Snowflake Document AI: Extract Data from Unstructured Documents

Document AI is a managed Snowflake capability that extracts structured data — text, numbers, dates, and more — from unstructured documents like PDFs and images. It combines pre-trained OCR (Optical Character Recognition) and NER (Named Entity Recognition) models so you can pull required fields from invoices, receipts, contracts, and medical records directly with SQL.

Document AI Architecture

Documents (PDF / images)
  -> stored in a Stage (Internal or External)
    -> referenced with BUILD_SCOPED_FILE_URL()
      -> Document AI model (inference via the PREDICT function)
        -> extraction result as VARIANT
          -> FLATTEN + SELECT into a structured table

Document AI runs on Snowflake's Serverless Compute, so you don't have to provision a warehouse yourself. Extraction results come back as a VARIANT value, and you can access individual fields using JSON path notation.

Creating a Document AI Model

-- Create a Document AI model (creating it in the Snowsight UI is recommended)
-- Example of creating it in SQL
CREATE OR REPLACE SNOWFLAKE.ML.DOCUMENT_INTELLIGENCE
  invoice_extractor
  FROM @doc_stage
  USING (
    SELECT
      BUILD_SCOPED_FILE_URL(@doc_stage, relative_path) AS file_url
    FROM DIRECTORY(@doc_stage)
    WHERE relative_path LIKE '%.pdf'
    LIMIT 10
  );

In the Snowsight UI you can visually annotate the fields (Values) you want to extract on the uploaded sample documents, then fine-tune the model from there.

Example Field Definitions

File Reference Functions

-- BUILD_SCOPED_FILE_URL: a session-scoped temporary URL
-- The most common input to the Document AI PREDICT function
SELECT BUILD_SCOPED_FILE_URL(
  @invoice_stage,
  'invoices/2026/INV-001.pdf'
) AS scoped_url;

-- BUILD_STAGE_FILE_URL: a stage reference URL
SELECT BUILD_STAGE_FILE_URL(
  @invoice_stage,
  'invoices/2026/INV-001.pdf'
) AS stage_url;

-- GET_PRESIGNED_URL: a pre-signed URL (for sharing externally)
SELECT GET_PRESIGNED_URL(
  @invoice_stage,
  'invoices/2026/INV-001.pdf',
  3600  -- expiry in seconds
) AS presigned_url;

Comparing File Reference Functions

Inference with the PREDICT Function

-- Inference on a single document
SELECT invoice_extractor!PREDICT(
  BUILD_SCOPED_FILE_URL(@invoice_stage, 'invoices/INV-001.pdf')
) AS result;

-- Extract individual fields from the result
SELECT
  result:invoice_number::VARCHAR AS invoice_number,
  result:invoice_date::DATE AS invoice_date,
  result:total_amount::NUMBER(12,2) AS total_amount,
  result:vendor_name::VARCHAR AS vendor_name,
  result:__confidence::FLOAT AS overall_confidence
FROM (
  SELECT invoice_extractor!PREDICT(
    BUILD_SCOPED_FILE_URL(@invoice_stage, 'invoices/INV-001.pdf')
  ) AS result
);

Batch Processing Pipeline

-- Batch inference combined with Directory Tables
INSERT INTO extracted_invoices (file_path, invoice_number, invoice_date, total_amount, vendor_name, confidence, extracted_at)
SELECT
  d.relative_path,
  r.result:invoice_number::VARCHAR,
  r.result:invoice_date::DATE,
  r.result:total_amount::NUMBER(12,2),
  r.result:vendor_name::VARCHAR,
  r.result:__confidence::FLOAT,
  CURRENT_TIMESTAMP()
FROM DIRECTORY(@invoice_stage) d,
LATERAL (
  SELECT invoice_extractor!PREDICT(
    BUILD_SCOPED_FILE_URL(@invoice_stage, d.relative_path)
  ) AS result
) r
WHERE d.relative_path LIKE '%.pdf'
  AND d.relative_path NOT IN (
    SELECT file_path FROM extracted_invoices
  );

Extracting Line Items (Array Data)

-- Expand the line items with FLATTEN
SELECT
  result:invoice_number::VARCHAR AS invoice_number,
  item.value:description::VARCHAR AS item_description,
  item.value:quantity::INT AS quantity,
  item.value:unit_price::NUMBER(10,2) AS unit_price,
  item.value:amount::NUMBER(12,2) AS line_amount
FROM (
  SELECT invoice_extractor!PREDICT(
    BUILD_SCOPED_FILE_URL(@invoice_stage, 'invoices/INV-001.pdf')
  ) AS result
),
LATERAL FLATTEN(input => result:line_items) AS item;

Using Confidence Scores

Each extracted field comes with a confidence score from 0.0 to 1.0. Low-confidence fields need human review, so the production best practice is to set a threshold and route anything below it to a review queue.

-- Send low-confidence extractions to a review queue
INSERT INTO review_queue (file_path, field_name, extracted_value, confidence)
SELECT
  relative_path,
  'total_amount',
  result:total_amount::VARCHAR,
  result:total_amount_confidence::FLOAT
FROM extracted_results
WHERE result:total_amount_confidence::FLOAT < 0.85;

Required Privileges

Best Practices

• Validate accuracy on sample documents before batch processing: run 5-10 test inferences to verify accuracy and cost, then move on to the production batch
• Set a confidence threshold and build a human-review workflow: for example, auto-approve anything at 0.85 or above and route the rest to a review queue
• Use Directory Tables and Streams for arrival-driven processing: build a pipeline that auto-detects new documents, runs PREDICT, and stores the results in your extracted-data table
• Keep the raw VARIANT result and structure it later: model updates can change the set of extracted fields, so retain the raw VARIANT in addition to the parsed columns

Check Your Understanding

Frequently Asked Questions