Support text-davinci-003

[danneu]

The text completion api is useful because it's not censored like the chat completion apis. Also, sometimes pure text completion is simply useful.

My uncommitted implementation generates a chat simulation based on the db.listHistory() messages that I normally was feeding into openai.chatCompletions.

For example, I turn this:

• { role: 'user', content: 'Give me a random color.' }
• { role: 'assistant', content: 'Yellow.' }
• { role: 'user', content: 'Now tell me a random animal.' }

Into something like this:

User: Give me a random color.
Assistant: Yellow.
User: Now tell me a random animal.
Assistant: 

It leaves a hanging "Assistant: " so the textCompletion api is solving text completion for something the assistant should say.

• [x] /model should allow davinci selection
• [x] I guess GPT4_ENABLED doesn't apply to davinci.
• [ ] The davinci path should also consider token account (esp since the generated prompt string is longer than chatCompletion token) and also set max_tokens since I don't think the API does it for us

[danneu]

I'll close this issue when the TODO checklist is complete.

Since I'm not in the mood to handle davinci token count, I didn't want that item to block me from launching davinci support into production.