01 / ACCEPT
Save the original.
The server checks the upload, its size and selected page range. Once accepted, the file and its job are saved together. Changing a field list creates a new definition version; earlier results retain theirs.
INSIDE TRACEWISE / EVIDENCE BEFORE APPROVAL
A reading you
can question.
Azure reads the document. Tracewise checks the evidence.
You decide what becomes an approved record.
01 / THE PROCESS
From original file
to reviewed data.
Your field list is the brief: names, descriptions, examples, data types and repeating tables. Tracewise uses it to organise a reading while retaining the original document, source evidence and review history.
02 / LOCATE
Read the page structure.
Azure Document Intelligence recognises words, layout, tables and selection marks. Word identifiers and page polygons supply the evidence used to locate extracted values.
03 / EXTRACT
Read the fields you chose.
Azure OpenAI receives the field definitions and recognised document evidence. It returns structured readings, raw text and source-word identifiers. Document instructions are treated as untrusted content.
04 / VERIFY
Check the reading.
Tracewise validates fields, rows, raw text and source references. Arithmetic and configured business rules flag concerns. Selected uncertain fields can receive an independent second model reading of the saved OCR evidence.
05 / REVIEW
Keep the judgment human.
Open the original beside the results, inspect source locations and review warnings. Correct a value or keep the printed reading with a reason. The initial extraction remains available.
06 / EXPORT
Use the approved version.
Confirmation saves a reviewed snapshot. Excel, CSV and JSON exports use eligible approved revisions, with the chosen list layout. Processing does not approve a document or download it for you.
02 / THE TWO AZURE SERVICES
Text evidence.
Structured interpretation.
The current deployment uses Azure Document Intelligence and the configured GPT-6 Astra deployment with high reasoning effort. These are separate jobs within one review workflow.
Model configuration is not a guarantee of accuracy. Processing locations and provider data-handling commitments should be checked against the deployed resources.
Document handlingDOCUMENT INTELLIGENCE / OCR
Recognised words and their positions.
The uploaded document is analysed for text and structure. Tracewise retains the recognised words, their identifiers and available page geometry. Some input formats can lack usable page positions.
AZURE OPENAI / FIELD READING
A structured answer to your field list.
The reader receives recognised evidence, enabled table information, field descriptions and reading settings. It has no action tools and cannot confirm a file. Product, customer and price datasets support application checks; they are not automatically added to the extraction request.
03 / AN INDEPENDENT READING
Ask again.
Keep both answers.
A second pass adds a comparison. It does not turn a doubtful source into a verified fact.
TRIGGER
Select uncertain fields.
Weak or unresolved source evidence and missing required readings can select fields for another pass. A supported reading does not automatically require a second model call.
ISOLATION
Withhold the first answer.
The same configured model reads the selected field definitions and saved OCR evidence again without receiving the first answer. This is a separate model reading, not new OCR or a fresh image-crop inspection.
COMPARISON
Preserve the primary result.
The saved comparison records both readings and whether value, source location and row presence agree. A disagreement adds a review warning. Agreement does not clear existing source warnings or create an approval.
The comparison is saved with the result. Background checkpoints retain its progress across retries. The dated live invoice studies on this site were recorded before this pass was added; they are not evidence that those historical files received it.
04 / CHECKS THAT EXPLAIN THE ISSUE
Warnings inform.
They do not rewrite.
Checks cover missing required values, unreadable numbers or dates, repeated files and inconsistent sums. Purchase-order checks also use the catalogue, customer and price data you supply.
Tax-rate comparison is optional and off by default. Generic arithmetic still compares subtotal plus printed tax with total. A configured rate is your business rule, not a legal determination.
Printed values are retained even when a calculation disagrees. A reviewer resolves the issue; blocking purchase-order checks must be cleared before confirmation.
Purchase-order checksA source box comes from evidence.
The application validates the returned word identifiers and raw text against the OCR evidence. Available matching word polygons supply a box. The model does not invent coordinates. Repeated matches, weak evidence or missing geometry remain visible for review.
Missing stays missing.
An absent reading is represented as Not on the page. An extracted value without a verified location is Place not found. An ambiguous date can retain its printed form and receive a warning rather than being guessed.
Your decisions stay separate.
A correction retains the original reading. Confirmation records the reviewed revision and reviewer provenance. Rereading does not silently replace reviewed corrections, and no AI stage confirms the file.
05 / THE WORDS BESIDE EACH VALUE
A clear status.
A next action.
Labels and explanations accompany color, so a status does not depend on color alone.
On the page
A supported source match has usable evidence and no remaining issue that changes its status. Inspect the source before relying on the result; the label is not a universal accuracy guarantee.
Check
A source or validation concern needs attention, such as repeated matches, weak OCR evidence, handwriting or a differing independent reading. Inspect the explanation and original page.
Place not found
A value was extracted, but its source location could not be verified. Review the original manually before accepting it.
Not on the page
No value was extracted, or the reading was explicitly rejected. The field remains empty; a required missing value also raises a warning.
Doesn’t match
Printed figures or extracted relationships do not reconcile. Review the compared amounts; Tracewise does not silently fix a document’s arithmetic.
Corrected
A person changed the value. The first reading remains available beside the correction.
Confirmed
A person inspected and accepted the value. Document confirmation is still required before an approved snapshot becomes exportable.
06 / WORK THAT SURVIVES A CLOSED TAB
Upload accepted.
You can step away.
A durable server queue and Azure scheduler advance OCR, field reading and review preparation. The workspace observes saved progress; it does not keep the job alive.
Wait for the upload-success message before closing. Uploading bytes still requires your connection. Return later to review the saved results and confirm them.
Saved checkpoints and bounded retries.
Completed stages are checkpointed. Temporary provider failures retry with backoff, and expired job leases permit recovery after an interrupted worker. Persistent failures remain visible rather than returning sample results.
Allowances remain explicit.
Server-side daily and monthly allowances limit processing. Up to three rereads per file per day are included within its paid page range; a wider range charges only added pages. Failed attempts refund their uncommitted reservations once.
Timing depends on the work.
The scheduler checks at intervals and advances bounded steps. Queue volume, document size and provider response time affect completion. The workspace warns when the scheduler heartbeat is stale.
8Synthetic invoice results across five single-file tests and one three-page PDF
3Split pages saved in Review while the workspace was closed
3Export formats downloaded and compared after human confirmation
Recorded 6 November 2026. These runs tested a limited invoice workflow, not general model accuracy, customer outcomes or every supported document type. They predate the independent second-reading feature.
08 / BEFORE YOU RELY ON A READING
Useful boundaries.
Can instructions inside a document control Tracewise?
Document content is treated as untrusted evidence. The extraction model has no action tools, and the application validates its output. A live synthetic instruction-injection test retained the expected identifier and amount without approving the file. That test is a limited observation, not a guarantee against every attack.
Does it automatically fix unclear scans?
There is no separate blur, darkness or cropping classifier. OCR may report orientation and source quality, but unreadable content still requires a clearer upload or human review. A second reading uses the same OCR evidence and cannot recover information that is absent from it.
Does a correction train the system?
No automatic learning from corrections is implemented. Improve your reusable field definitions or reference data deliberately, then request another reading when appropriate.
Can I skip review because two readings agree?
No. Agreement is another recorded signal. Existing warnings remain, and a person must confirm the document before it can be exported.
Understand the source.
Own the decision.
Follow one measured synthetic invoice from its reading to a confirmed export, then try the workflow on a document you are authorised to process.
Tracewise