Elicit now reviews the entire paper to get information for your table!
If you add the column "percentage of trials which met the inclusion criteria" to this query "What are the benefits of meditation?" Elicit finds the answer "3%" for one of the papers.
If you click on the 3%, you can see the part of the paper Elicit is most likely drawing from. You can also click the DOI or PDF links (if available) to confirm in the original text.
To get information from the body of the paper, we use semantic search to find the most relevant paragraph and then extract information from the paragraph using this prompt.
How well it works
You may not encounter answers from the full text very often:
Only ~ 30-40% of the papers in Elicit are open access papers whose texts are available
When you add a column to your table, we still mostly rely on the abstract to answer it. A rough heuristic is that ~ 85% of the answers you'll in the table for columns that are not metadata will come from the abstract. The remaining 15% or so will come from the full text.
We then estimate that these extractions are ~ 85% accurate so you do need to be careful about double checking the model's work. The context and links to original text should help.
All of this together means that you'll see a useful answer using the full text for ~ 1-5% of the cells in your table.
You'll be experimenting with this feature alongside us. If you find any mistakes, please let us know by clicking on the link at the bottom of the Context tab to send an email (or message us in Slack).