qaGPT

Inspiration

ChatGPT is one of the inspiration behind this project. The idea was to explore if we can combine power of retrieval models with ChatGPT generate one. Chat GPT generation is like magic however it lakes specificity and can get even better if it conditioned by a good retrieval based model. This is our first crude attempt to integrate these approaches.

What it does

We created a chat-bot backend and Avatar (Avatar graphic asset is not new). The focus of this project was the chatbot backend which is completely done during the hackathon time.

How we built it

We have used different components such as haystack (on top of hugging-face) , chatgpt API (through PyChatGPT) to build our MVP. Our model consists of followings: 1- A retrieval model to retrieve the relevant documents and passages. We used a Dense Reterival model from huggingface (through haystack). 2- A reading/comprehension model to focus more on important parts of the document. We used farm reader from haystack. 3- A non-chatgpt generative model. We create a model based on T5 using haystack and huggingface. 4- some logical operation to integrate chat gpt and the output of these models.

Challenges we ran into

To name a few: chat GPT API is not official and we get all sort of last minute problems. integrating our model and chat GPT interface was challenging and still need some works.

Accomplishments that we're proud of

We made this system with many different components work in roughly 2 days. It was specially challenging for a remote team like us.

What we learned

Although we have some initial success to make this integration works but we need to have more control over the chat gpt (or any generator) module to preform as we want to,