GPT-3 is a relativelly new NLP algorithm from OpenAI which produces very impressive results. It is very good in generalization and adapts very well to new inputs. It can be "trained" by feeding it custom knowledge material in text form. When providing Cisco's documentation as a source of knowledge, it is possible to create a chatbot that is able to answer detailed questions about Cisco's products and services. And this is exactly what I am trying to do right now.
This is a WiP. So far, it is a simple offline python chatbot trained only on the DNA Center data sheet. The end goal of course is to create a public available Webex bot based on a much bigger knowledge base.
But even with such a rather small "training set" the results are very promising.
If you would like to contribute, please contact me via e-mail (mdemikho@cisco.com) or Webex.
In order to run the bot you need Python 3 and the OpenAI python package and jsonlines.
Also, you need to have access to the GPT-3 API. More infos here: https://openai.com/blog/openai-api/
Just copy the files to your host and follow the setup steps below.
Open up the terminal and type in:
pip3 install -r requirements.txtJust insert your GPT-3 API key in config.json
Currently the bot is trained with the DNA Center data sheet. If you would like to add more training material, add it
to the sources folder. The knowledge base needs to be in text form and each chapter needs to be separated by ###.
Thereafter, the file needs to be converted into a JSONLINES format and uploaded to the OpenAI cloud. This can be done
by file2jsonl and uploadFilefunctions of the utils library. Or just use refreshtest.py. It deletes the current
knowledge base and replaces it with the one you define. In order to use it, just replace the file paths in the code with
your own and you are good to go (but do not change the paths of config.json and file_id).
In order to chat with the bot, just start main.py
python main.py