Document AI — Parse, Extract, and Classify
Snowflake Cortex provides three AI functions for document intelligence: AI_PARSE_DOCUMENT for text extraction, AI_EXTRACT for structured field extraction, and AI_CLASSIFY for document categorization. Together they enable end-to-end document processing without external tools.
Function Overview
| Function | Purpose | Input | Output |
|---|---|---|---|
AI_PARSE_DOCUMENT | OCR and text extraction | File on a stage | Full extracted text |
AI_EXTRACT | Structured field extraction | Text or file reference | JSON with named fields |
AI_CLASSIFY | Categorization / triage | Text | Matched category label |
AI_PARSE_DOCUMENT — OCR and Text Extraction
AI_PARSE_DOCUMENT reads PDFs, images, and scanned documents from a Snowflake stage and returns their textual content.
-- Parse a single document
SELECT AI_PARSE_DOCUMENT(
'@DOCS_STAGE/invoices/invoice_001.pdf',
{'mode': 'LAYOUT'}
) AS parsed_content;
Parse Modes
| Mode | Description |
|---|---|
LAYOUT | Preserves document layout and structure |
OCR | Optical character recognition for scanned/image documents |
Batch Processing from a Stage
-- Parse all PDFs in a stage directory
SELECT
RELATIVE_PATH,
AI_PARSE_DOCUMENT(
BUILD_SCOPED_FILE_URL(@DOCS_STAGE, RELATIVE_PATH),
{'mode': 'LAYOUT'}
):content::STRING AS doc_text
FROM DIRECTORY(@DOCS_STAGE)
WHERE RELATIVE_PATH LIKE '%.pdf';
AI_EXTRACT — Structured Field Extraction
AI_EXTRACT pulls specific fields from unstructured text into a structured JSON object. It works well for invoices, contracts, receipts, and forms.
SELECT AI_EXTRACT(
doc_text,
{
'vendor_name': 'The company that issued the invoice',
'invoice_number': 'The unique invoice identifier',
'total_amount': 'The total dollar amount due',
'due_date': 'The payment due date'
}
) AS extracted_fields
FROM PARSED_DOCUMENTS;
Accessing Extracted Fields
The output is a JSON object. Use standard Snowflake JSON notation to access individual fields.
SELECT
RELATIVE_PATH,
extracted_fields:vendor_name::STRING AS vendor,
extracted_fields:invoice_number::STRING AS invoice_no,
extracted_fields:total_amount::FLOAT AS total,
extracted_fields:due_date::DATE AS due_date
FROM (
SELECT
RELATIVE_PATH,
AI_EXTRACT(doc_text, {
'vendor_name': 'Company that issued the invoice',
'invoice_number': 'Unique invoice ID',
'total_amount': 'Total amount due as a number',
'due_date': 'Payment due date in YYYY-MM-DD'
}) AS extracted_fields
FROM PARSED_INVOICES
);
AI_CLASSIFY — Document Triage and Categorization
AI_CLASSIFY assigns a document to one of a set of categories you define.
SELECT
RELATIVE_PATH,
AI_CLASSIFY(
doc_text,
['Invoice', 'Contract', 'Receipt', 'Purchase Order', 'Other']
) AS doc_category
FROM PARSED_DOCUMENTS;
Accessing Classification Results
SELECT
RELATIVE_PATH,
AI_CLASSIFY(
doc_text,
['Invoice', 'Contract', 'Receipt', 'Purchase Order']
):label::STRING AS category,
AI_CLASSIFY(
doc_text,
['Invoice', 'Contract', 'Receipt', 'Purchase Order']
):score::FLOAT AS confidence
FROM PARSED_DOCUMENTS;
Building a Document Processing Pipeline
Combine all three functions into a pipeline that ingests, classifies, and extracts data from documents.
-- Step 1: Create a target table
CREATE OR REPLACE TABLE PROCESSED_DOCUMENTS (
file_path STRING,
doc_category STRING,
confidence FLOAT,
vendor_name STRING,
invoice_no STRING,
total_amount FLOAT,
raw_text STRING,
processed_at TIMESTAMP_NTZ DEFAULT CURRENT_TIMESTAMP()
);
-- Step 2: Process and insert
INSERT INTO PROCESSED_DOCUMENTS (file_path, doc_category, confidence, vendor_name, invoice_no, total_amount, raw_text)
SELECT
d.RELATIVE_PATH,
cls:label::STRING,
cls:score::FLOAT,
ext:vendor_name::STRING,
ext:invoice_number::STRING,
ext:total_amount::FLOAT,
parsed:content::STRING
FROM DIRECTORY(@DOCS_STAGE) d,
LATERAL (
SELECT AI_PARSE_DOCUMENT(
BUILD_SCOPED_FILE_URL(@DOCS_STAGE, d.RELATIVE_PATH),
{'mode': 'LAYOUT'}
) AS parsed
),
LATERAL (
SELECT AI_CLASSIFY(
parsed:content::STRING,
['Invoice', 'Contract', 'Receipt']
) AS cls
),
LATERAL (
SELECT AI_EXTRACT(
parsed:content::STRING,
{
'vendor_name': 'Company that issued the document',
'invoice_number': 'Document identifier',
'total_amount': 'Total amount as a number'
}
) AS ext
)
WHERE d.RELATIVE_PATH LIKE '%.pdf';
Automating with Streams and Tasks
Keep the pipeline running as new documents land on the stage.
CREATE OR REPLACE STREAM DOCS_STREAM ON STAGE @DOCS_STAGE;
CREATE OR REPLACE TASK PROCESS_NEW_DOCS
WAREHOUSE = ML_WH
SCHEDULE = '5 MINUTE'
WHEN SYSTEM$STREAM_HAS_DATA('DOCS_STREAM')
AS
INSERT INTO PROCESSED_DOCUMENTS (file_path, doc_category, confidence, vendor_name, invoice_no, total_amount, raw_text)
SELECT /* ... same pipeline query referencing DOCS_STREAM ... */
FROM DOCS_STREAM
WHERE RELATIVE_PATH LIKE '%.pdf' AND METADATA$ACTION = 'INSERT';
Key Takeaways
AI_PARSE_DOCUMENThandles OCR and text extraction from PDFs and images on Snowflake stages.AI_EXTRACTconverts unstructured text into structured JSON fields you define.AI_CLASSIFYtriages documents into categories with a confidence score.- Combine all three for an end-to-end document intelligence pipeline that runs entirely in Snowflake.