Unstructured Data Analysis in Snowflake: AI_PARSE_DOCUMENT, AI_EXTRACT, Cortex Search and Multimodal AI

1.Turning files into text: AI_PARSE_DOCUMENT

Most unstructured analysis starts with files such as PDFs, scans and filings sitting on a stage. AI_PARSE_DOCUMENT reads documents from internal or external stages and keeps the reading order and structural elements such as tables and headers. You can then pass its output to other Cortex AI functions for extraction, classification or enrichment, or index it for RAG.

The exam tests the difference between its two modes. LAYOUT mode is the default recommendation, especially for complex documents. It is optimized for structured content such as tables, headers and layout relationships, and it is the only mode that can extract images (set 'extract_images': true, which costs nothing extra). OCR mode is the faster choice when you only need high-quality text from scanned or text-heavy documents such as contracts, insurance claims and manuals.

SELECT AI_PARSE_DOCUMENT (
    TO_FILE('@docs.doc_stage','research-paper-example.pdf'),
    {'mode': 'LAYOUT' , 'page_split': true}) AS research_paper_example;

In both modes, page_split returns each page of a multi-page document separately, and page_filter limits processing to the pages you list. Setting page_filter turns on page_split automatically. LAYOUT output is Markdown, so headers come back as # lines and tables as Markdown tables. That matters later, because a Markdown-aware splitter can use that structure when it chunks the text.

Sources1

2.Pulling structured fields out: AI_EXTRACT

Use AI_EXTRACT when you want specific values out of a document, not the whole text. It runs on arctic-extract, a vision-based LLM, and it can read graphical content as well as text-heavy paragraphs, including logos, handwritten text such as signatures, tables and checkmarks. It returns three shapes of answer:

- Entity: a natural-language question or description, for example the city or the ZIP code. - List: a JSON schema for an array, for example every account holder on a bank statement. - Table: a JSON schema that gives the table title and the columns to extract.

Questions should be in plain English, specific, and ask for one value each. "What is the date?" is a poor question for a document that contains both an issue date and a signature date.

Most table-extraction accuracy comes from how you define the columns. List them in the order they appear in the document, and put repeated values such as Invoice Number first. Copy column names exactly as the document writes them (Product Code, not product_code). Add a Section column when the table is split into named sections. For hierarchical headers, join the parent names into the column name. The output limit is the constraint most people miss: table answers stop at 4096 tokens. For a table that runs across several pages, split the document into one-page documents and join the results afterwards. If even one page is too dense, split the table by columns.

Sources2

3.Preparing text for retrieval: recursive and Markdown splitting

For search and RAG, long parsed text has to be cut into chunks. SNOWFLAKE.CORTEX.SPLIT_TEXT_RECURSIVE_CHARACTER first splits on the highest-priority separator. Any chunk that is still longer than chunk_size is split again on the next separator, and this repeats until every chunk fits. The format argument sets the default separators. With 'none', only the separator list is used, and the default list is paragraph break, line break, space, then any character. With 'markdown', the text is also split on headers, code blocks and tables, so it works well on AI_PARSE_DOCUMENT LAYOUT output. The optional overlap repeats some characters from the previous chunk so each chunk keeps some context.

SELECT
   doc_id,
   c.value
FROM
   sample_documents,
   LATERAL FLATTEN( input => SNOWFLAKE.CORTEX.SPLIT_TEXT_RECURSIVE_CHARACTER (
      document,
       ? ,
      25,
      10
   )) c;

Chunk sizing is one of the few retrieval-quality settings you control yourself. Even though models such as snowflake-arctic-embed-l-v2.0-8k accept long inputs, smaller chunks usually give better retrieval and better downstream LLM answers. Two details lead to wrong answers on the exam. First, chunk_size in the splitter is counted in characters, while the 512 recommendation is in tokens, at roughly 4 characters per token. Second, text longer than the embedding model's context window is not rejected. Cortex Search truncates it before embedding, but keyword retrieval still uses the full text.

Sources34

4.Cortex Search: embedding models and semantic reranking

Cortex Search is the managed retrieval layer for RAG. Every query combines vector search for semantic matches, keyword search for lexical matches, and a semantic reranking step that reorders the top results. You choose the embedding model with EMBEDDING_MODEL when you create the service. Embedding, indexing and refresh are managed for you, and TARGET_LAG controls how often the service picks up changes.

CREATE OR REPLACE CORTEX SEARCH SERVICE transcript_search_service
  ON transcript_text
  ATTRIBUTES region
  WAREHOUSE = cortex_search_wh
  TARGET_LAG = '1 day'
  EMBEDDING_MODEL = 'snowflake-arctic-embed-l-v2.0'
  AS (
    SELECT
        transcript_text,
        region,
        agent_id
    FROM support_transcripts
);

Reranking is on by default and trades latency for quality. Turning it off with "reranker": "none" in scoring_config saves 100–300 ms per query on average. How much quality you lose depends on the workload, so Snowflake advises comparing results with and without reranking before you disable it. You can also change the relative weights of the text, vector and reranker scores. Reranking does not apply to batch search queries.

{
  "scoring_config": {
      "reranker": "none"
  }
}

Sources45

5.Multimodal analytics: audio, images and similarity

Unstructured data also includes audio and images, and the same AI functions handle them. AI_COMPLETE takes images as well as text, and TO_FILE creates the stage-file reference that AI_COMPLETE and other file-accepting functions need. AI_TRANSCRIBE converts staged audio and video into text with timestamps and speaker information, and from there the text-analysis functions take over. AI_PARSE_DOCUMENT with extract_images gives you a document's embedded images, which you can then tag or analyze with AI_EXTRACT or AI_COMPLETE. AI_PARSE_DOCUMENT names multimodal RAG as one of its use cases.

For comparing content, AI_EMBED produces an embedding vector for text or an image. AI_MULTI_EMBED creates embeddings from text, images, audio or video, with segment metadata, for semantic search across modalities. AI_SIMILARITY calculates the embedding similarity between two inputs, which is useful for questions like "how close is this ticket to that known issue". The sources here give only that one-line description of AI_SIMILARITY. They do not list its supported input types or options.

Sources61

Exam traps

Each one states something that sounds right. Open it to see what is actually true.

1. 1.OCR mode is the right choice whenever documents contain tables, because it is the higher-quality mode.Why is that wrong?

   OCR mode is for fast text extraction from scanned or text-heavy documents. Tables, headers, layout relationships and image extraction all need LAYOUT mode.

   Covered in Turning files into text: AI_PARSE_DOCUMENT

2. 2.Setting chunk_size to 512 in SPLIT_TEXT_RECURSIVE_CHARACTER gives you the recommended 512-token chunks.Why is that wrong?

   chunk_size counts characters, not tokens. Since a token is about 4 characters, a 512-token chunk is roughly 2,000 characters.

   Covered in Preparing text for retrieval: recursive and Markdown splitting

3. 3.Picking the longest-context embedding model and indexing whole documents gives the best retrieval quality.Why is that wrong?

   Smaller chunks usually give better retrieval and better downstream LLM answers, even when a long-context model is available.

   Covered in Preparing text for retrieval: recursive and Markdown splitting