Turn court filings into a checked table

The sample has six fictional civil filings from the same court, as PDFs, and expected.csv: a table of the right answers for each one, filled in by hand.

Click Open in Frisket Desktop or Open in Frisket Cloud at the top of this page, then Open in Frisket in the dialog. It imports both for you: the filings into one sheet, one row each, and expected.csv into a sheet called Expected.

If you import the files yourself, import expected.csv as its own sheet. It's the answer key you'll check the model against, so don't give it to the model as input.

Before a model can read the filings, Frisket needs their text. On Check text, click a filing and compare the text on the right with the page. These PDFs were made on a computer, so their built-in text is good: names, case numbers and line order all match the page.

12

On the Analyze tab, click Extract fields. You describe the facts you want, one column each, and a model reads each filing's text and fills them in. Set it up as below. The descriptions matter: the model follows them literally, so say exactly what you want, including the format for dates.

Find it under Analyze › Extract › Extract fields

1

Source

The text column the import created.

Set to: text

Is everything you need in the text column?

Yes Some of it is in the layout Some of it is missing from the text

Extract from the text. You get citations you can check.

Extract from text or from the PDF? →

2

Engine

A cheap, fast model. Start cheap and move up only if the check below shows it getting fields wrong.

Set to: Gemini 3.5 Flash-Lite

What did the sample result show?

Fields match the source Wrong or missing values Values absent from the source

Keep those settings for the next rows, then check their results too.

Which model should extract your fields? →

12

3

Columns it creates

Four columns: case_number, filed_at as a date, plaintiff and defendant_as_filed. Each has a description telling the model what to copy and how.

Name	Type	Description
case_number	text	The case number exactly as printed in the filing.
filed_at	date	The filing date in ISO format YYYY-MM-DD. Leave blank if absent.
plaintiff	text	The plaintiff or petitioner exactly as filed.
defendant_as_filed	text	The defendant or respondent exactly as filed.

3

4

Instruction

Tells the model to use only what the filing says, copy names exactly, and leave a field blank rather than guess.

Set to: Read only the supplied filing. Copy names and case numbers exactly. Use YYYY-MM-DD for dates. Extract only facts explicitly stated in the source; leave missing values blank. A requested hearing or remedy is not a court decision. Do not use outside knowledge.

5

Citations

Records the passage each value came from, so you can check it in Review.

Set to: Cite sources

45

Before running anything for real, try it on a few rows. Select the first three filings, then click Preview. The values appear in the sheet but aren't saved; close the preview and they're gone. Read them against the filings. If a field is wrong, change its description and preview again.

Now save results for the same three filings, so you can score them. Keep them selected, open the menu beside Run and choose Run on selected. The four columns fill in for those rows; the other three show as incomplete.

Since you have the right answers in the Expected sheet, let Frisket score the run.

1. On the Tables tab, click Join expected answers and set it up as below. This pairs each filing with its row in the answer key, and each extracted column with its answer.
2. On the Analyze tab, click Compare and choose the run whose name ends in vs expected.
3. Click each field to see which rows matched.

When the sample scores well, run the same settings on every filing. Open Extract fields again with the same settings, open the menu beside Run and choose Run all. Because the sample columns already exist, this run makes a new set with _2 on the end of each name, such as case_number_2. Your sample results stay as they were.

The _2 columns hold the full run, six rows. A good score on three filings doesn't guarantee the other three. Check them against the Expected sheet, and open the PDF for any value that differs.

To check values against their sources one by one, click Review in the status bar at the bottom of the window. If you have more filings like these, review a random sample of them before you rely on the column.