Vibe-coded ingestion of stuff from 13 polymer papers

[justaddcoffee]

Again just an experiment to see how well this works

@realmarcin @cmungall @ct-parker @davramov @azaidi06

[justaddcoffee]

Prompt:
"I want you to make nodes and edges for our KG project using 1) a directory of 83 scientific papers in ./polymer_papers (in the base directory) and 2) the data schema for this project, which lives here: src/matkg/schema/matkg_schema.yaml. You can make as many nodes and edges as you want, but they should be supported by the scientific paper. Prioritize nodes and edges that seem most important, and that reoccur in several papers. There are lots of examples of nodes and edges in kg/. "

[justaddcoffee]

Some observations:

• this didn't produce as many nodes / edges as I'd have expected
• new associations have DOIs, which is good
• we should watch out for hallucinations in has_evidence - for 1-2 spot checked examples, the exact string did not seem to be present in the pdf, and in some cases may be dubious?

[justaddcoffee]

@davramov @ct-parker @realmarcin @cmungall @azaidi06
I've rolled back previous vibe-coded papers, then re-vibe coded 13 papers:

10.1002/adfm.201002014
10.1002/adfm.201304216
10.1002/adfm.201801874
10.1002/adfm.201902238
10.1002/adfm.202112511
10.1002/adfm.202305611
10.1002/advs.201500095
10.1002/advs.201600117
10.1002/adma.201604603
10.1002/adma.201803045
10.1002/aenm.201701942
10.1002/aenm.201702831
10.1002/anie.202115585

with this (similar but better) prompt:

I want you to make nodes and edges for our KG project using 1) a directory of 83 scientific papers 
in ./polymer_papers (in the base directory) and 2) the data schema for this project, which lives 
here: src/matkg/schema/matkg_schema.yaml. Process each paper one by one. For each paper, 
you can make as many nodes and edges as you want, but they should be supported by the 
scientific paper. Prioritize nodes and edges that seem most important to the overall point of 
the paper. There are lots of examples of nodes and edges in kg/.

and then rebuilt the kg (106 nodes and 71 edges), which is here:
kg/matkg_graph.json

Maybe we could have a look at some specific papers and the nodes/edges it extracted, just for a manual QC check