feat(core): Implement Dynamic Parameters within regular nodes used as AI Tools

[OlegIvaniv]

Summary

This PR introduces a new feature that allows LLM to dynamically interact with n8n nodes when used as tools. The core of this feature is the new $fromAI() function, which allows for the definition of parameters that can be populated by AI models during workflow execution.

CleanShot.2024-09-17.at.21.50.55.mp4

Key Changes

1. Introduced a new $fromAI() function for defining AI-powered parameters in node configurations.
2. Updated the WorkflowDataProxy to handle $fromAI() calls and retrieve values passed by LLMs.
3. Modified the createNodeAsTool function to support AI tool conversions, including automatic description generation based on resource and operation.
4. Added support for various data types (string, number, boolean, json, date) in AI-powered parameters.
5. Updated several nodes to be usable as AI tools.

Implementation Details

$fromAI Function

The $fromAI() function is implemented in the WorkflowDataProxy class and accepts four parameters:

1. name: A string identifier for the parameter (required)
2. description: A string describing the parameter's purpose (optional)
3. type: A string specifying the data type (optional, defaults to 'string')
4. defaultValue: A fallback value if no AI input is provided (optional)

Example usage:

"inputText": "={{ $fromAI('inputText', 'Enter main text to process', 'string') }}"

Parsing and Schema Generation

• The parseFromAIArguments function uses a regex to extract all $fromAI() calls from the node parameters.
• Each extracted call is parsed into a FromAIArgument object using the parseArguments function.
• The generateZodSchema function creates a Zod schema based on the parsed arguments, which is used for validation and type-checking.

Node Conversion to AI Tools

The convertNodeToAiTool function has been updated to:

• Modify the node's name, displayName, and I/O configuration for AI tool usage.
• Add a descriptionType property allowing users to choose between automatic and manual description generation.
• Insert a toolDescription property for manual description input when needed.

WorkflowDataProxy Updates

• The $fromAI method in WorkflowDataProxy retrieves values passed by LLMs from the runExecutionData.
• It includes validation for parameter names and handles cases where no execution data is available.

CodeMirror Integration

• Added $fromAI() to the list of root dollar completions in the CodeMirror plugin.
• Updated the completion info to include examples and parameter descriptions.

Node Updates

Several core nodes (including Airtable, Baserow, Google Calendar, Gmail, Google Sheets, HackerNews, Postgres, Supabase, Telegram) have been updated with a usableAsTool: true property to indicate their compatibility with this new feature.

Review / Merge checklist

• PR title and summary are descriptive. (conventions)
• Docs updated or follow-up ticket created.
• Tests included.
• PR Labeled with release/backport (if the PR is an urgent fix that needs to be backported)

[jeanpaul]

Bravo! This is awesome work!

I have a couple of smaller comments / questions, but one thing that we definitely need to address is the parsing of the arguments to $fromAI. One thing that came to mind is that we may need to do more proper parsing, because users might be inclined to do things we don't anticipate yet, that are valid javascript.

Also, as discussed, I encountered an issue with Claude, when returning plain results from the HN node; I'm not sure what to do there, but maybe we need to wrap the results in something that's more easily digestible for some models?

[jeanpaul]

Does this handle escaped characters properly? I tried it, and got some weird results:

argsString: 'database-filter', 'the filter\\'s keyword to apply to the database', string, ['a\\'b', 'b']
result: {
  key: 'database-filter',
  description: "the filter\\\\'s keyword to apply to the database', string, ['a\\\\'b",
  type: "b']",
  defaultValue: undefined
}

[OlegIvaniv]

I improved the parsing method and added a few more tests; could you have another look?

[OlegIvaniv]

Thanks for the review @jeanpaul! I improved the parsing logic and added a few more tests, could you re-review, pls?

[jeanpaul]

Very nice! Awesome expansion of the tests as well.
Left another few small comments, but other than that looks great.

[jeanpaul]

should we add backticks as quote-chars as well?

[OlegIvaniv]

Updated + added unit tests

[OlegIvaniv]

@jeanpaul I updated the PR with the fixes product and reorganized unit tests + removed some duplicates. Could you have one last look? 🙏

[jeanpaul]

One small gotcha (and a quick check for my understanding :))

[github-actions]

✅ All Cypress E2E specs passed

[cypress]

n8n    Run #7127

Run Properties:   Passed #7127  •  193f913c0d: 🌳 🖥️ browsers:node18.12.0-chrome107 🤖 OlegIvaniv 🗃️ e2e/*

Project	n8n
Branch Review	ai-285-implement-placeholders
Run status	Passed #7127
Run duration	04m 33s
Commit	193f913c0d: 🌳 🖥️ browsers:node18.12.0-chrome107 🤖 OlegIvaniv 🗃️ e2e/*
Committer	Oleg Ivaniv
View all properties for this run ↗︎

---

Test results
Failures	0
Flaky	0
Pending	0
Skipped	0
Passing	437
View all changes introduced in this branch ↗︎

[janober]

Got released with [email protected]