A Question Answering chatbot powered by GPT-3 answer synthesis and sentence-transformers sentence embeddings that can answer questions based on your own data. Here, we use a PDF file about superheroes and their evolution and influence on pop-culture. The knowledge from the PDF is fed into the bot so that the bot can later answer questions about it when asked.
Currently works with FAISS but can easily be used with other vector DBs like qdrant, chroma, pinecone, etc for better scalability.
A notebook on how to build such a bot is given. Make sure all modules and libraries mentioned in requirements.txt is installed before starting! In this method, rather than using OpenAI api for generating embeddings for the text, we use open-source models which brings down the cost by a lot while maintaining the same level of quality. (Thanks to excellent models from sentence-transformers) Only during answer synthesis i.e. to generate the final answers we use the OpenAI api.
A Demo app can be found here
