Palimpzest test case: processing materials science papers for crystal recipes

[mikecafarella]

We have a colleague who wants to process papers from the materials science domain and extract "synthesis recipes". These recipes are passages that describe fairly complicated procedures for creating a novel chemical. This is useful because the recipes are extremely intricate and hard to discover. If we can build a model that suggests high-quality recipes for novel targets, it would be a big step.

The near-term goal is simply to extract these recipes from existing papers. So we want to populate a schema that looks like this:
(PaperIdentifier, TargetChemical, RecipeText)

After that works, we can populate a structured form of the recipe description. But just getting the raw text first would be helpful.

I have some annotated data we can use, though it's not shareable via Git so please don't commit it here.

Doing a basic but good job here involves:

1. Certainly writing code that extracts content from PDFs
2. Evaluating the accuracy of the initial task

After this basic version works, we want to evaluate runtime performance and maybe consider:

1. Using the PZ RAG operators like retrieve, but this depends on the runtime of the basic
2. Building a RAG index over the input papers

Finally, we would pursue the structured form of the recipe. I can share the domain experts' proposal for this structure.

[mdr223]

Next step: try to propose structured schema for info we want to extract from highlighted text