Output validation views

[shreyashankar]

One of the biggest pain points for users is to assess "correctness" of operations, at a high level. It does not need to be a number or accuracy fraction; that would just cause more questions (a la who validates the validators).

Our goal is to have the user sweep as much of the data/outputs as possible, and form a quick coarse assessment of accuracy.

In this vein, some of the following views may be interesting:

• Compare two (very different) documents and the LLM outputs, side by side. Here, it is hard to know what "very different" is---it is task specific; not necessarily two documents that have the smallest inner product in embedding space when embedding the entire documents. "Different" could mean structurally different documents, differences in a particular section of documents, etc.
• A visualization of a column/document attribute (e.g., histogram, bar chart). Discussed offline in research meeting 11/20.