How it works
Your assistant writes the rules once. The engine applies them every time.
Two jobs
The assistant authors. The engine executes.
Your AI assistant, once per document type
Reads the document, writes the rules that find each value, tests them against the real text, rehearses the extraction, and saves the result as a template. This is the part that needs judgement, and it happens with you in the loop.
The Docuoria engine, every run
Recognises the document, applies the template, validates the fields, and renders the file. No model is involved at this point. The same PDF with the same template produces the same bytes, on every run, on any machine.
That split is why a number in your spreadsheet can be traced to a rule and a position in the document, and why it is the same number next month.
Three terms
The three parts of every extraction
Match rules
How a document is recognised
Rules score whether a PDF is a known kind of document, from its text, layout, metadata, tables, or a combination. A template is chosen by score. Your assistant builds the rules from the document’s own text and structure, never from its file name.
Templates
The saved method for one kind of document
What to look for, where to find it, how to clean it up, and which fields the result must contain. Your assistant writes one per document type and stores it beside your PDFs.
Output
The file your tools open
CSV for spreadsheets, with one row per line item. JSON for software, with the structure kept. Both are checked against the template’s field list before anything is written.
The workflow
The steps, most of them skipped for a known document
- 0
Survey
Folders onlyGroups the PDFs by structure and reports which ones a stored template already covers, so a folder of one vendor’s bills gets one template, not twenty.
- 1
Classify
Always firstScores the PDF against every stored template. A strong match skips straight to step 5.
- 2
Inspect
New document typesReads every page: the text as the engine sees it, the tables, the metadata. Totals on page 1 and line items on page 3 are both accounted for.
- 3
Test
New document typesProves each pattern against the real text before it goes into a template, and reports exactly where a pattern stops matching.
- 4
Build
New document typesWrites the template, then checks that it does not also match a different kind of document, and that its field list is valid.
- 5
Dry-run
Every new or changed templateExtracts without writing a file. Every required field must be filled, every declared list must have rows, and where the document prints them, subtotal plus tax must equal the total.
- 6
Execute
Every runRuns the template and writes the CSV or JSON. A run that leaves a required field empty is reported as incomplete, not as a success.
- 7
Store
New or changed templatesSaves the template and confirms it ranks first for its own document and does not match its siblings.
Questions
It asks you at most one question, and only about the outcome
Every technical decision is the assistant’s to make silently: how to find a value, whether to rehearse, which format to write. It asks you only when the document supports more than one reasonable result and your request did not settle it. The question is about what ends up in your file, offers two to four choices, and marks one as recommended.
Your file is a 4-page invoice. Page 1 has the summary totals; pages 2 and 3 list the individual charges. What should I put in your spreadsheet?
- ◉ Everything: the totals plus every charge (Recommended)
- ○ Just the summary totals
- ○ Just the list of charges
Checks
What happens before a number reaches your file
- Every page of the document is read, not just the first.
- Every pattern is tested against the document’s actual text before it is used.
- Every required field must be filled and every declared list must have rows, or the run is incomplete.
- Where an invoice prints them, subtotal plus tax must equal the total, and line amounts must sum to the subtotal.
- A row in your spreadsheet is the unit you asked for, never a summary standing in for the detail.
- A changed template is re-run against the documents it was built on, and any regression blocks the change.
- Amounts in different currencies are kept in separate columns or rows, never added together.
- The assistant reports what it did from the engine’s own output: files written, rows added, PDFs skipped.
Outcomes
Every run ends in one of three states
Succeeded
The pipeline completed and the file is written. The result also states whether every required field was filled.
Failed
A step stopped part-way. The result names the step and the field, so the fix is to that one thing.
Rejected
The engine refused the request: an unreadable PDF, a malformed template, or an output shape the format cannot hold. Nothing partial is written.
Folders and months
Batches and running ledgers
A folder of PDFs
The survey groups documents by structure. Each group gets one template. A batch run then routes every PDF to its template and merges the rows into one CSV, with a column naming the source file of each row. Documents that match nothing are listed, not silently dropped.
The same folder, every month
A ledger records which PDFs it already contains. Sweeping the whole folder again adds only the new files; recorded ones are skipped before any work is done. A rebuild that would drop recorded rows is refused, and a correction replaces one document’s rows in place.
See these as requests and outputs · The engine surface, for developers