Add benchmark wireframes and extended evaluation flow

[nicooxxx]

Overview

This PR adds wireframes for the Benchmark module and extends the UI concepts from #118.

The goal is to visualize the broader benchmark workflow before implementing further UI changes.

The wireframes are based on:

• the current Benchmark UI and flow from Improve Benchmarking UI and dataset alignment flow #118
• the defined Benchmark use cases
• feedback and discussed requirements
• the current constraints around benchmark datasets, questions, question types and coverage

Flow covered

The wireframes currently cover the following flow:

Upload & map → Evaluation → Results → Annotation

This includes:

• Uploading a dataset
• Aligning dataset columns with the benchmark format
• Reviewing the detected mapping
• Choosing between Ranking and Classification
• Checking how many questions can be evaluated
• Configuring and starting an evaluation
• Reviewing evaluation metrics
• Revisiting saved evaluation results
• Comparing multiple evaluations
• Annotating retrieved evidence

Upload and mapping

The Upload & map screen combines dataset preparation, alignment and mapping review.

The user can:

• upload their own dataset
• see the expected benchmark structure
• align their columns with the benchmark fields
• validate the alignment
• review the mapped dataset before continuing
• save the prepared dataset for later evaluations

Evaluation

The Evaluation screen provides a clearer distinction between Ranking and Classification and briefly explains what both modes evaluate.

The current Ranking flow also shows benchmark coverage before the evaluation starts.

For example, the user can see:

• how many questions match the benchmark set
• how many questions will be included
• which questions are missing or have a type mismatch
• that lower coverage can make the benchmark result less representative

Results and comparison

The Results screen shows the metrics of a completed evaluation and allows saved evaluation runs to be compared.

The current wireframe includes:

• evaluation metadata and benchmark coverage
• Ranking metrics such as MAP, MRR, Precision@K and Recall@K
• comparison of multiple saved evaluation runs
• metric differences between selected runs

Annotation

The Annotation screen explores a question-level annotation workflow.

The user can:

• select a saved evaluation
• review retrieved evidence for individual questions
• mark the relevance of retrieved chunks
• add an annotation note
• move between questions
• revisit saved annotations

Constraints considered

The wireframes are informed by the current benchmark constraints, including:

• evaluated reports and questions need to be comparable with the benchmark dataset
• question types may need to match
• scales may need to be aligned
• partial evaluation of reports or questions should be communicated clearly to the user

Not all of these constraints are represented one-to-one in the current screens yet and they should guide further iterations.

[suung]

#127 @nicooxxx maybe it makes sense to check that PR and see if we update main firsst with it / build on top it

[nicooxxx]

Yes, that's true. I will review the first PR for the merge.