| name | intent-router |
| description | Classify the information-need of a query and dispatch it to the appropriate retrieval or reasoning strategy. Use before read-side memory access, before multi-strategy retrieval, or any time you'd otherwise default to "one retriever for everything". Returns a strategy label, a token budget, and a retrieval depth so downstream handlers can be specialised. Backed by Pre-Route (arxiv 2605.10235v2) and MemFlow (arxiv 2605.03312v1), which together show LLMs possess latent routing ability elicitable via a structured prompt — and that externalising the routing decision improves small-model performance by ~2x. Triggers: "route this", "what strategy", "before retrieving", "intent classification", or any query whose ideal handling depends on what KIND of question it is.
|
| user-invocable | true |
| version | 1.0.2 |
| format | "2025-10-02T00:00:00.000Z" |
| triggers | ["route this query","what retrieval strategy","what reasoning strategy","intent classification before retrieval","front-run the memory access"] |
| updated | "2026-05-16T00:00:00.000Z" |
| status | ACTIVE |
| source | arxiv 2605.10235v2 (Pre-Route), 2605.03312v1 (MemFlow) |
Intent Router
Why
The 2026 frontier converges on one architectural shift: route → strategy → act, replacing the historical one-retriever-for-everything pattern. Different query intents demand categorically different handlers:
- lookup — direct factual recall, single retrieval, small budget
- multi-hop — chain through intermediate facts, deeper retrieval, larger budget
- global — summarise/aggregate across a broad slice, retrieve broadly + compress
- verification — check a claim against evidence, retrieve narrowly + compare
- deep-reason — escalate to a higher-tier model, retrieve generously, allow chain-of-thought
- lexical — exact-match / keyword (BM25 territory), no embedding, narrow budget
The empirical claim: a single fixed strategy is dominated on every axis by the conditional strategy.
How
Before any memory access or multi-strategy retrieval, run a structured-prompt classification pass that returns:
{
"strategy": "lookup|multi-hop|global|verification|deep-reason|lexical",
"token_budget": 2000,
"retrieval_depth": 5,
"refinement_policy": "none|verbal-rerank|consensus-check",
"confidence": 0.0
}
If confidence < 0.6, fall back to the broadest strategy (global) so the system degrades gracefully rather than mis-routing.
Output format vs confidence
The full JSON envelope is the canonical form when downstream consumers read it programmatically (typed pipelines, gating code). For inline human-facing routing decisions, taper the ceremony to match the confidence:
- confidence ≥ 0.9 — emit just the strategy label on one line (e.g., ). The other fields default; don't print the envelope.
