Module 3 · Lesson
How an assistant works
Every assistant platform, whatever its brand name, combines the same parts.
| Part | What it is | Your job |
|---|---|---|
| Instructions | Standing directions the model follows in every conversation (sometimes called a system prompt). | Define role, scope, rules, tone and escalation. |
| Knowledge | Documents the assistant can draw on. | Supply current, clean, well-organised sources. |
| Capabilities or tools | Optional abilities such as web search, data analysis or connections to other apps. | Enable only what is needed. For a policy assistant, usually turn web search off so answers come only from your documents. |
| Guardrails | Rules about what not to do, plus the platform’s own safety features. | Decide the refusals and escalations. |
| Part | What it is | Your job |
|---|---|---|
| Feedback loop | A way to learn from real use. | Review questions and failures; update knowledge and instructions. |
Retrieval, in plain language Your handbook is usually too long to fit in every conversation. So when a user asks a question, the platform searches your documents for the passages most relevant to that question, adds them to the prompt behind the scenes, and the model writes its answer from them. This is called retrieval- augmented generation (RAG). Three consequences follow:
- Document quality determines answer quality. If two documents disagree, or one is out of date, the assistant may retrieve the wrong one.
- Retrieval can miss. If the answer is buried in a table, a scanned image, or phrased very differently from the question, it may not be found.
- The model can still hallucinate. If nothing relevant is retrieved, a poorly instructed assistant may answer from its general training. Your instructions must forbid that.
Jot down key ideas in your own words. Saves automatically.