Retrieval routing is the difference between "search everything the same way" and "choose a retrieval path because the query demands it."
For Cortagent, the important part is not naming many retrieval methods. The important part is keeping the choice visible.
What routing decides
A routing layer can decide whether a query should use:
- vector retrieval for semantic similarity,
- keyword retrieval for exact terms and identifiers,
- graph-style retrieval for relationship-heavy questions,
- or a more complex path when decomposition is justified.
Constraint
More retrieval is not automatically better retrieval. Extra retrieval can add latency and noise. The router has to justify fan-out.
Public-safe routing questions
The router should be able to answer:
- Why this retrieval path?
- Why now?
- What was skipped?
- What diagnostics were produced?
Those questions matter because retrieval errors often look like answer errors. If routing is opaque, debugging starts too late.
What we avoid saying publicly
We do not publish private routing thresholds, corpus-specific heuristics, customer data, or internal traces. The public point is the architecture boundary: retrieval choice is explicit and reviewable.