SnowPro Specialty: Gen AI Question 187
Single answerTranscriptsA customer support team stores call-center conversation transcripts in Snowflake and wants to build a retrieval-augmented chatbot that answers questions using the most relevant parts of each conversation. The transcripts are long, multi-speaker documents, and users often ask about specific promises made by an agent during a call. Which approach is the MOST appropriate to improve retrieval quality before creating embeddings?
- A
Split each transcript into smaller semantically meaningful chunks, preserving local context such as speaker turns and timestamps where useful, then generate embeddings for those chunks
- B
Generate one embedding per full transcript so the vector captures the entire conversation in a single representation
- C
Convert each transcript into a single sentiment score and embed that score for similarity search
- D
Store transcripts as raw text only and rely on SQL pattern matching instead of embeddings for semantic retrieval
Show answer and explanation
Correct answer: A
Explanation
For transcript-based retrieval use cases, best practice is to preprocess long documents into smaller, semantically coherent chunks before generating embeddings and building retrieval. This is especially important for multi-speaker transcripts, where one call may cover several topics and only a small section may answer the user’s question. Chunk-level embeddings typically improve recall and precision compared with whole-document embeddings. Metadata such as speaker, timestamp, conversation ID, or topic labels can also help with filtering and grounding results. This aligns with common retrieval-augmented generation design guidance in Snowflake-based GenAI workflows: prepare text carefully, generate embeddings over appropriately sized chunks, and retrieve the most relevant chunks rather than entire long documents.
- A. Correct.
Correct. Long transcripts usually contain multiple topics, speakers, and context shifts. Chunking the transcript into smaller, coherent segments before embedding generally improves semantic retrieval because each vector represents a focused portion of the conversation. Preserving metadata such as speaker turns, call identifiers, and timestamps can further improve downstream retrieval and grounding, especially when users ask about a specific statement made during part of a call.
- B. Incorrect.
Incorrect. A single embedding for an entire long transcript often loses fine-grained detail and reduces retrieval precision. When a user asks about a specific promise or topic buried inside a lengthy call, the full-document embedding may be too coarse to surface the exact relevant passage.
- C. Incorrect.
Incorrect. Sentiment is only one narrow attribute of a transcript and does not preserve the actual semantic content needed for question answering. Embedding a sentiment score would not help retrieve the portion of the transcript discussing refund terms, escalation commitments, or delivery dates.
- D. Incorrect.
Incorrect. SQL pattern matching can help with exact keyword search, but it is not a substitute for semantic retrieval in a GenAI chatbot. Users may ask questions with wording that differs from the transcript text, and embeddings are designed to support similarity-based retrieval across paraphrases and related concepts.