Her & Samantha

The name of this repo is reference from here.

Usage

Slackbot test

1. Make a copy of config_template.yaml and name it config.yaml.
2. Set the following variables to prevent any impact on product environment.
   1. pubsub_subscription.launch: False
   2. pubsub_publish.send_userID: test
   3. chat_hist_pubsub.topic: "trash-collector"
3. Modify memory_type, days_of_week, prompt, prompt, violation to suit your usage scenario
4. Modify pach.sh to suit your usage scenario.
5. Run sh pach.sh to upload your model registry
6. Deploy the endpoint on Vertex AI

Broadcast on Production

1. Make a copy of config_template.yaml and name it config.yaml.
2. Set the following variables to run AVLiver on the production environment.
   1. pubsub_subscription.launch: True
   2. pubsub_subscription.topic: "media17-live-events" or "media17-live-events-test" (The latter is for testing use)
   3. pubsub_subscription.pull_userID: <AVLiver userID>
   4. pubsub_publish.send_userID: <AVLiver userID>
   5. tts_endpoint.speaker_name: <speaker name>
   6. chat_hist_pubsub.topic: AI-Vliver-chat-hist
   7. (optional) Set moderator.launch: True to launch moderator.
3. Modify memory_type, days_of_week, prompt, prompt, violation to suit your usage scenario
4. Modify pach.sh to suit your usage scenario.
5. Run sh pach.sh to upload your model registry
6. Deploy the endpoint on Vertex AI

Plug-in (airflow & cloud function)

Airflow (option)

Airflow is used to tell AI-VLiver daily information that you want AI-VLiver to know in the prompt.

1. develop your owninformation_collector.py, which would collect information from real world and summary into a short message and return byinformation_collector.
2. modify daily_information_dag.py to set up your aivliver's endpoint.
3. put these codes to GCP composer: ai-vliver-daily-information to trigger airflow

airflow will inject message into variable: daily_mail in prompt.py

Cloud function (option)

Cloud function is used to remember the long-term memory in redis and inject it into variable: long_term_memory in prompt.py

cloud function will call LLM to extract useful information from each conversation to construct the structure data(dict) to save in the redis and inject into prompt ultimately.

1. write down your own LLM prompt in the prompt.py (give few shot example to LLM)
2. set up you AIVLiver's endpoint id
3. set a new key name to replace long_term_memory

---

Note: The name of this repo is reference from here.