Databricks Document Parsing

Parse unstructured documents into structured text using ai_parse_document — the foundation for document processing and custom RAG pipelines on Databricks.

When to Use

Use this skill when:

• Parsing PDFs, DOCX, PPTX, or images into text
• Extracting structured data from unstructured documents
• Building a custom RAG pipeline (parse → chunk → index → query)
• Ingesting documents from Unity Catalog Volumes for search or analysis

Overview

ai_parse_document is a SQL AI function that extracts content from binary documents. It runs on serverless SQL warehouses and supports PDF, DOC/DOCX, PPT/PPTX, JPG/JPEG, and PNG.

Aspect	Detail
Function	ai_parse_document(content) or ai_parse_document(content, options)
Input	Binary document content (from read_files with format => 'binaryFile')
Output	VARIANT with document.pages[], document.elements[], metadata
Requirements	Databricks Runtime 17.1+, Serverless SQL Warehouse
Tool	Use via execute_sql — no dedicated MCP tool needed

Quick Start

Parse all documents in a Volume:

SELECT
  path,
  ai_parse_document(content) AS parsed
FROM read_files('/Volumes/catalog/schema/volume/docs/', format => 'binaryFile');

Common Patterns

Pattern 1: Parse with Options

SELECT ai_parse_document(
  content,
  map(
    'version', '2.0',
    'imageOutputPath', '/Volumes/catalog/schema/volume/images/',
    'descriptionElementTypes', '*'
  )
) AS parsed
FROM read_files('/Volumes/catalog/schema/volume/invoices/', format => 'binaryFile');

Options:

Key	Values	Description
version	'2.0'	Output schema version
imageOutputPath	Volume path	Save rendered page images
descriptionElementTypes	'', 'figure', '*'	AI-generated descriptions (default: '*' for all)

Pattern 2: Parse + Extract Structured Data

Combine ai_parse_document with ai_query to extract specific fields. Use transform() + try_cast() to concatenate element text, then pass the full text to ai_query with returnType => 'STRING'.

WITH parsed_documents AS (
  SELECT
    path,
    ai_parse_document(content) AS parsed
  FROM read_files('/Volumes/catalog/schema/volume/invoices/', format => 'binaryFile')
),
parsed_text AS (
  SELECT
    path,
    concat_ws('\n\n',
      transform(
        try_cast(parsed:document:elements AS ARRAY<VARIANT>),
        element -> try_cast(element:content AS STRING)
      )
    ) AS text
  FROM parsed_documents
  WHERE try_cast(parsed:error_status AS STRING) IS NULL
)
SELECT
  path,
  ai_query(
    'databricks-claude-sonnet-4',
    concat(
      'Extract vendor name, invoice number, and total due from this document. ',
      'Return the result as a JSON object with keys: vendor, invoice_number, total_due. ',
      text
    ),
    returnType => 'STRING'
  ) AS structured_data
FROM parsed_text
WHERE text IS NOT NULL;

Pattern 3: Custom RAG Pipeline

End-to-end: parse documents → chunk text → store in Delta table → create Vector Search index.

Step 1 — Parse and chunk into a Delta table:

ai_parse_document returns a VARIANT. You must use variant_get with an explicit ARRAY<VARIANT> cast before calling explode, since explode() does not accept raw VARIANT values.

CREATE OR REPLACE TABLE catalog.schema.parsed_chunks AS
WITH parsed AS (
  SELECT
    path,
    ai_parse_document(content) AS doc
  FROM read_files('/Volumes/catalog/schema/volume/docs/', format => 'binaryFile')
),
elements AS (
  SELECT
    path,
    explode(variant_get(doc, '$.document.elements', 'ARRAY<VARIANT>')) AS element
  FROM parsed
)
SELECT
  md5(concat(path, variant_get(element, '$.content', 'STRING'))) AS chunk_id,
  path AS source_path,
  variant_get(element, '$.content', 'STRING') AS content,
  variant_get(element, '$.type', 'STRING') AS element_type,
  current_timestamp() AS parsed_at
FROM elements
WHERE variant_get(element, '$.content', 'STRING') IS NOT NULL
  AND length(trim(variant_get(element, '$.content', 'STRING'))) > 10;

Step 1a (production) — Incremental parsing with Structured Streaming:

For production pipelines where new documents arrive over time, use Structured Streaming with checkpoints for exactly-once processing. Each run processes only new files (tracked via checkpoints), then stops with trigger(availableNow=True).

See the official bundle example: databricks/bundle-examples/contrib/job_with_ai_parse_document

Stage 1 — Parse raw documents (streaming):

from pyspark.sql.functions import col, current_timestamp, expr

files_df = (
    spark.readStream.format("binaryFile")
    .option("pathGlobFilter", "*.{pdf,jpg,jpeg,png}")
    .option("recursiveFileLookup", "true")
    .load("/Volumes/catalog/schema/volume/docs/")
)

parsed_df = (
    files_df
    .repartition(8, expr("crc32(path) % 8"))
    .withColumn("parsed", expr("""
        ai_parse_document(content, map(
            'version', '2.0',
            'descriptionElementTypes', '*'
        ))
    """))
    .withColumn("parsed_at", current_timestamp())
    .select("path", "parsed", "parsed_at")
)

(
    parsed_df.writeStream.format("delta")
    .outputMode("append")
    .option("checkpointLocation", "/Volumes/catalog/schema/checkpoints/01_parse")
    .option("mergeSchema", "true")
    .trigger(availableNow=True)
    .toTable("catalog.schema.parsed_documents_raw")
)

Stage 2 — Extract text from parsed VARIANT (streaming):

Uses transform() to extract element content from the VARIANT array, and try_cast for safe access. Error rows are preserved but flagged.

from pyspark.sql.functions import col, concat_ws, expr, lit, when

parsed_stream = spark.readStream.format("delta").table("catalog.schema.parsed_documents_raw")

text_df = (
    parsed_stream
    .withColumn("text",
        when(
            expr("try_cast(parsed:error_status AS STRING)").isNotNull(), lit(None)
        ).otherwise(
            concat_ws("\n\n", expr("""
                transform(
                    try_cast(parsed:document:elements AS ARRAY),
                    element -> try_cast(element:content AS STRING)
                )
            """))
        )
    )
    .withColumn("error_status", expr("try_cast(parsed:error_status AS STRING)"))
    .select("path", "text", "error_status", "parsed_at")
)

(
    text_df.writeStream.format("delta")
    .outputMode("append")
    .option("checkpointLocation", "/Volumes/catalog/schema/checkpoints/02_text")
    .option("mergeSchema", "true")
    .trigger(availableNow=True)
    .toTable("catalog.schema.parsed_documents_text")
)

Key techniques from the official example:

• repartition by file hash — parallelizes ai_parse_document across workers
• trigger(availableNow=True) — processes all pending files then stops (batch-like)
• Checkpoints — exactly-once guarantee; no re-parsing on re-runs
• transform() + try_cast — safer than explode + variant_get for text extraction
• Three-stage pipeline — separate parse/text/structured stages with independent checkpoints

Step 1b — Enable Change Data Feed (required for Vector Search Delta Sync):

ALTER TABLE catalog.schema.parsed_chunks
SET TBLPROPERTIES (delta.enableChangeDataFeed = true);

Step 2 — Create a Vector Search index and query it:

Use the databricks-vector-search skill to create a Delta Sync index on the chunked table and query it. Ensure CDF is enabled first (Step 1b above).

Output Schema

ai_parse_document returns a VARIANT with this structure:

document
├── pages[]          -- page id, image_uri
└── elements[]       -- extracted content
    ├── type         -- "text", "table", "figure", etc.
    ├── content      -- extracted text
    ├── bbox         -- bounding box coordinates
    └── description  -- AI-generated description
metadata             -- file info, schema version
error_status[]       -- errors per page (if any)

Common Issues

Issue	Solution
Function not available	Requires Runtime 17.1+ and Serverless SQL Warehouse
Region not supported	US/EU regions, or enable cross-geography routing
Large documents	Use LIMIT during development to control costs
explode() fails with VARIANT	explode() requires ARRAY, not VARIANT. Use variant_get(doc, '$.document.elements', 'ARRAY<VARIANT>') to cast before exploding
Short/noisy chunks	Filter with length(trim(...)) > 10 — parsing produces tiny fragments (page numbers, headers) that pollute the index
ai_query returns markdown fences	Use returnType => 'STRING' for clean output. If fences still appear, strip with `regexp_replace(result, '```(json)?\s*
Re-parsing unchanged documents	Use Structured Streaming with checkpoints — see Pattern 3, Step 1a

Related Skills

• databricks-vector-search — Create indexes and query embeddings (Step 2 of RAG)
• databricks-agent-bricks — Pre-built Knowledge Assistants (out-of-the-box RAG without custom parsing)
• databricks-spark-declarative-pipelines — Production pipelines for batch document processing
• databricks-dbsql — Full AI functions reference including ai_query, ai_extract, ai_classify