Provide Additional Context
Context engineering is supplying the facts, examples, and constraints a model needs at inference time so it reasons over your reality instead of its training-data average.
Why it matters
A model knows nothing about your codebase, your customer, or last week’s incident — without that grounding it confabulates plausible-but-wrong answers. Supplying relevant context is the single highest-leverage move against hallucination, and it is the entire premise of RAG: retrieve the right facts, paste them in, and let the model do the reasoning. In an agent, the observation-reflection output of one step becomes the context of the next.
How it works
Layer context by stability — static at the top (cacheable), volatile near the question:
| Layer | Holds | Example |
|---|---|---|
| System / role | Persona, rules | ”You are a SQL assistant” |
| Reference | Facts, docs, schema | table DDL, API spec |
| Few-shot | Input→output pairs | 2-3 worked examples |
| Query | The actual ask | ”write the join for…” |
Two patterns dominate. Few-shot prompting shows 2-5 examples so the model infers the pattern — far more reliable than describing it. Retrieval (embeddings-and-vector-search) pulls only the top-k relevant chunks instead of dumping everything, because every token competes for the window and models lose recall in the middle of long inputs (“lost in the middle”). Delimit injected context clearly (XML tags, ---, triple backticks) so the model never confuses data for instructions — the same boundary that defends against prompt-injection-jailbreaks.
Example
Grounding a support bot with retrieved facts instead of trusting memory:
SYSTEM: Answer ONLY from <docs>. If absent, say "I don't know."
<docs>
Refund policy: digital goods refundable within 14 days of purchase.
Order #8821: purchased 2026-05-02, type=digital.
</docs>
USER: Can I refund order 8821? (today: 2026-05-30)
→ "No — order #8821 was purchased on 2026-05-02, past the
14-day window for digital goods."
Without the <docs> block the model guesses a generic policy; with it, the answer is correct and auditable.
Pitfalls
- Dumping everything — stuffing the whole wiki buries the relevant chunk and triggers lost-in-the-middle recall loss; retrieve, don’t paste.
- Stale context — pasted facts are a snapshot; an agent that caches them across turns will answer from outdated state.
- No source boundary — un-delimited context lets injected text (“ignore previous instructions”) read as a command.
- Contradictory layers — a system rule fighting a few-shot example yields unpredictable precedence; keep them consistent.