Expand description
Turning retrieval results into text an LLM can use.
The daemon answers in JSON built for programs: record ids, distances, nested traversal nodes. Handing that to a model wastes context on syntax and buries the parts that matter. Everything here renders the same data as compact prose-with-structure, keeps the numbers a reader would act on (retrieval score, similarity, edge confidence, utility) and drops the ones nobody reads.
Every renderer is total: an empty result set produces guidance about what to try next, not an empty string.