project root

cd gpt-embeddings
git submodule update --init

Adding secret keys and seed data

First, we need to create a few .env files, so that both the frontend and backend have access to gpt models for conversation and embedding, respectively.

Please create the following file zep/.env and populate it with the following keys:

ZEP_OPENAI_API_KEY="<OPENAI API KEY>"
ZEP_LLM_OPENAI_ORG_ID="OPTIONAL OPENAI ORG"

Then, create another env file in the following path web/.env and populate it with the same key values. Notice that the key name has changed though.

OPENAI_API_KEY="OPENAI API KEY"
OPENAI_ORGANIZATION="OPTIONAL OPENAI ORG"

Finally, we need to create a few documents that the database can embed into search vectors for the context retrieval. Please create a new directory in the repo root named documents/ and create a file in there named documents/raw_convo.txt. (TODO: add support for auto-chunking any text, json, or yml documents)

Deploy from docker

Once the above steps have been completed, you should be able to start the service on your machine by running the following command from the project root. Make sure you have docker and docker-compose installed.

docker compose up