Retriever Routing over Heterogeneous Indexes
Specify a router that selects one or more candidate retrievers per query based on retriever metadata and query shape.
A router specification: retriever descriptions, selection prompt or rule, multi-select policy, merge behaviour and fallback.
Role
You are designing query routing across several knowledge sources.
Task
Specify a retriever router over {{retriever_inventory}}. For each retriever write the metadata description from {{retriever_metadata}} that the selector will read: content domain, granularity, freshness and known gaps. Define the selection rule {{selection_rule}} over the query signals {{query_signals}}, requiring the selector to output the chosen retriever names and a one-line reason. Apply {{multi_select_policy}} when a query spans sources, including how results are merged and deduplicated. Route unmatched queries to {{fallback_retriever}} and state how the fallback answer is labelled.
Context
A router is only as good as the metadata it reads, so retriever descriptions must state coverage boundaries and gaps rather than advertising capability.
Inputs
- {{retriever_inventory}}
- {{retriever_metadata}}
- {{query_signals}}
- {{selection_rule}}
- {{multi_select_policy}}
- {{fallback_retriever}}
Constraints
- Every retriever description must state what it does not cover
- Selector output must be machine-parseable names plus reason
- Merged results must be deduplicated by source identity
- No query may exit the router unrouted
Output Format
Markdown: retriever metadata table, selection rule block, merge and fallback rules.
Quality Criteria
- Descriptions permit a correct choice without trial retrieval
- Multi-select path defined for cross-domain queries
- Deduplication key specified
- Fallback outputs are visibly labelled