Playground

Not yet checked on the product

This chapter was written from the product's source code and has not yet been checked by a person on a running Cogneris. If what you see on screen differs, the screen is right — please tell your Cogneris contact so the chapter can be corrected.

The Playground is where you try Vision on a real document before you put it in a workflow or an API integration. On one screen you can extract a document against your templates (Template Based), extract one without any template (Zero Shot), find where each document starts in a multi-page PDF (Split), detect and straighten documents in a photo (Crop), and burn a redaction into a copy of a document (Redaction).

Every run is a real call to the Vision service. It is billed and rate-limited just like an API call, and every tool except Template Based extraction also fires your webhooks (see What every run costs and triggers).

How to get here	Vision left menu › Playground, or the Playground card on the Vision Dashboard
Who can use it	The shell grants the screen to the roles cogneris-admin, cogneris-tech, customer-admin and customer-tech. The service only checks that you are signed in. Redaction has its own conditions (see Redaction).
Content status	From source; not observed (vision-fe 0b724a0 + cogneris-vision-fe#192, vision-be fce7fcc + cogneris-vision-be#504, 2026-10-22)

The start screen

The screen is titled Choose Extraction Mode, with the subtitle Upload a document to start extraction. It shows two groups of tiles. Clicking a tile opens your computer's file picker. The tool starts as soon as you pick a file; there is no separate "run" button.

Group	Tile	Tile text	Accepted files	Section
Extraction mode	Template Based	Classify the document first, then extract fields using a trained template · Click to upload	Any image, or PDF	Template Based
Extraction mode	Zero Shot	Auto-detect the document type and extract all key fields in a single pass · Click to upload	Any image, or PDF	Zero Shot
Other tools	Split	Identify each distinct document in a multi-page PDF and show where it begins · Click to upload a PDF	PDF only	Split
Other tools	Crop	Detect documents in an image and crop each detected region. · CLICK TO UPLOAD JPG, PNG OR PDF	JPG, PNG, PDF	Crop
Other tools	Redaction	Permanently hide sensitive content in a rasterized copy. · see the table below	PDF, JPEG, PNG, WEBP	Redaction

Under the tiles: Or drag and drop a file on either tile. Dropping a file on Template Based or Zero Shot picks that mode and runs it, the same as clicking the tile and choosing the file. Crop also accepts a drop; Split does not.

The Redaction tile appears only when the service reports that redaction is available, or when it could not find out. Its last line changes with its state:

Tile text	Meaning	Clicking it
Open Redaction	At least one template has a published redaction policy you can use, or you hold cogneris-admin or cogneris-platform-admin	Opens the policy step
Requires a published Redaction policy	Redaction is on, but no template has a published policy yet	Nothing. The tile is disabled.
Could not check availability. Try again	The availability request got no answer	Asks the service again

Each Playground run handles one file. If you pick several, only the first is used.

Common problems on the start screen

Message	Cause	What to do
Unsupported file format. Please upload an image or PDF.	The file picked for Template Based or Zero Shot is not an image or a PDF	Convert the file to PDF, JPG or PNG
Split only accepts PDF files.	A non-PDF was picked for Split	Use a PDF
Crop only accepts JPG, PNG or PDF files.	Another format was picked for Crop	Use JPG, PNG or PDF
Unable to process the PDF. Please try again.	The Split call failed (any service error, including a rejected file or an exhausted quota)	Try again. If it keeps failing, check the file and your plan.
Unable to process the file. Please try again.	The Crop call failed	As above
The model did not detect any documents.	Crop returned no document and no image	Use a sharper photo with the whole document in frame

The service enforces its own file checks on every tool: the content must match the extension, and the file can be at most 50 MB (Redaction: 100 MB, checked in the browser). A file that fails these checks shows the error message of the tool you used.

Template Based extraction

Two calls run in sequence:

1. Classify. The file is sent to the classifier, which returns a document type and a confidence. The service then looks for one of your templates whose name matches that type. It first tries an exact match (case-insensitive), then a normalized match.
2. Extract. If a template matched, the file is extracted against that template's schema. Because the template id is sent, the service builds the prompt from the template's fields and asks the model to return only the aliases the template declares (see Fields and schema).

Once the classifier answers and finds no template, the Playground stops. It does not extract with an empty schema; it shows the No Template state instead (see No template for this document type).

Template Based is the only mode that can show the Document trust check panel (see Document trust check).

Zero Shot extraction

A single call classifies the document and extracts its key fields at the same time, without a template. The result has the same layout as Template Based. The mode chip reads Zero Shot, and no Document trust check panel is shown: the service never runs the fraud side-car for zero-shot.

Use it to see what Vision reads on a document type you have not modeled yet. Then create a template and switch to Template Based to get stable field names.

The extraction result screen

Template Based and Zero Shot share one result screen. The document preview is on the left, the extracted data on the right.

While it runs

Element	What it shows
Status badge	Identifying Document... while classifying, then Extracting Data... (Template Based) or Extracting (Zero Shot)..., with animated dots and a running timer in seconds
Progress line	A message that changes every 5 seconds. Template Based: still processing, identifying document type, getting document information, almost there, just a little more. Zero Shot: analyzing document structure, identifying key fields automatically, extracting data without template, almost there, just a little more
Step list	Uploaded › Classified document type › Extracting fields… › Trust & fraud check, each marked done, current or pending
Data panel	Six placeholder rows

When it finishes

Element	What it shows
Status badge	Document Processed Successfully and the total time, Process Error on failure, or No Template when no template matched
New Upload (round upload button)	Clears everything and returns to the start screen
Type chip	The detected document type and its confidence, e.g. "Invoice · 92.50%". The confidence has two decimals and is colored: green from 90, amber from 50 to 89.99, red below 50, neutral when the service sent none.
Mode chip	Template-based or Zero Shot
Document trust check	Template Based only, when the fraud check ran (see below)
Form / JSON	Switches the data panel between field cards and raw JSON. Hidden while running, on error, and on No Template.

Preview toolbar

Button	What happens
+ (Zoom In)	Enlarges by 20 % per click
− (Zoom Out)	Shrinks by 20 % per click, down to about 40 %
⤢ (Reset)	Back to 100 % and no rotation
⟳ (Rotate)	Rotates 90° clockwise

PDFs are drawn page 1 only, on a canvas, at the zoom you choose. Images are scaled and rotated in place.

Form view: the extracted fields

The Extracted Data panel lists what the model returned:

Kind of value	How it is shown
Single value	A card with the field label, the value and, when the model returned one, the confidence as a whole number followed by % with a colored dot. The card's accent uses the same bands as the type chip: 90 and above high, 50 and above medium, below 50 low.
Group of values (an object)	Each inner key becomes a card of its own
List (an array, e.g. invoice lines)	A block titled with the field name, one row per item, each cell with its label and value. A cell that is itself a list becomes a nested table.

Click a card to select it. On an image, the area the model read the value from (its bounding box) is outlined on the preview. Click the card again to clear the outline. On a PDF no outline is drawn, even when the model returned a box.

The confidence is what the model reports about itself. It is not a measured accuracy. If nothing was extracted, the panel reads No data extracted.

JSON view

Shows the extracted values as indented JSON, with the confidence, label, page and box wrappers removed: you see the values only, in the shape your API integration will receive them.

No template for this document type

When the classifier names a type that none of your templates matches, the badge shows No Template and the panel shows No Template for This Document Type with one of two texts:

• Document type detected as …, but no predefined template exists for this type. Use the Template creation module and then try again.
• This document couldn't be matched to any of your templates. Use the Template creation module to create one, then try again. (when the type is empty or "UNKNOWN")

Create Template opens the New Template screen with the document you tested already attached and, when the classifier recognised a type, that type as the title.

Common problems with extraction

Message	Cause	What to do
Failed to run classification process. Please try again.	The classifier call failed (any error from the service, including a rejected file, a rate limit or an exhausted plan)	Try again later. If it persists, check your plan usage.
Unable to identify the document type.	The classifier answered with no result	Try a clearer scan, or use Zero Shot
Unable to extract data. The document was classified but extraction failed.	The extract call failed after a successful classification. This also covers the model returning truncated output, and a document the input-quality gate rejected as too poor to read.	Try again with a sharper, complete document
Unable to extract data. Zero-shot extraction failed.	The zero-shot call failed	Try again

Document trust check

An advisory panel that shows the fraud assessment of a Template Based extraction. It never blocks or hides the extracted data in the Playground.

When it appears

The service runs the fraud check (the fraud-detection agent) only when all of these hold. Otherwise no panel is shown at all:

1. Fraud-on-extract is enabled for the deployment.
2. The extraction was pinned to a template (Template Based, with a match).
3. Your tenant's agent guards allow the fraud-detection agent, agents are not paused, and fraud enforcement is set to Advisory (record only) (see Tenant agents).
4. The template's agent configuration enables the fraud-detection agent (see Agent).

The fraud check makes one extra model call, which estimates how likely the document is to be AI-generated. Its cost is added to the run.

What the panel shows

Element	What it shows
Document trust check	The panel title
Badge	Trusted, Review recommended, Suspicious, or Verification unavailable
Risk score	The combined score, 0–100 %
AI-generated likelihood	The model's estimate, 0–100 %, that the document was produced by AI
Policy action (advisory only)	allow, flag or block: what enforcement would have done. In the Playground the check only runs in advisory mode, so this always reads allow.
Analysis ID	The assessment id. Quote it to support, or find the run in the review queue.
Recommendation	Trusted: Continue with the extracted data. · Review: Human review is recommended before relying on the result. · Suspicious: Pause the decision and perform human review; the extraction remains available.
Coverage is incomplete. These checks could not complete:	Names of the checks that failed, plus download: followed by an id for any document the pipeline could not read. Only shown when something failed.
Evaluated checks	One row per check: its name in words with the technical name beside it, and Clear, Flagged or Failed, followed by the signals it raised
Signal	Severity (Low, Medium, High), the signal code, Signal score as a percentage, and the evidence text

When no assessment exists, or the request fails, the badge reads Verification unavailable and the panel says The fraud assessment is unavailable. Extracted data remains available. along with the Analysis ID. The panel never shows Trusted for an assessment it could not read.

For an assessment recorded before the checks were listed, the panel shows Coverage is unknown because this assessment predates the recorded check roster. and lists the signals without grouping them by check.

The check names, signal codes and evidence texts come from the service as they are, in English and not translated.

How the risk level is decided

The service combines three terms and clamps the result to 0–1:

• 0.6 × the highest score among the signals below
• 0.3 × the AI-generated likelihood
• 0.4 × the prompt-injection magnitude (0.7 when an injection phrase is found in the extracted text, otherwise 0)

At 0.35 or more the level is Review recommended; at 0.7 or more it is Suspicious; below 0.35 it is Trusted. These are the global defaults. Your tenant's fraud scoring settings and the template's sensitivity can change them. A broken or unverifiable PDF signature never yields Trusted: the level is raised to at least Review recommended, while the score itself is left unchanged.

The checks and their signals

The checks run in this order. A score of 0 % means the signal is informational and does not raise the risk.

Check	Signal code	Severity · score	What it means
semantic	invalid_tax_id	High · 90 %	A field whose name contains cpf or cnpj holds a number that fails the check-digit validation
semantic	impossible_date	Medium · 60 %	A date field (name contains date, data, emiss or vencim) holds a date in the future
semantic	value_inconsistency	Low · 40 %	An amount field (name contains amount, total, valor or price) holds something that is not a number
pdf_structure	pdf_incremental_updates	Medium · 50 %	The PDF has more than one end-of-file marker: it was edited and re-saved after creation
pdf_structure	pdf_date_mismatch	High · 80 %	The PDF's modification date is earlier than its creation date
pdf_structure	pdf_producer_anomaly	Medium · 50 %	The PDF's creator and producer differ (e.g. made in one tool, re-saved in another)
pdf_structure	pdf_missing_metadata	Low · 30 %	The PDF has neither creator nor producer metadata
image_ela	ela_inconsistency	Medium · 55 %	Error-level analysis: after a JPEG re-compression, more than 20 % of pixels changed noticeably, which suggests pasted or edited regions. At most one per document.
perceptual_hash	duplicate_document	High · 80 %	The exact same file (same SHA-256) was already submitted in your tenant
perceptual_hash	layout_reuse	Medium · 50 %	A visually near-identical document was already submitted in your tenant
perceptual_hash	repeat_reuse	Medium or High · 60 % up to 95 %	The same visual fingerprint has 3 or more earlier submissions in your tenant. The score rises 10 % per extra submission, and the signal is High from 70 %.
c2pa_provenance	c2pa_present_invalid	High · 70 %	The file carries a content-provenance (C2PA) manifest that is broken
c2pa_provenance	c2pa_present_valid, c2pa_present_untrusted, c2pa_absent	Low · 0 %	A valid manifest, a manifest from an issuer you have not trusted, or no manifest. Informational only.
c2pa_provenance	c2pa_unavailable	Low · 0 %	Provenance was not evaluated for this document
pdf_signature	pdf_signature_broken	High · 70 %	The PDF's digital signature no longer matches its content. Forces at least Review recommended.
pdf_signature	pdf_signature_unverifiable	High · 70 %	The signature could not be checked. Forces at least Review recommended.
pdf_signature	pdf_signature_untrusted	Medium · 40 %	Signed, but by a certificate that is not trusted
pdf_signature	pdf_signature_valid, pdf_signature_absent	Low · 0 %	Validly signed, or not signed. Informational; a valid signature does not lower the risk.
pdf_signature	pdf_signature_unavailable	Low · 0 %	The signature was not evaluated for this document
prompt_injection	prompt_injection_attempt	High · 0 %	The extracted text contains an instruction aimed at the AI (e.g. "ignore previous instructions", "system prompt", "approve this"). It is listed at 0 %, but it raises the risk through the prompt-injection term above.
history_velocity	velocity_anomaly	Medium · 60 %	Your tenant submitted more than 20 documents for fraud assessment so far today (default window and limit)

What to do with the result: on Review recommended or Suspicious, have a person look at the original document before relying on the data. Confirmed cases and false positives are recorded in the review queue (see Tenant agents).

Split

For a multi-page PDF that contains several documents (an ID, a proof of address, a payslip…). The service identifies each distinct document and the page where it begins. It does not cut the PDF into files.

Input: one PDF. The browser rejects anything else, and the service rejects it again by extension and content.

While it runs: Processing document... with the file name, a timer, and a step that advances every 2.5 seconds: Uploading document…, Analyzing pages with AI…, Detecting document types…, Building the document index…, Opening the original PDF…. These steps are timed in the browser and do not report the service's real progress. Cancel returns to the start screen; the request already sent to the service is not withdrawn.

Result:

Element	What it does
Document index and file name	Header
New Upload	Back to the start screen
Document index table: Document type, Start page	One row per document found. Click a row to jump to its first page. The row of the document you are viewing is highlighted (the last document whose start page is at or before the current page).
No distinct documents were identified. You can still review the original PDF.	The service found nothing usable. If any entry has no type or an invalid start page, the whole index is discarded.
‹ / › and the page box, of …	Previous and next page, or type a page number and press Enter
− / + and the percentage	Zoom from 50 % to 200 % in steps of 25 %
The original PDF could not be displayed. Try a different file.	The browser could not render the PDF

Crop

Detects the documents in a photo or scan, corrects their perspective, and shows the straightened image. This is useful for a phone photo of an ID card on a table.

Input: JPG, PNG or PDF. A PDF of exactly two pages (typically front and back) is merged in the browser into one PNG image, page 1 above page 2, before it is sent. Any other PDF is sent as it is.

While it runs: Processing document..., the file name, a timer and timed steps: Uploading file…, Detecting documents with AI…, Computing bounding boxes…, Cropping detected regions…, Preparing crop result…. Cancel works as in Split.

Result: two panels.

Panel	What it shows
Original	The file you sent. On an image (including a merged two-page PDF), each detected document is outlined. A PDF that was not merged is shown without outlines.
Image Cropped	The composite the service returns when there is one; otherwise every detected document's image, one under the other. If there is none: Select a row to preview the cropped image.
New Upload	Back to the start screen

When several documents are detected, only one image is shown here. The detection table (type, coordinates, angle, confidence) that the translation catalog still describes is not shown in this version.

Redaction

Produces a rasterized copy of a document in which sensitive regions are painted over in the pixels, so they cannot be recovered by copying text. The categories, style and retention come from a redaction policy published on a template (see Redaction policy).

When you can use it

• The service must report redaction as available for your tenant.
• At least one template must have a published policy. Otherwise the tile reads Requires a published Redaction policy.
• Platform administrators (cogneris-admin, cogneris-platform-admin) can open the tool regardless, but can only submit a file when a policy is available.

Step 1 — Redaction policy

Field	What it does	Required	Values	Validation messages
Policy template	The template whose published policy is applied	Yes	Placeholder Choose a template with a published policy, then every template with a published policy, by name. The list loads all pages automatically, up to 20 pages.	Unable to load policies · No template has a published redaction policy yet. Publish one from a template's Redaction tab. · Redaction is unavailable for this template.

Button	What happens
Back	Back to the start screen
Select document	Opens the file picker; enabled once a valid policy is selected. The upload starts as soon as you pick a file.

File checks in the browser: PDF, JPEG, PNG or WEBP, at most 100 MB. The service processes at most 100 pages.

Step 2 — Redacting document

Redaction runs as a background job. The screen polls its status (every 1 s at first, slowing to every 5 s) and shows the current stage: Queued, Rasterizing pages, Detecting sensitive regions, Consolidating regions, Burning redactions into pixels, Verifying redacted output, Storing temporary artifacts. It also shows pages done out of the total, once known.

Cancel asks the service to stop the job. If the service does not confirm: The cancellation could not be confirmed. Please try again.

Step 3 — the result

Before anything is shown, the browser checks both files against the job's manifest (fingerprints, page counts, categories) while showing Verifying downloaded artifacts. If anything does not match, the result is withheld: The result could not be verified and is unavailable.

Element	What it shows or does
Redacted document ready	Header, with New Upload
Original / Redacted result	Side by side, on the same page and zoom. For PDFs: ‹ / ›, Page … of …, zoom 50–200 %.
Safe redaction summary	Protected categories (each category with its count), Affected pages (or None), Low-confidence regions were redacted and require review (with the count, only when there are any), Policy version, Duration (… seconds), Available until (local date and time), Output hash (the first 12 characters)
Download redacted document	Saves redacted- followed by the first 8 characters of the job id, as .pdf, .jpg or .png
Download safe manifest	Saves redaction-manifest- followed by the same 8 characters, as .json. The manifest lists regions, categories and fingerprints, not the hidden content.

Possible categories: Full name, Tax ID, Document number, Date of birth, Address, Phone, Email, Bank account, Payment card, Face, Signature, QR or barcode, Passport.

The result is temporary. When Available until passes, the screen drops it and shows The temporary result has expired.

Copy to ECM

Shown only when the service allows copying for your tenant.

Field	What it does	Required	Format	Validation
Destination folder ID	The ECM folder that receives the copy	Yes	A folder UUID (not the all-zero one)	Copy to ECM stays disabled until it is valid
Document name	Name of the new ECM document	Yes	1 to 200 characters, no control characters	As above
I confirm that this redacted artifact should be copied to ECM.	Explicit confirmation	Yes	Checkbox	As above

Copy to ECM (shows Copying to ECM while it runs) copies the redacted file only, never the original. On success: Copied as document …, version …., and the form locks. The copy is accepted only if the service returns valid ids and a fingerprint equal to the verified output.

Common problems with redaction

Message	Cause	What to do
Redaction only accepts PDF, JPEG, PNG or WEBP files.	Wrong file type	Convert the file
This file is larger than the 100 MB maximum.	File too large	Split or compress it
Redaction is unavailable for this template.	The policy could not be read, or the feature is off	Pick another template, or ask an administrator
Sensitive data was still visible after redaction, so the result was withheld. Widen the categories on this template and run it again.	The final check still found sensitive data	Add categories to the policy and run again
The redacted document did not pass verification, so it was not released.	Output verification failed	Run again; report it if it repeats
Sensitive-data detection could not finish on this document.	Detector failure	Run again later
The redacted document could not be stored securely.	Storage failure	Run again later
The pages of this document could not be rendered for redaction.	The pages could not be rasterized	Try another export of the file
This file could not be read as a redactable document.	Corrupt or unsupported content	Check the file
The redaction policy on this template is not valid. Ask an administrator to review it.	Invalid policy	Fix the policy (see Redaction policy)
This redaction stopped because the job was changed somewhere else.	The job was changed elsewhere	Start again
Redaction is temporarily unavailable. Please try again in a few minutes.	Any other service-side failure	Wait and retry
This redaction expired before it finished.	The job expired	Start again
Failed	A request failed without a specific reason	Retry
The document preview could not be rendered.	The browser could not draw the files	The downloads may still work. Otherwise run again.
The authenticated download could not be completed.	Download failed or returned an unexpected file	Retry
The ECM copy could not be completed safely.	The copy failed or returned an unexpected answer	Check the folder ID and your ECM access, then retry

On the failure card, the title shows the final state (Completed, Failed, Cancelled or Expired), the reason, and New Upload.

The old Chat address

/playground/chat used to hold an in-app chat. It now redirects to the MCP screen, which connects an editor to the Cogneris MCP server (see MCP). Old bookmarks and links land there.

What every run costs and triggers

Topic	What happens
Billing	Classify, Extract, Zero Shot, Split and Crop are billable calls, charged as usage of your plan. Template Based makes two billable calls (classify, then extract), plus the AI-likelihood model call when the fraud check runs. On a plan that stops at its limit, the service refuses the call once the monthly allowance is used, and the Playground shows that tool's generic failure message. The Playground does not display a cost.
Rate limit	These calls share the service's document-processing quota. When you exceed it, calls are refused and the tool shows its failure message.
Webhooks	The Playground asks the service not to publish webhooks on every call — classify, extract, zero-shot, split and crop — so a try never fires *.processed or *.failed to your configured endpoints (see Developers).
What is stored	The fraud assessment and the fingerprint of the document (used later for duplicate detection) are stored when the fraud check runs. A redaction job's files are kept until Available until.

Worked example: create a template, test it, read the result

1. Create the template. In Templates, create a template. Give it a Title that names the document type the classifier will return, for example Invoice: the Playground finds your template by comparing its name with the detected type. Upload 2 to 5 sample invoices and wait for processing to finish (see Templates).
2. Shape the fields. In the template editor, keep the fields you need, with the aliases your integration expects (see Fields and schema).
3. Test it. Open Playground, click Template Based, and pick an invoice that was not among the samples.
4. Read the chips. Invoice · 96.00% in green with Template-based means the document was classified and matched to your template. If you see No Template, the detected type differs from your template's title. Note the type shown in the message and rename the template to match, or create one with that name.
5. Read the fields. Each card shows a value and the model's confidence. Click a card to see where it was read (on images). Treat amber and red cards (below 90) as candidates for clearer field descriptions in the schema, or for a better sample.
6. Check the JSON. Switch to JSON: these are the keys and values your API call will return.
7. Read the trust check if it appears. Trusted means continue. Review recommended or Suspicious lists the signals that raised the score; for example, pdf_incremental_updates means the PDF was edited after creation.
8. Change the schema, run the same file again, and compare.

Related

• Vision overview
• Templates · Fields and schema · Redaction policy · Agent
• Tenant agents: agent guards, fraud enforcement and scoring, fraud review queue
• Developers: API keys, webhooks and request logs
• MCP
• Workflow phases that run the same tools: Extraction, Classification, Zero Shot, Split, Crops

---

Last reviewed 2026-10-22 · Version 1.0.0 · Owner: product-documentation