Lesson Content

Application prompts should separate developer instructions, user input, and retrieved data. Never concatenate untrusted text into high-priority instructions without clear delimiters and validation. Treat model output as untrusted until parsed and checked.

Structured output allows the application to request fields matching a schema. Validate types, required fields, ranges, and allowed values after generation. Tool calling means the model proposes a tool and arguments; your application decides whether to execute it, validates arguments, enforces authorization, performs the action, and returns the result.

Retrieval systems find relevant chunks from approved documents and include them as context. Quality depends on document freshness, chunking, metadata, source priority, and citation behavior. The model can still misinterpret retrieved material, so high-stakes claims need verification.