Caching And Corpus Ordering

Radar sends large, stable source material before the smaller changing question.

That message order helps providers with prefix caching:

1. System prompt.
2. Corpus or document.
3. Question or focus instruction.

If you ask several questions about the same files, the corpus prefix can remain stable while only the final question changes.

Why It Matters

Provider caching can reduce latency and cost when the beginning of a request matches a previous request. Radar's tool implementations preserve that ordering by design.

Practical Advice

• Group related questions around the same file set.
• Avoid adding unnecessary files to the corpus.
• Use search first when you do not know which files matter.
• Use RADAR_VERBOSE=1 to inspect input size and timing.