arcrouter-classifier

Fast prompt classifier for LLM routing. Classifies prompts by topic and complexity in <1ms without any LLM calls.

Used internally by ArcRouter to route 345+ models across 24 topic categories.

Install

npm install arcrouter-classifier

Quick Start

import { scorePrompt, detectTopic, normalizeRoutingBudget } from 'arcrouter-classifier';

// Complexity scoring
const complexity = scorePrompt('Write a binary search tree in TypeScript');
console.log(complexity.tier);       // "COMPLEX"
console.log(complexity.confidence); // 0.92
console.log(complexity.isAgentic);  // false

// Topic detection
const topic = detectTopic('Explain quantum entanglement');
console.log(topic); // "science/physics"

// Budget normalization
console.log(normalizeRoutingBudget('low'));    // "economy"
console.log(normalizeRoutingBudget('medium')); // "auto"
console.log(normalizeRoutingBudget('high'));   // "premium"

API

scorePrompt(prompt, hasToolsArray?)

Classify a prompt into one of four complexity tiers.

const result = scorePrompt('Prove that P ≠ NP using a step-by-step argument');

result.tier        // "SIMPLE" | "MEDIUM" | "COMPLEX" | "REASONING"
result.score       // Raw score used to determine tier
result.confidence  // 0-1 sigmoid-calibrated confidence
result.isAgentic   // true if tools[] present or agentic keywords detected
result.reason      // Human-readable explanation

Tiers: | Tier | Threshold | Typical prompts | |------|-----------|-----------------| | SIMPLE | score ≤ 12 | "What is X?", single-word lookups | | MEDIUM | 12 < score ≤ 35 | Explanations, comparisons, how-to guides | | COMPLEX | score > 35 | Code, math, long analysis | | REASONING | COMPLEX + ≥2 reasoning markers | Proofs, step-by-step derivations, trade-off analysis |

detectTopic(prompt)

Detect the topic category using two-pass detection (top-level then subcategory).

detectTopic('Debug this React useEffect hook')
// → "code/debugging"

detectTopic('Calculate the integral of e^x from 0 to 1')
// → "math/calculus"

detectTopic('What are the symptoms of COVID-19?')
// → "science/medicine"

Categories:

• code / code/frontend / code/backend / code/algorithms / code/devops / code/security / code/debugging
• math / math/calculus / math/algebra / math/statistics / math/discrete
• science / science/physics / science/chemistry / science/biology / science/medicine
• writing / writing/creative / writing/technical / writing/business / writing/academic
• reasoning / reasoning/logical / reasoning/multistep
• general

detectTopicDetailed(prompt)

Returns both primary and secondary categories with confidence scores.

const result = detectTopicDetailed('Fix the authentication bug in my Express middleware');
result.primary    // "code"
result.secondary  // "code/security"
result.confidence // 0.8

normalizeRoutingBudget(raw?)

Normalize any budget string to a canonical tier.

normalizeRoutingBudget('low')     // → "economy"
normalizeRoutingBudget('medium')  // → "auto"
normalizeRoutingBudget('high')    // → "premium"
normalizeRoutingBudget('free')    // → "free"
normalizeRoutingBudget(undefined) // → "auto"

compressMessages(messages)

Lossless prompt compression — reduces token count for long conversations.

const { messages: compressed, stats } = compressMessages(messages);

stats.original_chars    // e.g. 8420
stats.compressed_chars  // e.g. 7600
stats.saved_chars       // e.g. 820
stats.ratio             // e.g. 0.903 (lower is better)
stats.layers_applied    // e.g. ["L1:dedup", "L2:whitespace"]

Compression fires above 5000 chars and applies:

• L1 — Deduplicates repeated assistant messages
• L2 — Normalizes whitespace (trailing spaces, extra blank lines)
• L5 — Compacts pretty-printed JSON in tool results

Agentic Detection

scorePrompt sets isAgentic: true when the prompt contains tool-use keywords (read file, run command, git commit, function_call, etc.) or when hasToolsArray = true.

scorePrompt('Read file config.json and update the port').isAgentic  // true
scorePrompt('What is 2 + 2?').isAgentic                             // false
scorePrompt('Any prompt', true).isAgentic                           // true (tools[] present)

How ArcRouter Uses This

1. scorePrompt() → complexity tier → selects model tier (cheapest to most capable)
2. detectTopicDetailed() → topic + subcategory → semantic model selection
3. normalizeRoutingBudget() → budget tier → cost sensitivity weighting
4. compressMessages() → reduce tokens on long conversations → saves cost

The full ArcRouter stack (semantic routing, benchmark scoring, council consensus) runs on Cloudflare Workers at api.arcrouter.com.

Links

• ArcRouter API: https://api.arcrouter.com
• SDK: https://github.com/ArcRouterAI/arcrouter-sdk (npm install @arcrouter/sdk)
• Docs: https://arcrouter.com/docs
• MCP server: claude mcp add arcrouter --transport http https://api.arcrouter.com/mcp

License

MIT