SnowPro Specialty: Gen AI Question 146
Single answerUnstructured dataA retail company stores thousands of product manuals as PDF files in an internal stage in Snowflake. The Gen AI team wants to build a retrieval-augmented chatbot that answers support questions using the content of those manuals. They need a solution that minimizes data movement, keeps processing inside Snowflake where possible, and allows the text to be indexed for semantic search. Which approach is the MOST appropriate?
- A
Use Snowflake directory tables on the stage to read the PDF contents directly and create embeddings from the file metadata.
- B
Create an external function to send each PDF to a third-party OCR service, store the returned text outside Snowflake, and query it remotely at runtime.
- C
Use Snowflake Document AI to extract text from the staged PDF files, store the extracted text in Snowflake tables, and then generate embeddings for search and retrieval.
- D
Load the PDF files into a VARIANT column and use standard SQL string functions to tokenize the binary content for semantic search.
Show answer and explanation
Correct answer: C
Explanation
For unstructured documents such as PDFs, the key challenge is converting file content into usable text before applying Gen AI techniques like chunking, embedding generation, and semantic retrieval. In Snowflake, Document AI is designed to extract information from documents stored in stages, making it a strong fit when an organization wants to keep processing inside Snowflake and reduce external data movement. After extraction, the text can be stored in relational tables and used in downstream retrieval workflows, including vector embeddings and semantic search patterns used in RAG applications. Directory tables are useful for file metadata management, not document content extraction. Likewise, raw binary PDF content is not appropriate for direct SQL tokenization. A third-party OCR path may work technically, but it is less aligned with the scenario requirements around minimizing movement and maximizing Snowflake-native processing. These patterns are consistent with Snowflake guidance for handling unstructured data, staged files, and document-processing workflows for AI applications.
- A. Incorrect.
Incorrect. Directory tables provide metadata about files in a stage, such as file names and paths, but they do not parse or extract the textual contents of PDF documents. Creating embeddings from metadata alone would not support accurate retrieval over manual content. This distractor reflects a common misunderstanding that staged file metadata is sufficient for unstructured document understanding.
- B. Incorrect.
Incorrect. While an external function and third-party OCR workflow could be built, it does not meet the stated goal of minimizing data movement and keeping processing inside Snowflake where possible. It also adds operational complexity, external dependencies, and potential governance concerns. Candidates may choose this if they know OCR is needed but overlook Snowflake-native options.
- C. Correct.
Correct. For PDF manuals stored in stages, a Snowflake-native pattern is to use Document AI to extract structured text from unstructured documents, persist that text in Snowflake tables, and then create embeddings for semantic retrieval. This approach aligns with Gen AI best practices for keeping data in-platform, reducing unnecessary movement, and preparing content for vector-based search in a RAG architecture.
- D. Incorrect.
Incorrect. PDF files are not meaningfully searchable by applying SQL string functions to binary or raw loaded content. PDFs require parsing or extraction of text before downstream NLP tasks such as chunking, embedding generation, or semantic search. This option represents the misconception that unstructured binary files can be treated like plain text without preprocessing.