Augmenting the prompt — the most overlooked step

Most RAG tutorials write:

f'Use this context.\n\n{context}\n\nQuestion: {q}'

That's a starting point, not production. Better prompts:
- Tell the model when there's no relevant info ('say I don't know')
- Number chunks and ask for citations
- Constrain answer length and format
- Set a system role for tone

The prompt is half the system. A great retrieval with a sloppy prompt still gives meh answers.

Mine, distilled from many tries:

System: 'You are an expert assistant. Answer using only the context. If context is insufficient, say I don't know.'

User: '<context>\n[1] {chunk1}\n[2] {chunk2}\n...\n</context>\n\nQuestion: {q}\n\nRules:\n- Cite sources as [1], [2]\n- Quote relevant snippet for each cite\n- If context insufficient, say so'

That's it. Three rules, one structure, every chunk numbered. The model now knows what to do.

Same shape every time. Numbered context, system role, instructions block. Take this; tweak the system prompt to your domain; ship.

RAG is a knowledge task. You want consistent, grounded answers. Set temperature=0 (or close).

Higher temperatures = more creative phrasing but also more hallucinations and inconsistencies. For 'answer this question from these docs', creativity is a bug, not a feature.

For agent-style flows that need exploration, raise temperature. For Q&A, keep it cold.

Day 62 done.

- Stuffing matters
- Numbered context, citation rule
- Reusable prompt builder
- Temperature = 0 for Q&A

Tomorrow (Day 63): evaluating RAG. Without it, you're shipping vibes. Then we close week 9.