IDP Extract

Extracts everything an extraction project asks for from a document, routes to the profile of the format in front of it and looks each value up in the page, so you know where every value came from and how much it deserves your trust.

It is the normal path of the IDP: one activity. Inside it loads the document if needed, routes to the profile, extracts all fields in a single call to the model, grounds and persists the result to the store — but that is its business, not yours.

Idempotent: doesn’t re-pay

IDP Extract saves its result by reference (defaults to the file name). If that document was already extracted, it reuses the stored result and does not call Azure/LLM again — a retry after a downstream failure costs nothing. It also persists the document + extraction, so you don’t need Documents Upload afterward.

IDP Extract   document   = = doc
              project    = "device-record"
              credential = = credential("anthropic-key")
              → output: res

The project carries the configuration, not the playbook

This activity used to ask for taxonomy, model, mode and endpoint. Now it asks for an extraction project and everything else comes from it. The project — which you edit in the portal — carries:

• the canonical schema (the fields you deliver, the contract),
• N format profiles, one per layout, each with its overrides on top of the canonical,
• a router that picks the profile by the document’s printed format code,
• the config: model, reading mode, endpoint and OCR credential.

So the person who knows which fields matter configures it in the portal and does not touch the playbook. The playbook only says which project and where the document comes from.

Parameters

Parameter	Type	What it is
project	text	The name of an extraction project of this environment. It carries the canonical schema, the format profiles, the model, the reading mode and the OCR endpoint. For CI it also accepts an inline project-resolve JSON.
credential	credential	Zoan Cloud credential holding the model provider’s API key. Use = credential("name").
document	object	A document that IDP Classify or IDP Load Document already read. Pass it and the file is not read again — with a scan that means not paying Azure twice.
path	text	The path, if you do not pass document.
dpi	int	The dpi the boxes come back in. Use the same one you give PDF To Images if you are going to upload the document, or the boxes land in the wrong place.
pages	text	Which pages to read. Empty = all of them.
azureCredential	credential	Zoan Cloud credential holding the Azure Document Intelligence key. Only for scans.
maxRetries	int	How many times to retry a transient provider failure. A backfill is thousands of calls, so the odd 429 or 5xx is a certainty, not a risk.

Give it the document, not the path

If you come from IDP Classify, use document = = doc. The file is read only once.

It routes to the format profile

The same document type arrives in several formats, and each format paints the fields somewhere else. The project has one profile per format; IDP Extract reads the printed code of the document (for example F-GOP-007), picks the matching profile and merges its overrides on top of the canonical. The output always comes out with canonical names — the mapping is identity.

A format that matches no profile is not extracted

If the document matches no profile of the project, the activity fails on purpose. Classify it first with IDP Classify and send whatever you do not recognise to review: extracting a document of an unknown format yields plausible, false data, and that is the expensive part.

It does not classify

It extracts what it is asked for, always. Checking that the document is the right one is your decision, and it happens earlier with IDP Classify.

It is not a whim: an activity that decides on its own not to do its job hides the control flow, and it also prevents routing between several projects — which is exactly what you need when the queue brings mixed documents.

The output

Field	What it is
values	The data, clean: { serial: "ABC-123", accessories: [...] }. It is what the rest of the playbook consumes.
fields	One field per value, with its box, its confidence and why it needs review. It is what the validation station paints.
needsReview	Whether any field fell below its threshold.
reviewFields	Which ones.
documentType	The type the project declares.
format	The profile it routed to (e.g. F-GOP-007). Use it to trace and show which format was used.
source	text or azure-di.
dpi	The dpi of the boxes.
pageCount	Pages read.
values	The data, clean: { serial: "ABC-123", accessories: [...] }.
documentId	The id of the document persisted in the store.
reference	The key it was saved under (idempotency).
usage	The consumption of the document: { model, inputTokens, outputTokens, calls }.

Usage comes in tokens, not dollars

Extraction asks for all fields in one call (no strict grammar, so the document is read once and not re-sent per batch as before). usage carries that call’s tokens. It is in tokens and not dollars on purpose: the price per model changes, so the cost is computed by whoever consumes it (or the measurement harness) with the current rate.

Confidence is not asked of the model

A model that grades itself says “95%” with the same certainty when it is right as when it makes it up. That confidence is not calibrated.

The one here comes from looking the literal up among the document’s words. If it is there, there is confidence and a box. If it is not, the value cannot be verified and the field goes to review. It is a mechanical check, not an opinion.

Related activities

• IDP Classify — which type it is, before extracting
• IDP Ground — grounding on its own, if you assemble the pipeline by hand
• Documents Upload — upload it to review it

Model escalation

If the project configures an escalation model, when more than 10% of the fields land below threshold the activity retries once with that model and keeps the result that leaves fewer fields to review. An extra call worth cents that saves reviewer minutes. Usage of both calls is added up in usage.

It slices a given document

If you pass document (a reading already done) and pages (a range, "4-7"), the activity slices the reading to that range instead of re-reading the file: that is how Take Confirmed Bundles makes extracting a bundle’s segments not pay Azure again.