fix(js/vertexai): dedupe, improve types, fix rouge

[10kc-emilyk]

While investigating the appropriate types for vertex evaluators, I'm adding them here too, and discovered an incorrect type used for the ROUGE score evaluator. Also there was some code duplication with the types.

Checklist (if applicable):

• PR title is following https://www.conventionalcommits.org/en/v1.0.0/
• Tested (manually, unit tested, etc.) not yet, but i'll be doing some manual testing to verify that rouge was failing before this change
• N/ADocs updated (updated docs or a docs bug required)

[MichaelDoyle]

/gemini summary

[MichaelDoyle]

/gemini review

[gemini-code-assist]

Summary of Changes

This pull request refactors the Vertex AI evaluation plugin to improve type safety and maintainability. By leveraging official protobuf definitions and centralizing shared types, the codebase is now more robust and less prone to configuration errors. Additionally, the changes address specific implementation bugs in the ROUGE evaluator and fix minor naming inconsistencies.

Highlights

• Type Safety Improvements: Replaced generic 'any' types with specific Google Cloud AI Platform protobuf types for metric specifications across all evaluators.
• Code Deduplication: Centralized evaluation metric types and configurations into a shared 'types.ts' file, removing redundant definitions from the main evaluation logic.
• Bug Fixes: Corrected the ROUGE evaluator implementation to use the expected array format for instances and fixed a typo in the 'FLUENCY' metric enum value.

Activity

• The pull request was created by 10kc-emilyk.
• The author identified and addressed code duplication between evaluation files.
• The author noted a typo in the 'FLUENCY' metric enum value.
• A summary and review were requested via automated commands.

[MichaelDoyle]

@ssbushi can you take a peek?

[gemini-code-assist]

Code Review

This pull request enhances type safety within the Vertex AI evaluation plugin by migrating evaluation types to a separate file and replacing any types with specific Protobuf definitions from @google-cloud/aiplatform. The evaluator factory and specific evaluator functions were updated to utilize these stricter types. Review feedback pointed out that type-casting datapoint.reference as a string in the BLEU and ROUGE evaluators could mask runtime errors if the value is undefined, suggesting the addition of explicit runtime validation checks.

[ssbushi]

Thank you for your contribution. I made a few trivial tweak to get this ready for merging