Guide · The ontology
Extraction runs
In a nutshell
An extraction run is one pass of the lab over a set of filings, turning their text into facts. Every fact records the run that produced it, and the run records the model and prompt version used, so a result can be traced and compared with a later run.
What happens in a run
- The run takes the passages of each release.
- A language model is given the passages and the ontology's vocabulary, and must answer in a structure generated from the ontology, citing a passage for every item.
- The answers are checked: numbers must appear in their passage, quotes must be verbatim, and each fact must fit the ontology's rules. Anything that fails goes back for correction, at most twice, and whatever still fails is dropped.
- Each release's facts are loaded into the graph, replacing that release's earlier facts.
What a run records
The model it used and the version of the prompt. Facts from different runs can then be told apart, and the effect of a new prompt or model can be measured by running again and comparing.
On the page
The page shows one run. Its name is under Extraction run in a fact's panel.