GES-C02 · Domain 4 of 4

GES-C02 Domain 4: Snowflake Document Processing

Domain 4 covers document processing with Cortex: AI_PARSE_DOCUMENT in OCR and LAYOUT modes, AI_EXTRACT with response formats, staging and preparing files, automated pipelines with streams and tasks, and troubleshooting errors, privileges, cost and arctic-extract fine-tuning.

Exam weight estimate
15%
Official objectives estimate
4
Estimated questions estimate
~8
Suggested study time estimate
5–8 hrs

About 11 min read Reviewed October 2026

Overview

Snowflake Document Processing is 15% of GES-C02, about 8 questions, and replaces the Document AI content of GES-C01 with function-based processing. It has four objectives. Objective 4.1 covers the parsing functions: AI_PARSE_DOCUMENT with OCR and LAYOUT modes and options such as page_split, and AI_EXTRACT with its response format and prompt design.

Objective 4.2 covers preparing documents: uploading files to a stage, meeting supported format and size limits, and setting up the stage correctly. Objective 4.3 covers automated pipelines that detect new files and process them with streams and tasks, often combined with other Cortex functions. Objective 4.4 covers troubleshooting and optimization: reading query errors, using GET_PRESIGNED_URL to inspect files, meeting privilege requirements, controlling cost and fine-tuning arctic-extract models for better extraction accuracy.

How it's tested

Domain 4 questions follow a document from upload to table. They ask which function or mode to use, what the stage needs, how to automate processing of new files, or why a call failed. Read for the document type, whether layout such as tables matters, and whether the goal is full text or specific fields.

  • Parse or extract

    A document need is described and you choose AI_PARSE_DOCUMENT for full content or AI_EXTRACT for specific fields.

  • Automate new files

    Documents arrive continuously and you design a stream on a directory table with a task that calls the functions.

  • Fix the failure

    A document call errors or returns poor results and you identify a stage, format, privilege or model cause.

Reasoning it through

A finance team needs invoice number, vendor and total from thousands of scanned invoices. They want fields, not full text, so AI_EXTRACT with a response format naming those three fields fits. AI_PARSE_DOCUMENT would return all the content, leaving more parsing work. If accuracy on an unusual layout is weak, fine-tuning an arctic-extract model on labeled examples is the documented path.

Objectives 4

Official objectives from the Snowflake outline, in outline order.

4.1 Use document parsing functions.AI_PARSE_DOCUMENT modes and options, and AI_EXTRACT response formats.
What it means
This objective covers AI_PARSE_DOCUMENT, which returns document content in OCR mode (text) or LAYOUT mode (text with structure such as tables and headings), with options including page_split and page limits, and AI_EXTRACT, which pulls named fields from documents or text using a response format and well-written prompts.
What to understand
Choose OCR mode for plain text such as scanned letters, and LAYOUT mode when tables and reading order matter, for example before chunking for RAG. page_split returns content per page, which helps with large files and citations. AI_EXTRACT takes a response format that can be a list of questions, an object of field names to questions, or a JSON schema, and returns JSON. Clear, specific field questions improve accuracy. Files are passed with TO_FILE referencing a stage path.
How it shows up
A document and desired output are described and you choose the function, mode or option.
Common confusion
Using OCR mode when the task depends on preserving table structure.
4.2 Prepare and manage documents and implement extracting workflows.Uploading documents and meeting format and size requirements.
What it means
This objective covers getting documents ready for processing: uploading files to an internal or external stage, meeting supported file types and size or page limits, and implementing a basic extraction workflow from staged files into tables.
What to understand
Internal stages used with Cortex document functions should use server-side encryption, and a directory table lists staged files so you can query them and process each one. Check that file types such as PDF, images, Word or PowerPoint are supported and that files stay under documented size and page limits; split oversized documents. Store extraction results in a table with the file path so results can be traced back to the source.
How it shows up
A workflow fails or needs setup for new documents and you choose the stage, format or preparation step.
Common confusion
Overlooking stage encryption or directory table setup when calls fail on staged files.
4.3 Build automated document processing pipelines with Cortex AI integration.Streams and tasks for continuous processing.
What it means
This objective covers automating document processing: detecting newly uploaded files with a stream on a stage's directory table, then running a task that calls AI_PARSE_DOCUMENT or AI_EXTRACT and other Cortex functions and writes results to tables.
What to understand
Enable a directory table on the stage and refresh it automatically or on a schedule, create a stream on it to capture new files, and create a task that runs when the stream has data, processing only new rows. Chain further tasks to classify, summarize or chunk the parsed text and load it into Cortex Search. This processes each document once and keeps costs predictable.
How it shows up
Files arrive continuously and you choose the objects and order to process each new file automatically.
Common confusion
Scheduling a task to reprocess every file in the stage instead of consuming a stream of new files.
4.4 Troubleshoot and optimize document processing.Errors, presigned URLs, privileges, cost and fine-tuning.
What it means
This objective covers diagnosing document processing problems and improving results: interpreting query errors, using GET_PRESIGNED_URL to view a source file, meeting requirements and privileges, applying cost and best-practice guidance, and fine-tuning arctic-extract models for better extraction.
What to understand
Common failures come from unsupported formats, oversized files, missing stage privileges, missing encryption or missing Cortex roles. GET_PRESIGNED_URL produces a temporary link so a reviewer or app can open the original file. Cost is token-based, often by page, so limit pages, avoid reprocessing and pick the lighter mode when structure is not needed. When AI_EXTRACT is inaccurate on a document type, fine-tune an arctic-extract model with labeled examples.
How it shows up
A document call errors, returns weak fields or costs too much, and you choose the fix.
Common confusion
Rewriting prompts endlessly when the fix is fine-tuning, or blaming the model for a privilege error.

High-yield concepts

The terms, rules and formulas this domain turns into questions most often.

  • AI_PARSE_DOCUMENT

    Returns full content from documents.

    Remember
    OCR or LAYOUT mode.
    Tested as
    Parsing questions.
    Trap
    Using it when only fields are needed.

  • OCR mode

    Text extraction without structure.

    Remember
    Lighter choice for plain text.
    Tested as
    Mode selection.
    Trap
    Using it for table-heavy files.

  • LAYOUT mode

    Text with tables and structure.

    Remember
    Good input for RAG chunking.
    Tested as
    Mode selection.
    Trap
    Choosing it for simple letters.

  • page_split

    Returns content per page.

    Remember
    Helps with large files and citations.
    Tested as
    Option questions.
    Trap
    Confusing it with a page limit.

  • AI_EXTRACT

    Pulls named fields into JSON.

    Remember
    Uses a response format.
    Tested as
    Extraction questions.
    Trap
    Writing vague field questions.

  • TO_FILE

    References a staged file in a function.

    Remember
    Takes stage and path.
    Tested as
    Function call questions.
    Trap
    Passing a raw URL instead.

  • Directory table

    Queryable list of staged files.

    Remember
    Base for streams on new files.
    Tested as
    Pipeline questions.
    Trap
    Forgetting to refresh it.

  • Stream and task

    Capture new files and process them.

    Remember
    Processes each file once.
    Tested as
    Automation questions.
    Trap
    Reprocessing the whole stage.

  • Server-side encryption

    Stage encryption setting.

    Remember
    Needed for internal stages with these functions.
    Tested as
    Troubleshooting questions.
    Trap
    Using client-side encryption.

  • GET_PRESIGNED_URL

    Temporary link to a staged file.

    Remember
    Lets users view the source.
    Tested as
    Review workflow questions.
    Trap
    Expecting a permanent link.

  • arctic-extract fine-tuning

    Improves extraction on specific layouts.

    Remember
    Uses labeled examples.
    Tested as
    Accuracy questions.
    Trap
    Only tweaking prompts.

Side by side

Concepts candidates mix up, compared on the point that decides the answer.

AI_PARSE_DOCUMENT vs. AI_EXTRACT

ConceptMeaningKey differenceCommon trap
AI_PARSE_DOCUMENT Full document content Returns text or layout Using it for a few fields
AI_EXTRACT Specific fields as JSON Driven by a response format Using it to get all text

OCR vs. LAYOUT mode

ConceptMeaningKey differenceCommon trap
OCR Plain text output Lighter, ignores structure Losing tables
LAYOUT Structured output with tables Keeps reading order and tables Paying for structure you do not need

Ways to process new documents

ConceptMeaningKey differenceCommon trap
Stream on directory table + task Event-like incremental processing Each file processed once Forgetting directory refresh
Scheduled full reprocess Task over all staged files Repeats work and cost Choosing it for continuous arrival
Manual query Ad hoc processing Fine for one-off tests Using it in production

Exam traps

Each trap is an answer that looks right. Learn the clue that should change your mind.

  1. Old Document AI build

    What candidates pick
    Train a Document AI model in the old interface.
    Why it looks right
    GES-C01 material taught it.
    Why it's wrong
    GES-C02 focuses on AI_EXTRACT and AI_PARSE_DOCUMENT; accuracy work uses arctic-extract fine-tuning.

    Clue to watch for Exam answers mention AI_EXTRACT and arctic-extract.

  2. OCR for tables

    What candidates pick
    Parse a financial statement in OCR mode.
    Why it looks right
    OCR sounds like the standard scan option.
    Why it's wrong
    OCR mode returns text without structure; LAYOUT mode preserves tables and reading order.

    Clue to watch for Tables or columns must stay intact.

  3. Parse then regex

    What candidates pick
    Parse whole documents and pull fields with string functions.
    Why it looks right
    It seems flexible.
    Why it's wrong
    AI_EXTRACT returns named fields directly as JSON and handles varied layouts better than pattern matching.

    Clue to watch for Only a handful of fields are required.

  4. Full reprocess task

    What candidates pick
    Schedule a task to parse every file nightly.
    Why it looks right
    Tasks are the automation tool.
    Why it's wrong
    Without a stream, already processed files are parsed again, wasting tokens; a stream on the directory table captures only new files.

    Clue to watch for Files arrive continuously.

  5. Model blamed for access

    What candidates pick
    Switch models after a stage access error.
    Why it looks right
    Errors look like model failures.
    Why it's wrong
    The role needs privileges on the stage and the Cortex database role; changing models will not fix access.

    Clue to watch for The error mentions privileges or the stage.

  6. Permanent link

    What candidates pick
    Store GET_PRESIGNED_URL output for long-term access.
    Why it looks right
    It returns a URL.
    Why it's wrong
    Presigned URLs expire, so generate them when needed rather than storing them as permanent links.

    Clue to watch for Users open the file later.

Answer framework

Use these steps to answer Domain 4 document processing scenarios.

  1. Decide whether the goal is full content for search or analysis, which suggests AI_PARSE_DOCUMENT, or specific fields for tables, which suggests AI_EXTRACT.
  2. If parsing, pick OCR mode for plain text and LAYOUT mode when tables, headings or reading order must survive, and use page_split for large files.
  3. Check preparation: supported format, size and page limits, a stage with server-side encryption, a directory table, and the right privileges.
  4. For ongoing intake, create a stream on the directory table and a task that processes only new files and writes results with the file path.
  5. When results are wrong or costly, inspect files with GET_PRESIGNED_URL, read the error, limit pages, and fine-tune arctic-extract for weak extraction.

Practice question

Supplier invoices are uploaded to an internal stage throughout the day. The team needs the invoice number, date and total stored in a table for each new file, without reprocessing older files. Which design is BEST?

Want more like this? The GES-C02 practice test has questions for every domain.

Checklist

0 of 11

Tick each one you can do without notes. Your ticks stay in this browser.

Study strategy

At 15% this is the smallest domain, but it is concrete, so a few focused sessions can secure most of its questions. Start in a trial account: create an internal stage with server-side encryption and a directory table, upload a few PDFs including one with tables, and run AI_PARSE_DOCUMENT in both OCR and LAYOUT modes to compare output. Try page_split on a longer file.

Then run AI_EXTRACT on invoices or forms with each response format style: a list of questions, an object of field names and a JSON schema. Note how question wording changes accuracy, and read the documentation on fine-tuning arctic-extract so you know when it is the right step.

Next, automate it: create a stream on the directory table and a task that extracts fields from new files into a table, then upload another file and confirm only it is processed. Generate a presigned URL for a result row.

Finish by reviewing limits, privileges and cost guidance, and practice troubleshooting questions. Ignore older Document AI material, which GES-C02 no longer covers.

FAQ

How much of the SnowPro Gen AI exam is Snowflake Document Processing?

Snowflake Document Processing is 15% of GES-C02, the smallest domain, or about 8 of 55 questions. It has four objectives on parsing functions, document preparation, automated pipelines and troubleshooting.

Is Document AI still on the exam?

GES-C02 covers document processing through AI_PARSE_DOCUMENT and AI_EXTRACT rather than the older Document AI model-building interface tested in GES-C01.

When should I use LAYOUT mode?

Use LAYOUT when tables, headings or reading order must be kept, such as financial statements or content you will chunk for RAG. Use OCR for plain text where structure does not matter.

How do I tell AI_EXTRACT what to return?

Pass a response format, either a list of questions, an object mapping field names to questions, or a JSON schema. The function returns the answers as JSON.

How do I process new documents automatically?

Enable a directory table on the stage, create a stream on it to capture new files, and run a task that calls the document functions on the stream's rows.

Why might a document function call fail?

Typical causes are an unsupported format, a file over size or page limits, a stage without server-side encryption, or a role missing stage privileges or the Cortex database role.

What is arctic-extract fine-tuning for?

Improving AI_EXTRACT accuracy on document types it handles poorly, by training on labeled examples of your own documents.

Other Artificial Intelligence & Machine Learning certifications

Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.

Test your Domain 4 knowledge

You have covered AI_PARSE_DOCUMENT modes, AI_EXTRACT, stage preparation, streams and tasks, and troubleshooting. See whether you can design a document pipeline under exam timing.

Take the GES-C02 practice test Check your GES-C02 readiness

Keep going