SnowPro Specialty: Gen AI Question 57
Single answerUnstructured data use casesA financial services company stores thousands of quarterly reports, investment research PDFs, and earnings call transcripts in a Snowflake stage. Analysts want to ask natural-language questions such as "What risk factors were mentioned for Company X last quarter?" and receive answers grounded in the source documents. The team wants to minimize data movement and keep both the documents and derived search data inside Snowflake. Which approach is the MOST appropriate?
- A
Build a retrieval-augmented generation pipeline by extracting text from the staged documents, generating embeddings, storing them in Snowflake, and using vector search to retrieve relevant chunks for the LLM
- B
Convert all PDFs and transcripts into structured relational tables only, then use SQL pattern matching instead of embeddings because unstructured documents cannot be used directly in Gen AI workflows in Snowflake
- C
Move the documents to an external vector database because Snowflake can store unstructured files but cannot support similarity search over document embeddings
- D
Use only a large language model with the full text of every report in the prompt for each analyst question, because this avoids the need to preprocess documents or store embeddings
Show answer and explanation
Correct answer: A
Explanation
The best answer is the RAG-based design in Option 1. For unstructured data use cases such as question answering over PDFs, transcripts, and reports, Snowflake best practices align with keeping files in stages, extracting text, chunking content, generating embeddings, and using vector similarity search to retrieve relevant passages. Those passages are then provided to an LLM to generate a grounded answer. This pattern is preferred over keyword-only approaches because it supports semantic retrieval, and it is preferred over prompting with the full corpus because it scales better and reduces cost and hallucination risk. Snowflake documentation and Gen AI guidance emphasize unstructured data handling in stages, embedding-based retrieval, and vector search as key building blocks for enterprise Gen AI solutions.
- A. Correct.
Correct. This is the standard and most practical architecture for question answering over unstructured enterprise documents. The workflow is to keep documents in Snowflake stages, extract and chunk text, generate embeddings, store those embeddings in Snowflake, and perform vector similarity search to retrieve the most relevant chunks at query time. Those retrieved chunks are then supplied to the LLM as grounding context in a retrieval-augmented generation (RAG) pattern. This approach minimizes hallucination risk, reduces prompt size, and keeps data and derived search artifacts in Snowflake.
- B. Incorrect.
Incorrect. While some metadata or extracted fields may be modeled relationally, the claim that unstructured documents cannot be used directly in Gen AI workflows in Snowflake is false. Snowflake supports unstructured data storage and enables patterns where text is extracted from files and used for embedding generation and retrieval. SQL pattern matching alone is also a poor fit for semantic question answering over long documents because it does not capture meaning the way embeddings and vector search do.
- C. Incorrect.
Incorrect. This is a plausible misconception because many teams historically used external vector databases for RAG, but Snowflake supports storing embeddings and performing vector similarity search within Snowflake. If the goal is to minimize data movement and keep both source documents and derived search data inside Snowflake, moving everything to an external vector store is not the most appropriate approach.
- D. Incorrect.
Incorrect. Sending the full text of every report in every prompt is inefficient, expensive, and often infeasible due to context window limits. It also increases latency and does not scale as the document corpus grows. Best practice for enterprise document QA is to retrieve only the most relevant chunks and ground the model on those results rather than prompting with the entire corpus.