OpenAI has a number of models available through APIs. OpenAI agent is basically a wrapper around models offered. The agent has a webservice component that works with the APIs. To create an OpenAI agent:
$ cd agents/openai
$ python src/openai_agent.py --session <session_id> --properties '{"openai.api":"ChatCompletion","openai.model":"gpt-4","output_path":"$.choices[0].message.content","input_json":"[{\"role\":\"user\"}]","input_context":"$[0]","input_context_field":"content","input_field":"messages"}'
As you see above details of the model (e.g. gpt-4), api (e.g. ChatCompletion), etc. are provided as input to the agent. In addition to prepare input data for the API and to fetch response you need to provide additional properties such as input_field, input_context, input_concext_field, and output_path. input_field is the name of the field to where text data is provided to the API (e.g. prompt for text_davinci_003, messages for gpt-4). output_path is json_path to where the API returns the resulting text in the response (e.g. $.choices[0].message). Some of the API calls require more structured input such as ChatCompletion. The JSON structure is specified in input_json, and where to insert input text data is provided as json_path in input_context and the associated input field name is provided in input_context_field. Note that any properties with the prefix openai. will be passed on to OpenAI API, everything else is used by the agent itself for data processing.
For the OpenAI agent you would also need to start a webservice that will proxy to OpenAI APIs. To deploy the service :
$ cd services/openai
$ ./deploy_service.sh --port_mapping 8001:8001 --service openai --image blue-service-openai
Make sure the openai.service property of the agent is set to ws://blue_service_openai:8001 given above deployment of the OpenAI service.
To build the service:
First, edit openai.env and replace <openai_api_key> with your API Keys.
Then, build the service:
$ cd services/openai
$ ./docker_build_service.sh