Configure single-task and pro-mode Content Understanding pipelines

Standard Content Understanding answers one question about one file: apply this schema, return these fields. Pro mode answers a question that requires reasoning across several files and against reference material — is this invoice consistent with the contract? It is a different shape of pipeline, and it carries one restriction that decides most questions about it: pro mode is currently available for document data only.

Why This Matters

Reference data is the defining feature. Pro mode "enables reasoning over both input content and reference data", where standard mode sees only the input.

Reference documents attach at analyzer creation. They are configured once and applied at analysis time, not passed per request.

The modality restriction eliminates designs. "Currently, pro mode is exclusively available for document data" — so any pro-mode proposal over images, audio, or video is wrong on that basis.

Prerequisites

• The analyzer model, fieldSchema, and the three field methods.
• That extract is documents only, and confidence and grounding come from estimateFieldSourceAndConfidence.
• The API versions: 2025-11-01 GA, 2026-06-01-preview.
• That Content Understanding runs on your own model deployments.

Learning Objectives

By the end of this lesson you will be able to:

1. Distinguish standard and pro mode by what each reasons over.
2. Configure reference data and know when it is supplied.
3. Design a multi-file pro-mode request.
4. Apply the document-only and preview restrictions.
5. Choose the cheaper mode where it suffices.

Building Blocks

Standard mode — the single task. One analyzer, one input, one schema. Content extraction, then field extraction producing extract, classify, or generate values, with confidence scores 0–1 and grounding where enabled. This covers intelligent document processing, captioning, classification, and RAG ingestion — the great majority of work.

Pro mode. "Designed for advanced use cases that require multi-step reasoning and complex decision-making, such as identifying inconsistencies, drawing inferences, and making decisions." It "enables reasoning over both input content and reference data", and offers "multi-step reasoning capabilities, and multi-input document support".

Reference data. "During analyzer creation, you can provide reference documents that add context at analysis time." The guidance: "Reference documents should be concise and focused, prioritizing essential documents and ensuring they're as brief as possible to enhance retention and recall."

The canonical example. "To analyze invoices for consistency with a contractual agreement, you can provide the invoice and related documents (such as a purchase order) as inputs, and provide the contract files as reference data."

The restriction. "Currently, pro mode is exclusively available for document data."

Availability. Pro mode is documented as preview.

Deep Dive

What "multi-step reasoning" actually buys

Standard mode answers questions about a document. Pro mode answers questions across documents and against a standard.

The documented framing is precise: pro mode is for "identifying inconsistencies, drawing inferences, and making decisions". Each names a task standard mode cannot do in one pass.

Inconsistencies require at least two sources — a value here compared to a value there. A standard analyzer extracts an invoice total and a contract rate; something else must compare them, which means writing that comparison logic yourself.

Inferences require combining facts that no single document states. Whether a mortgage applicant supplied everything needed is not a field on any one document; it is a conclusion from a set.

Decisions require applying a standard to evidence — approve, flag, reject — which is what makes pro mode's output actionable rather than descriptive.

The practical read: if you find yourself planning to run several standard analyzers and then write cross-document comparison code, that is the shape pro mode exists to absorb.

Inputs against reference data

The split is the core configuration decision, and the documented example makes it concrete: to check invoices against a contract, supply "the invoice and related documents (such as a purchase order) as inputs", and "the contract files as reference data".

Inputs are the documents being examined — the ones that vary per request, the ones the question is about.

Reference data is the stable material they are judged against — contracts, policies, rate cards, standards. It changes rarely, which is why it belongs at analyzer creation rather than in every request.

Getting the split backwards produces a pipeline that works and costs far more than it should: passing the contract as an input on every request re-sends stable material each time, while attaching the varying invoices as reference data would be wrong outright.

The brevity guidance matters more than it first appears. Reference documents should be "concise and focused, prioritizing essential documents and ensuring they're as brief as possible to enhance retention and recall". A 200-page master agreement attached whole is worse than the handful of clauses that actually govern the check — not merely more expensive, but less effective, because retention and recall degrade with volume.

When standard mode is enough

Pro mode costs more, is preview, and is document-only, so the default should be standard unless the requirement genuinely needs cross-document reasoning.

Standard mode is sufficient whenever the question is answerable from one document at a time: extracting invoice fields, classifying document type, captioning an image, describing a video segment, producing RAG-ready Markdown. Adding pro mode to any of these buys nothing.

Standard mode is also the right choice where the comparison logic is simple and deterministic. If the check is "does the invoice total equal the sum of line items", that is arithmetic — extract the fields with a standard analyzer and compare in code, which is exact, cheap, auditable, and repeatable. Pro mode earns its cost when the comparison requires judgement over unstructured language: whether an invoiced service falls within a contract's scope, whether supporting documentation satisfies a stated requirement.

That mirrors the hybrid principle from orchestration: deterministic checks belong in code; judgement over unstructured text belongs in the model.

Verification, and why it matters more here

A decision needs to be defensible in a way a description does not.

Confidence scores — "reliability estimates from 0 to 1" — and grounding, which "identifies the specific regions in the content where each value was extracted or generated", let a reviewer trace a conclusion to the clause and the line item that produced it. In a standard extraction that is convenient; in a pro-mode decision it is close to essential, because the output is an assertion about a relationship rather than a copied value.

The operational pattern is the same as elsewhere: route high-confidence decisions straight through, and queue low-confidence ones with their grounding attached so a reviewer sees exactly where to look. That is the documented purpose — accuracy "while minimizing the cost of human review".

Worked Examples

Example 1 — invoice against contract. A finance team must confirm each invoice is consistent with the governing contract, with related purchase orders considered.

Pro mode, matching the documented example exactly: the invoice and purchase order as inputs, the contract files as reference data attached at analyzer creation. This is "identifying inconsistencies" over multiple documents — a standard analyzer per file plus hand-written comparison logic is the workaround pro mode exists to replace.

Example 2 — comparing product photographs to an approved set. A brand team wants to reason across submitted photographs against a reference library of approved imagery.

Not available: "pro mode is exclusively available for document data". The route is standard analyzers producing classifications and descriptions for each image, plus your own comparison against the approved set — or a multimodal model at question time for a small number of assets.

Example 3 — arithmetic dressed as reasoning. A team proposes pro mode to check that invoice totals equal the sum of their line items.

Standard mode plus code. The check is deterministic arithmetic, so extract the fields with a standard analyzer and compare in application logic — exact, cheap, auditable, and repeatable. Pro mode is warranted when the comparison needs judgement over unstructured language, such as whether an invoiced service falls within a contract's scope.

Visual Explanations

The two shapes:

Choosing the mode:

Common Mistakes

Proposing pro mode for images, audio, or video. Documents only.

Passing reference material as an input on every request. It attaches at analyzer creation.

Attaching large reference documents whole. Brevity aids retention and recall.

Using pro mode for deterministic arithmetic. Standard plus code is exact and cheaper.

Using pro mode where one document answers the question.

Ignoring that pro mode is preview under a production-only constraint.

Skipping confidence and grounding on a decision output.

Building cross-document comparison logic by hand where pro mode fits.

Practice Exercises

1. State the defining difference between standard and pro mode.
2. When is reference data supplied, and what guidance governs it?
3. Split the canonical invoice example into inputs and reference data.
4. Give two situations where standard mode is the better answer.
5. What single restriction eliminates most pro-mode proposals?

Summary & Concept Map

Standard Content Understanding is the single task: one analyzer, one input, a schema producing extract, classify, or generate values with confidence and grounding — and it covers the large majority of work, including everything non-document. Pro mode adds multi-step reasoning across multiple input documents and against reference data attached at analyzer creation, for "identifying inconsistencies, drawing inferences, and making decisions" — the canonical shape being invoice and purchase order as inputs with contract files as reference data. Keep reference material concise and focused to preserve retention and recall, express the conclusion as a field, and enable confidence and grounding so a decision traces to its evidence. Two restrictions decide most questions: pro mode is documents only and is preview — and where a comparison is deterministic, standard plus code remains exact, cheaper, and auditable.

Sources and freshness

Written against current Microsoft Learn documentation for the AI-103 skills measured (16 April 2026), reviewed 2026-08-20. Microsoft Learn controls every changing product contract — availability, preview status, quotas, limits, regional support, naming, and retirement dates all move independently of this lesson. Where a scenario turns on a specific number or a GA/preview boundary, confirm it against the product's own page before relying on it.