3.3.3. Advanced Responses: Prompts, Knowledge, and HTTP
💡 First Principle: Inside a topic, you can manufacture a response from three different raw materials: model reasoning (a custom prompt), retrieved content (scoped knowledge through generative answers), or live system data (an HTTP request). Choosing correctly is asking: does this answer come from thinking, from documents, or from an API?
| Mechanism | Raw material | Configure | Reach for it when |
|---|---|---|---|
| Custom prompt node | Model reasoning over supplied inputs | Prompt instructions, input bindings, model (2.4.2) | Summarize, classify, extract, draft — transform what you already have |
| Generative answers (scoped) | Retrieved knowledge chunks | Data sources per node, custom instructions (3.3.4) | Answer from documents/indexes, cited |
| Send HTTP request node | Live API response | Method, URL, headers, body; parse response (JSON schema) into variables | A system must be queried or commanded mid-topic, no connector at hand |
The custom prompt node embeds an AI Builder prompt as a step: conversation variables flow in as prompt inputs, the model's output lands in a variable — reasoning as plumbing. Scoped generative answers ground a specific node on specific sources, overriding the agent's general knowledge behavior for that moment of the conversation (the priority rule from 2.4.1). The Send HTTP request node is the raw escape hatch: full control over the request, response parsed against a schema into typed variables — powerful, but unmanaged (no connector governance, credentials handled in-node), which is why "is there a connector for this?" remains the first question.
⚠️ Exam Trap: The three mechanisms have non-overlapping signatures — "summarize what the user said" is never HTTP; "answer from the policy index" is never a bare prompt; "fetch the live shipment status from a REST endpoint" is never generative answers. Distractors swap them; the raw material gives it away.
Reflection Question: One topic must fetch an order (API), summarize its history for the user (reasoning), and answer a follow-up policy question (documents). Sequence the three nodes and name what each passes to the next via variables.
