DiscoveryVIP
All guides
AI DOCUMENT-TO-DATA AUTOMATION STUDIO

Turn documents
into data you can defend.

From Unstructured Documents to Reliable, Actionable Data.

Define the fields. Find the evidence. Catch the errors. Build a document workflow that keeps you in control from the first upload to the final spreadsheet.

Try an invoice →Learn the workflow
▤   A document is a starting pointText · PDF · CSV · JSON · email
↓ extract · normalize · validate
total 113.00Source found · arithmetic checked
↓ review before export
✓   A record you have verifiedIllustrative workflow, not live results
Local-first · free learning studioNo automatic AI uploads. Your documents stay in this browser until you approve a connected extraction.
01 / COLLECT YOUR SOURCES

Start with a document, not a guess.

Choose an original fictional template to load its schema, documents and verified reference data into a new workspace.

Paste a document or email

PDF, TXT, CSV, JSON, Markdown: up to 5 MB per file; text PDFs up to 100 pages. PNG/JPEG: up to 2 MB, extraction requires connected Gemini vision. Scanned PDFs need external OCR first.

Document library

02 / DEFINE WHAT TO EXTRACT

A clear schema is your contract.

Generated JSON schema

This transport schema permits null for missing facts. Local validation additionally enforces required values, formats, ranges and cross-field rules.

Rules: {"type":"sum","parts":["subtotal","tax"],"target":"total","tolerance":0.01} or {"type":"dateOrder","first":"invoice_date","last":"due_date"}.
Visual field editor
Use the same field definitions as the schema export. Tables are arrays of these row objects. Lists contain text items.


03 / PROCESS, NEVER BLINDLY TRUST
Follow every stage.
Choose a stage to inspect how this workflow handles your data.
Local extraction recognizes exact field_name: value lines and CSV/JSON keys. It does not understand arbitrary prose. Each connected batch asks before transmitting source text or images. Processing is sequential to keep requests bounded.
Ready.


04 / SOURCE EVIDENCE & HUMAN REVIEW
Check the value. Open the proof.
Source document
Choose a record.
Process a document to see its extracted fields and evidence.


05 / TRANSFORM WITH A PREVIEW
Shape the data without hiding the changes.
Transform all active, non-rejected records. Preview first. Applying changes resets review, preserves an audit entry, updates relevant schema fields and revalidates the results. Deduplication keeps the first record; filtered-out records are marked rejected rather than silently deleted.
No preview yet.


06 / COMPARE EXTRACTION APPROACHES
Measure against a known reference.
Save a configuration, change the schema, prompt or model, then save the other. Comparison uses documents with verified expected data (included with templates). It does not replace your reviewed records.
Accuracy means exact equality to verified reference fields after extraction normalization. It is not a model confidence score. Local processing has no provider cost. Estimates use entered prices and reported input/output usage only.


07 / EXPORT REVIEWED RECORDS
Make the output useful.
Only accepted records that still pass validation and duplicate checks are exported. Rejected, unreviewed and superseded records stay out. Nested data is serialized as JSON in spreadsheet cells.
Spreadsheet workflow
Upload the exported JSON file to your own Google Drive, then run the included Apps Script from a Google Sheet. Review its permissions first. This studio does not write to Google Sheets or external APIs directly.
Download working Apps Script importer
Workflow controls
Validation and review remain mandatory. The exported recipe records the schema, rules, normalization choice, approach and instructions, without credentials.


08 / LEARN BY CORRECTING REAL EXAMPLES
Nine practical skills for dependable extraction.


Optional AI connection & privacy
AI and image extraction are unavailable until a Gemini gateway is configured.
Your sources, records, corrections and schemas are stored in this browser using IndexedDB. They are not encrypted or isolated from other users of this browser profile. Export a backup before clearing storage. Deleting a workspace removes its local sources and records.
Documents are untrusted data. The model receives a separate extraction policy and no tools. Exact quotation checks detect nonexistent quotes, but do not prove that a quotation supports a value. Human review is still required. Never send confidential or personal information without authorization.