SnowPro Specialty: Gen AI Question 85
Single answerSnowflake CortexA support operations team stores product manuals and troubleshooting guides in a Snowflake table. They want to build a chatbot that answers employee questions using only this internal content, while minimizing custom infrastructure and keeping the workflow inside Snowflake as much as possible. Which approach is the most appropriate with Snowflake Cortex?
- A
Create embeddings for the documents, store them in Snowflake, use vector similarity search to retrieve relevant chunks, and pass the retrieved context to a Cortex LLM function to generate the final answer
- B
Call a Cortex complete function directly on the employee question without retrieval, because Cortex models automatically search all tables in the account for relevant business data
- C
Export the manuals to an external vector database first, because Snowflake Cortex cannot support retrieval workflows on data stored in Snowflake
- D
Fine-tune a model on the full set of manuals before deployment, because Cortex requires model training before it can answer questions about proprietary documents
Show answer and explanation
Correct answer: A
Explanation
The best answer is the RAG-based approach using Snowflake-native components and Cortex capabilities. In practice, teams commonly ingest and chunk source documents, generate embeddings, store both text chunks and vectors in Snowflake, perform vector similarity search to find relevant passages, and then invoke a Cortex LLM function with the retrieved context to produce the answer. This design aligns with best practices for enterprise GenAI: keep data close to where it is governed, ground model outputs in trusted sources, and avoid unnecessary model customization when retrieval is sufficient. Snowflake documentation on Cortex and vector-based retrieval patterns supports using embeddings and similarity search in Snowflake to build grounded question-answering workflows.
- A. Correct.
Correct. This is the standard retrieval-augmented generation (RAG) pattern for grounding responses in enterprise data. In Snowflake, a practical approach is to chunk documents, generate embeddings, store those embeddings in Snowflake, retrieve the most relevant chunks using vector similarity, and then send the question plus retrieved context to a Cortex LLM function for answer generation. This keeps both data and orchestration largely within Snowflake and helps reduce hallucinations by grounding the model in approved internal content.
- B. Incorrect.
Incorrect. Cortex LLM functions do not automatically discover and search arbitrary tables in your Snowflake account. If you call a completion function with only the user prompt, the model responds based on the prompt and its underlying model knowledge, not by implicitly querying enterprise tables. This option reflects a common misconception that managed LLM services automatically have semantic access to all warehouse data.
- C. Incorrect.
Incorrect. Snowflake can support retrieval workflows on data stored directly in Snowflake, including vectorized content and similarity search patterns. While external vector databases are possible in some architectures, they are not required here and would violate the team's goal of minimizing custom infrastructure and staying inside Snowflake as much as possible.
- D. Incorrect.
Incorrect. Fine-tuning is not required for answering questions over proprietary documents. For most enterprise knowledge assistant scenarios, RAG is the preferred approach because it grounds answers in current source material without the cost, complexity, and maintenance burden of model training. This option represents the misconception that proprietary-document Q&A always requires training or fine-tuning.