Routing Heuristics
Description
Section titled “Description”A router should classify query intent before retrieval begins. The objective is to separate focused lookups from exhaustive synthesis and route each to the cheapest path that preserves answer quality.
Strong Signals for RAG
Section titled “Strong Signals for RAG”- Query targets a specific document, person, or known artifact.
- User asks for a single fact or narrow lookup.
- Expected answer likely exists in one or a few sources.
Examples:
- “What is our enterprise pricing tier?”
- “Find references to the Chicago office.”
Strong Signals for RLM
Section titled “Strong Signals for RLM”- Query asks for “all”, “every”, “across”, “patterns”, or “compare”.
- User asks for summary over a broad period or corpus segment.
- Query requires cross-source aggregation before answering.
Examples:
- “Summarize all postmortems from last quarter.”
- “Compare onboarding quality across departments.”
Signals for Hybrid Routing
Section titled “Signals for Hybrid Routing”- Query has a broad synthesis goal plus a clear topical filter.
- A pre-filter can reduce corpus size before recursive analysis.
Example:
- “Analyze all support tickets about billing and list systemic failure patterns.”
Hybrid path:
- Retrieve likely billing-related material.
- Run exhaustive reasoning on that narrowed corpus.
Operational Guidance
Section titled “Operational Guidance”- Start with simple heuristics and explicit keywords.
- Add model-based routing only when heuristic recall/precision is insufficient.
- Log route decisions and answer quality outcomes for iterative tuning.