Review and update the current state of our AI integration

[jdunkerley]

Need to review the work previous done by Marcin and Jan.
Look at what works / doesn't as it stands.

Goal of this ticket is to understand where we currently are and then we can work out a plan to get the functionality back to fully capable.

[jdunkerley]

@JaroslavTulach please add references:

• AI Completions #5910
• discussing REST calls in Visualizations

[enso-bot]

Dmitry Bushev reports a new STANDUP for today (2023-11-24):

Progress: Started working on the task. Obtained an API key. Investigated how the AI completion is enabled in the language server. Investigated how the AI completion is implemented in the IDE. Found out how to open the AI completion prompt. Played with the completion a bit It should be finished by 2023-11-27.

Next Day: Next day I will be working on the #8348 task. Continue working on the task

[4e6]

How to use

• The AI completion is enabled by providing the OPENAI_API_KEY environment variable. It is required for the Language server to make calls to openai.
• The AI completion prompt is toggled by cmd+tab key combination in IDE.
  ⚠️ on Linux, I have to hold the tab key and then press the ctrl button to make the shortcut work. There is an open issue about that Handle conflicting shortcuts in a nicer way. #7528

Implementation

The AI completion is implemented by creating a query from the user prompt and sending it to https://api.openai.com/v1/completions.

To form a query, each user prompt is wrapped in a template describing the table API

enso/distribution/lib/Standard/Visualization/0.0.0-dev/src/AI.enso

Lines 23 to 29 in 8516ed5

	ops_prompt = "Operations available on Table are: " + (ops . join ",")
	aggs_prompt = "Available ways to aggregate a column are: " + (aggs . join ",")
	joins_prompt = "Available join kinds are: " + (joins . join ",")
	base_prompt = ops_prompt + '\n' + aggs_prompt + '\n' + joins_prompt + '\n' + examples
	columns = self.column_names . map .to_text . join "," "Table[" "];"
	goal_line = "goal=" + goal_placeholder + "==>>`"
	base_prompt + '\n' + columns + goal_line

and providing examples

enso/distribution/lib/Standard/Visualization/0.0.0-dev/src/AI.enso

Lines 13 to 22 in 8516ed5

	Table["id","category","Unit Price","Stock"];goal=get product count by category==>>`aggregate [Aggregate_Column.Group_By "category", Aggregate_Column.Count Nothing]`
	Table["ID","Unit Price","Stock"];goal=order by how many items are available==>>`order_by ["Stock"]`
	Table["Name","Enrolled Year"];goal=select people who enrolled between 2015 and 2018==>>`filter_by_expression "[Enrolled Year] >= 2015 && [Enrolled Year] <= 2018`
	Table["Number of items","client name","city","unit price"];goal=compute the total value of each order==>>`set "[Number of items] * [unit price]" "total value"`
	Table["Number of items","client name","CITY","unit price","total value"];goal=compute the average order value by city==>>`aggregate [Aggregate_Column.Group_By "CITY", Aggregate_Column.Average "total value"]`
	Table["Area Code", "number"];goal=get full phone numbers==>>`set "'+1 (' + [Area Code] + ') ' + [number]" "full phone number"`
	Table["Name","Grade","Subject"];goal=rank students by their average grade==>>`aggregate [Aggregate_Column.Group_By "Name", Aggregate_Column.Average "Grade" "Average Grade"] . order_by [Sort_Column.Name "Average Grade" Sort_Direction.Descending]`
	Table["Country","Prime minister name","2018","2019","2020","2021"];goal=pivot yearly GDP values to rows==>>`transpose ["Country", "Prime minister name"] "Year" "GDP"`
	Table["Size","Weight","Width","stuff","thing"];goal=only select size and thing of each record==>>`select_columns ["Size", "thing"]`
	Table["ID","Name","Count"];goal=join it with var_17==>>`join var_17 Join_Kind.Inner`

The query looks like:

Operations available on Table are: <example code>
Available ways to aggregate a column are: <example code>
...
goal: <user prompt>

Then based on the provided examples, ChatGPT tries to come up with a code sequence that would satisfy the goal.

It works better than I expected given how little information is provided to ChatGPT

enso-ai.mp4

[4e6]

Related PRs:

• AI Completions #5910
• New Prototype AI Searcher #7146