About

This project is based on FastAPI and represents a chatbot server comprising of multiple API's for processing user input prompt in various text and audio formats. The responses is generated using openAI and multiple apporaches are accomodated to send the response back to the user. Here are a few details that will come in handy while navigating this repository.

.
├── assets                  # Domain specific files provided to OpenAI Assistant
├── interface               # Utility classes used within the server code
├── legacy                  # Generalized version of the chat bot server
├── services                # Services used within the server files
    ├── aws
    ├── openai
    ├── redis
    ├── utility
        ├── constants
├── server.py               # Chatbot main server file
├── server-test.py          # Chatbot test server file
├── requirements.txt        # List of dependency used for deploying 
├── Procfile                # Procfile (equired by Heroku only)
└── README.md

Getting Started

Text Based Conversational Bot Server

There are two forms of servers available under this repository

• The first one is a basic one found under the legacy folder, this can be re-used for multiple scenarios depending on your application domain and is built upon OpenAI Assistants and depicts the workflow. Follow the steps below to make the legacy server work

  • Create a .env file comprising of the folling details

    ASSISTANT_ID=XXXXXXXXXXXX
    OPENAI_API_KEY=XXXXXXXXXX
    THREAD_ID=XXXXXXXXXXXXXXX
    
    AWS_ACCESS_KEY_ID=XXXXXXXXXXXXXXXX
    AWS_DEFAULT_REGION=XXXXXXXXXXXXXXX
    AWS_SECRET_ACCESS_KEY=XXXXXXXXXXXX
    

  • use pip install to install all dependencies.

  • You can test by using using a server-test.py within the root directory by running the following command on your terminal given that the aforementioned keys are provided in the .env file

    uvicorn server-test:app --host=127.0.0.1 --port=8000
    

  • Once you're server is running you can visit the API doc file here http://localhost:8000/docs where you can test it

  • To understand how an assistant works, we recommend that you go through the official doc from OpenAI https://platform.openai.com/docs/assistants/how-it-works .

  • Here's the interaction diagram of the workflow

    

• The second one is found the main directory by the name of server.py and is an example of a specialized bot for the any domian's policy documents. This bot is built on Conrad Lab's policy documents which are maintained by the human rescources. Follow the steps below to make the server work

  • Create a .env file comprising of the folling details

    ASSISTANT_ID=XXXXXXXXXXXX
    OPENAI_API_KEY=XXXXXXXXXX
    THREAD_ID=XXXXXXXXXXXXXXX
    
    AWS_ACCESS_KEY_ID=XXXXXXXXXXXXXXXX
    AWS_DEFAULT_REGION=XXXXXXXXXXXXXXX
    AWS_SECRET_ACCESS_KEY=XXXXXXXXXXXX
    
    REDIS_ENDPOINT=XXXXXXXXXXXXXXXX
    REDIS_PORT=XXXXXXXXXXXXXXXXXXXX
    REDIS_PASSWORD=XXXXXXXXXXXXXXXX
    

    • use pip install to install all dependencies.

    • use the following command to run the server

      uvicorn server:app --host=127.0.0.1 --port=8000 --reload (--reload is optional for hot reload)
      

    • The introduction of redis instace is the only difference between the aforementioned legacy server, in this use-case we need it to traclk and persist the thread Ids of the OpenAI assistants, which saves the context of the conversation. These thread Ids will be saved against a unique identifier which refers to an individual user, whenever we need to update the assistant we have to flush these thread Ids to impose a new context and hence new threads will be created in order to achieve that and will be saved. We also keep track of the file Ids for each uploaded file to the assistant, these come in handy whenver we are updating the OpenAI assistant. In order to achieve that the new file needs to be uploaded first on the OpenAI platform, the file Id against the new file would be replaced with the previous and store in the redis. The files not provided to the OpenAI will be deleted from the context, which means we have to keep track of all files that exist in a particular context while replacing old files with the new ones in the process and refreshing the threads along the way.

    • Here's the interaction diagram of the workflow

      

Voice Based Response Server Endpoints

The voice based endpoints are common between both legacy and main directory server files but are subject to changes in the future. The two tools AWS Transcribe and AWS Polly are used in these endpoints to transcribe the audio file send by the user, the generated text is used to generate a response throgh OpenAI with a stream=True parameter within the completion API. The stream parameters helps in generating stream based response which are received in forms of chunks and can be immediately processed by Polly to be spoken out to the user. Make sure you have the following keys present within your .env file

AWS_ACCESS_KEY_ID=XXXXXXXXXXXXXXXX
AWS_DEFAULT_REGION=XXXXXXXXXXXXXXX
AWS_SECRET_ACCESS_KEY=XXXXXXXXXXXX

• For Official Use /processAudioFile
  This endpoint follows the complete workflow as stated within the aforementioned paragraph and is depicted in the diagram below. The voice chunks are uploaded as a file and is transcribed by AWS Transcribe and is served to AWS Polly which is streamed to the users.

  

• For Official & Testing both /processText
  This endpoint can be used for testing purposes, you can provide a test prompt which will not require transcribing and just be sent to OpenAI completion API to generate the response and be spoken out by Polly. This is also being used in one of the iterations of the server where the transcription is done through PicoVoice which improves the latency by a great deal.

Creating the Policy Assistant

You can create your own assistant following the example within services/openai/create_policy_assistant.py which is a functional example for creating the assistant as per the example stated above. If you want to run this code locally, make sure you have the following environment variables intact, you can observe an addition of three set of new redis variables which correspond to the files Db shown in the diagram above. The files db stores the filenames along with their OpenAI file Id

ASSISTANT_ID=XXXXXXXXXXXX
OPENAI_API_KEY=XXXXXXXXXX
THREAD_ID=XXXXXXXXXXXXXXX

AWS_ACCESS_KEY_ID=XXXXXXXXXXXXXXXX
AWS_DEFAULT_REGION=XXXXXXXXXXXXXXX
AWS_SECRET_ACCESS_KEY=XXXXXXXXXXXX

REDIS_ENDPOINT=XXXXXXXXXXXXXXXX
REDIS_PORT=XXXXXXXXXXXXXXXXXXXX
REDIS_PASSWORD=XXXXXXXXXXXXXXXX

REDIS_ENDPOINT_FILES_DB=XXXXXXXXXXXXXXXXXXXXXXX
REDIS_PORT_FILES_DB=XXXXXXXXXXXXXXXXXXXXXXXXXXX
REDIS_PASSWORD_FILES_DB=XXXXXXXXXXXXXXXXXXXXXXX