AI and Generative AI Solutions
What to know
Generative AI questions focus on architecture choices around large models. Key concepts: foundation models accessed via API (Alibaba's Qwen family, offered through Model Studio/DashScope-style platforms), inference hosting on GPU ECS or PAI (Platform for AI) for training and custom models, and retrieval-augmented generation (RAG) to ground answers in private data. RAG combines a vector store (embeddings in a vector database) with a retriever and the model, reducing hallucination without retraining. Also understand prompt engineering, fine-tuning versus RAG, guardrails/content moderation, and cost/latency trade-offs between model sizes.
Traps to watch
A common trap is reaching for fine-tuning when the real need is fresh private knowledge-RAG is usually the better fit for frequently changing data, while fine-tuning shapes style/behavior. Another is assuming a bigger model is always better; smaller models cut latency and cost for simple tasks. Watch for answers that ignore data privacy (sending sensitive data to external endpoints without controls), skip evaluation, or claim RAG eliminates all hallucination. Distinguish training (needs GPUs, PAI) from inference-only (API calls).
Rule of thumb
Ask "is the gap knowledge or behavior?" Knowledge that changes → RAG with a vector store. Behavior/tone/format → fine-tuning or prompting. Start with prompting, add RAG for grounding, fine-tune only when needed.
Worked example
"A support bot must answer from an internal, frequently updated knowledge base." Answer: embed documents into a vector store, retrieve relevant chunks per query, and pass them to a foundation model via API with a grounded prompt and moderation-no retraining required.