Databricks Certified Generative AI Engineer Associate free question 4 / 10 · Domain 2: Data Preparation
A team parses specification PDFs with many tables using ai_parse_document (output schema version "2.0"). The design keeps each table in one chunk without cutting it mid-row and loads the chunks into an AI Search index. In what format should the chunking code expect the content of a part recognized as a table?
- A TSV string with columns separated by tabs, one table row per line
- A string that lists each row as a separate JSON array
- A string of an HTML table (table, tr and td tags)
- A string in Markdown table notation (columns separated by |)
Show the answer and explanation
Answer: C
Why this is correct
The description of the return value of ai_parse_document (Returns) says that for version 2.0, tables are represented in HTML. The content of a part recognized as a table comes back as a string in HTML form, so the code that builds chunks should be written to treat it as HTML. It can use the string as the body of one chunk as it is, or read rows and columns from the tags and tidy them. Because the relationship between rows and columns is kept in the tags, a table with header rows or cells that span several rows passes to the next stage without losing its structure.
Why the other options are wrong
"A string in Markdown table notation (columns separated by |)" … Tables in the 2.0 output are not in Markdown notation. Code that splits columns on | cannot split a string containing tags correctly, and the chunks would have broken table content.
"A string that lists each row as a separate JSON array" … The output as a whole has a nested, JSON-like shape, but the contents of a table are not an array per row. The content of a table is a single HTML string.
"A TSV string with columns separated by tabs, one table row per line" … Tables are not returned with a delimiter between columns. Code that splits on tabs would put the whole string with its tags into one column and could not extract the rows and columns of the table.
Key point
The output schema carries a major.minor version. A minor update is backward compatible and only adds fields, but a major update can add, remove or rename fields. Databricks says the default version may be raised, so pinning the version with "2.0" in the version argument keeps the treatment of tables from changing suddenly when the default changes.