Rerun and resume
pydelity saves each result in the output folder as soon as it has it. When you run it again on the same folder, it reuses what still fits and only redoes the rest. So you never pay twice for the same survey, and a run that stopped can carry on where it left off.
What gets reused
Section titled “What gets reused”The measured facts (pdf_facts.json) are reused as long as the PDF’s
content is unchanged. pydelity compares a SHA-256 hash of the file, so renaming
or moving the PDF doesn’t matter here. A newer pydelity that measures PDFs
differently measures them again.
The survey (pdf_survey.json) is reused as long as everything it was made
from is unchanged:
- the PDF’s content and its file name
- the measured facts
modelinstructionsandlanguagesadvanced.survey.effort,check_retriesandmax_output_tokens- the survey instructions built into pydelity, which a newer release may change
Changing any of these runs the survey again, at full cost. Changing anything
else, such as max_cost, provider_retries, request_timeout_seconds or
expiry_hours, keeps the saved survey.
By default the output folder is named after the PDF, so a renamed PDF also gets
a new, empty output folder unless you set output_dir.
The upload of the PDF to the provider (uploads.json) is reused while at
least an hour is left before the provider deletes it. See
How long the provider keeps your PDF.
When a run reuses the survey, it makes no model request and costs nothing. The
command’s summary shows the step as reused.
Redo everything
Section titled “Redo everything”To measure and survey again even though nothing changed, use overwrite:
pydelity document.pdf --overwriteimport pydelity
result = pydelity.run("document.pdf", overwrite=True)The new survey costs as much as the first one. The upload is still reused.
After a run that stopped or failed
Section titled “After a run that stopped or failed”A run can end early because you answered no, it hit max_cost, the provider
failed, the survey had no usable answer or you pressed Ctrl+C. In each case
run.json records how it ended and why, and anything already finished stays
saved.
To continue, run the same command again. pydelity reuses the saved facts, upload
and survey, and redoes only what is missing. Money already spent in the folder
still counts toward max_cost.
Each run replaces run.json and report.html, so they describe the latest run.
cost.jsonl keeps every run’s requests.
Damaged or leftover files
Section titled “Damaged or leftover files”pydelity writes each file in full to a temporary file and then moves it into place, so an interrupted run never leaves a half-written output. Temporary files left by an interruption are removed at the start of the next run.
If a saved file is damaged anyway, for example after being edited by hand,
pydelity warns you and redoes the step that writes it. A damaged last line in
cost.jsonl is skipped with a warning.
One PDF per folder, one run at a time
Section titled “One PDF per folder, one run at a time”An output folder holds the results for one PDF. If you point a run at a folder
that holds results for a different PDF, pydelity warns you and replaces them.
The folder’s earlier spending still counts toward max_cost, because it stays
in cost.jsonl.
Run only one pydelity run on an output folder at a time. pydelity doesn’t detect two runs on the same folder, and they can overwrite each other’s files. Runs on different folders can go at the same time.
On Windows, pydelity can’t replace a file while another program has it open. If a run fails with an error naming a file in the output folder, close the program that has it open and run again.