FAQ
Browse by topic
Every FAQ section, grouped by topic. Click any item to jump to it.
Getting Started
Extraction Modes & Templates
Review & Evidence
Getting Started
Which workflow should I use?
Use Extraction when you want named fields back in a table from one or many PDFs, such as findings, dates, metrics, or conclusions.
Choose Flat mode for one row per file. Choose Relational mode when each PDF can contain repeated records such as devices, endpoints, cohorts, or line items.
How should I set up my first extraction?
Keep the first run small. Start with 2-3 fields you can verify quickly, then expand once the output looks right.
Use representative PDFs for the first run rather than the biggest batch. The goal is to validate the setup, not maximize volume on the first attempt.
How do I describe my document set?
Data type tells the extractor what kind of records the PDFs represent, such as safety reports, endpoint tables, study summaries, or requirement documents.
Write it the way you would describe the set to another person. This helps instruction generation and relational template inference stay on the right track.
Good Data type examples
post-market surveillance findings clinical study summaries contract payment terms
How should I use Document profile?
Leave Document profile on Auto for the first run. It is the safest default when you are still learning how the job behaves.
Use Structured Numeric for row-heavy or amount-heavy PDFs like invoices, contracts, or estimate tables. Use Research for journals, studies, and narrative evidence extraction.
How should I choose columns for the first run?
Columns define the exact fields you want returned in the output table.
For the first run, choose fields that are easy to spot and easy to verify. Add interpretation hints only when a field name could be read more than one way.
Extraction Modes & Templates
When should I use Relational mode?
Use Relational mode when one PDF contains repeated linked entities, such as multiple devices, endpoints, cohorts, or adverse events.
Relational output can contain multiple records per PDF, and each record carries the same required field keys.
Relational mode is currently in Beta — the record-template inference is still being refined; flag any unexpected groupings so we can tune it.
Example output: one trial PDF, two device rows
| Record | Device name | Manufacturer | Risk class | |
|---|---|---|---|---|
| trial.pdf | device-1 | CardioSense Monitor | ACME Med | Class IIb |
| trial.pdf | device-2 | NeuroTrack Sensor | ACME Med | Class III |
Example template for repeated devices
record=device fields=device_name,udi,manufacturer,risk_class cardinality=multi
Example output (multiple records from one PDF)
{
"records": [
{
"record_type": "device",
"cells": {
"device_name": { "value": "CardioSense Monitor" },
"udi": { "value": "00812345678901" },
"manufacturer": { "value": "ACME Med" },
"risk_class": { "value": "Class IIb" }
}
},
{
"record_type": "device",
"cells": {
"device_name": { "value": "NeuroTrack Sensor" },
"udi": { "value": "00812345678944" },
"manufacturer": { "value": "ACME Med" },
"risk_class": { "value": "Class III" }
}
}
]
}When should I use Flat mode?
Use Flat mode when you need one consolidated row per PDF.
This is best for document-level summaries where repeated entities do not need separate rows.
Example output: 3 invoices in, 3 rows out
| Invoice # | Total | Due date | |
|---|---|---|---|
| invoice_apr.pdf | INV-1042 | $4,820.00 | 2026-06-01 |
| invoice_may.pdf | INV-1078 | $1,200.00 | 2026-07-01 |
| invoice_jun.pdf | INV-1095 | $9,150.50 | 2026-07-15 |
Example flat output (one row per file)
PDF file | study_id | primary_endpoint | adverse_events trial_a.pdf | ST-101 | 12-month MACE rate 4.3% | Not found trial_b.pdf | ST-102 | 6-month restenosis 2.1% | Mild bleeding (3 events)
When should I care about the Record template (DSL)?
Most users can rely on the auto-generated "What we’ll extract" summary shown in the new-job form — it describes the same template in plain English.
The DSL only matters when you are doing relational extraction and need precise control over how repeated entities are grouped. Open the advanced editor in the form to access it. `record` is the entity anchor, `fields` lists the required field keys, and `cardinality` sets whether you expect one record (`single`) or many (`multi`).
Example A: device-level extraction (multi)
record=device fields=device_name,udi,manufacturer,risk_class cardinality=multi
How that appears in review table
IFU_2025.pdf | device-1 | device | CardioSense Monitor | 00812345678901 | ACME Med | Class IIb IFU_2025.pdf | device-2 | device | NeuroTrack Sensor | 00812345678944 | ACME Med | Class III
Example B: trial summary (single)
record=trial fields=trial_id,primary_endpoint,safety_summary cardinality=single
When should I add Rules?
Rules are optional setup instructions that influence how values are interpreted and normalized.
Keep them concrete. For example: use endpoint table first, normalize percentages, return Not found when unsure.
How do saved instruction setups work?
After reviewing an instruction pack, choose Save setup for reuse. The saved setup contains the fields, extraction settings, and reviewed instructions, but not uploaded PDFs, document samples, or raw AI responses.
When you select that setup for a later job, Extract reuses the saved instructions without calling the instruction-generation AI. You still review and approve the setup before extraction begins, and the extraction AI still processes each new PDF.
Review & Evidence
What am I checking during instruction approval?
Check three things: the objective matches the document set, the field plans match the outputs you expect, and the missing-value behavior is strict enough.
You do not need to rewrite everything. If something is off, revise with one narrow sentence such as "Prefer endpoint table wording" or "Treat score as numeric plus unit when present."
You can request up to 2 instruction-pack revisions for each job.
How do I verify a result?
Select a populated or otherwise reviewable value in the results table to open its evidence view. Review the source quote and page when available. A highlight may be exact, approximate, or unavailable.
If the document does not support the result, use the feedback actions to flag it or suggest a correction.
Does every result include a quote, page, and highlight?
No. Evidence is shown when an extracted value can be grounded to the source PDF. Some cells may have a value without a complete quote, page reference, or positional highlight.
Scanned or image-only PDFs often contain little embedded text, so their highlights can be sparse or unavailable. Approximate highlights and best-matched excerpts are labelled in the evidence view.
Extraction is AI-assisted and can be incomplete or incorrect. Check important results against the original PDF before relying on them for publication, financial, legal, clinical, or regulatory work.
What happens if one PDF in a multi-file job fails?
The job can finish as Partially complete. Results from PDFs that succeeded remain available for review and CSV export when your plan allows it.
A failed PDF is identified in the results view and is excluded from evidence review because no grounded highlights were generated for it. Re-export or correct that PDF, then submit it in a new job.
Files, Privacy & Limits
Can I extract data from scanned or image-only PDFs?
Yes. Direct PDF is the recommended and currently available handling mode for both normal text PDFs and scanned or image-only PDFs. Text-only mode is temporarily unavailable while local parsing is improved.
Direct PDF sends the complete original PDF to the AI provider configured for extraction so it can process embedded text, page images, scans, and visual tables.
Instruction setup still relies on locally extracted text samples, which can be sparse for scans. Extraction may still work, but positional highlight boxes may be limited or unavailable.
What files are allowed?
Upload PDF files only. Extraction jobs accept up to 10 PDFs per run, with a maximum size of 15 MB for each file.
The app must be able to verify the PDF binary and determine its page count. If it cannot, re-export the document as a standard PDF and try again.
Do not upload confidential or sensitive documents.
Who can access my files, and where are they processed?
Viewing a source file requires your authenticated account, and the server checks that the file belongs to one of your jobs.
Uploaded files are stored in the service cloud storage during the retention window. In Direct PDF mode, the complete original PDF is also sent to the AI provider configured for extraction.
Use Extract only for public or non-confidential documents that you are authorized to process. It is not a HIPAA-compliant environment or a GDPR compliance tool.
What are the demo, trial, plan, and download limits?
The public interactive sample requires no account or card. The signed-in demo workspace is read-only: you can explore the workflow and starter templates, but you cannot upload files, run live extraction, or download results.
The Stripe-managed trial is available once per account, lasts 14 days, and includes 20 pages in total, with up to 20 columns per job. A card is required. Trial results can be reviewed in the app, but artifact downloads are disabled.
Active paid plans allow up to 20 columns per job and use these monthly page quotas: Starter: 100 pages; Premium: 1,000 pages; Professional: 10,000 pages. A job must fit within the remaining page quota.
CSV is the only result download currently available, and it requires an active paid plan. Provenance JSON and highlighted-PDF downloads are not currently offered.
Are files and evidence deleted automatically?
Each job is retained for a limited period determined by our current retention settings. After that period, uploaded PDFs, extraction results, evidence, and generated artifacts are scheduled for automatic deletion.
Final deletion may be delayed while a job is queued or processing, or because of provider delays, retries, backups, or technical issues. You can delete a job earlier from its detail page, or delete your account data from Settings.