A large enterprise is implementing an internal knowledge base chatbot for its employees. The chatbot needs to answer questions by retrieving relevant information from a vast collection of internal documents, policies, and technical manuals. Crucially, the chatbot must provide answers that are grounded in these specific documents and avoid hallucination. Which generative AI approach, combined with which Azure service, is most suitable?
- AImplementing a basic QnA Maker knowledge base without external document integration.
- BFine-tuning a large language model (LLM) using Azure OpenAI Service.
- CLeveraging Azure AI Language for text summarization and sentiment analysis.
- DUsing a Retrieval Augmented Generation (RAG) approach with Azure AI Search and Azure OpenAI Service.
Show answer & explanationAnswer & explanation
Correct answer: D. Using a Retrieval Augmented Generation (RAG) approach with Azure AI Search and Azure OpenAI Service.
The requirement to provide answers grounded in specific documents and avoid hallucination strongly points to a Retrieval Augmented Generation (RAG) approach. This involves using Azure AI Search to retrieve relevant document snippets and then feeding those snippets, along with the user's query, to an Azure OpenAI Service model to generate a grounded answer.
Why the other options are wrong
- A. QnA Maker can build a knowledge base, but it's less flexible for dynamic retrieval from vast unstructured documents and doesn't inherently use advanced generative AI for nuanced answers.
- B. Fine-tuning can help with style/domain, but doesn't inherently prevent hallucination or ground answers in specific, dynamically retrieved documents as effectively as RAG.
- C. Text summarization and sentiment analysis are analytical tools, not generative approaches for answering questions from a knowledge base.
Retrieval Augmented Generation (RAG)
An architectural pattern that combines information retrieval with generative AI. A search component retrieves relevant documents, which are then used as context for a large language model (LLM) to generate grounded and accurate responses.
- Reduces LLM hallucination.
- Grounds answers in specific, verifiable sources.
- Leverages existing data without retraining LLMs.
- Often uses semantic search for retrieval.
Memory trick: Search first, then Generate – RAG for reliable responses.