Router Complexity Grader

Complexity Grader

Purpose

The Complexity Grader assigns every incoming request to exactly one complexity bucket before any other processing occurs. The bucket drives:

1. Model selection: short-answer requests do not need the highest-capability model; deep-research and full-document requests do
2. Token budget: estimated output token count allows the inference layer to pre-allocate compute and enforce latency budgets
3. Tool activation: only certain buckets unlock RAG, web search, or agentic tool calls
4. Latency targeting: web surface has hard latency SLAs (≤4s for short-answer; ≤12s for medium); the grader ensures requests are not over-processed

Inputs / Signals

Complexity Buckets

short-answer

• Estimated output: ≤ 200 tokens
• Typical requests: single factual question ("What is the limitation period for contract claims in the UAE?"), greeting, yes/no legal question, simple definition, quick calculation with a single variable
• Tools activated: none; model answers from training
• Latency target: ≤ 4 seconds first token on web; ≤ 2 seconds on API

medium

• Estimated output: 200–1,500 tokens
• Typical requests: drafting a single clause, reviewing a 1–2 page excerpt, answering a single-jurisdiction legal question with some elaboration, producing a structured short memo
• Tools activated: RAG if user references their own documents; no web search unless freshness is critical
• Latency target: ≤ 12 seconds on web

deep-research

• Estimated output: 1,500+ tokens
• Typical requests: multi-jurisdiction comparison, recent amendment lookup, regulatory landscape overview, case law research, comparative analysis with recommendations
• Tools activated: RAG (firm KB + public legal corpus), web search for recent amendments/rulings, legal data hunter if needed
• Latency target: up to 30 seconds acceptable with a progress indicator; warn user if exceeding 20 seconds

full-document

• Estimated output: 3,000+ tokens, structured output
• Typical requests: full contract draft, complete memorandum of law, due diligence report, full lease or NDA generation
• Tools activated: RAG for precedents; template engine if available; web for any jurisdiction-specific verification
• Latency target: streaming output expected; total generation may take 60–120 seconds; user should see progress

agentic

• Estimated output: multi-step, variable
• Typical requests: any request referencing a workflow.* skill, multi-step M&A diligence, automated clause-by-clause review with output to a structured report, deposition prep workflow, automated regulatory filing sequence
• Tools activated: full tool suite as specified by the workflow; may include connector calls (Linear, CRM, document storage), calculator tools, multi-turn confirmation gates
• Latency target: asynchronous; user should be informed that results will be delivered in stages

Logic — Grading Rules

Apply rules in order; first match wins:

1. If the message is a greeting, chitchat, admin question, or feature question about Louis → short-answer
2. If the message contains a workflow.* reference or multi-step procedure description → agentic
3. If the message attaches multiple documents (>2) or asks for a complete contract draft → full-document
4. If the message contains "compare", "all jurisdictions", "comprehensive", "thoroughly", "case law search", "recent amendments", or ≥3 jurisdictions → deep-research
5. If the message attaches a single document for review or asks for a single clause draft → medium
6. If none of the above match → short-answer (default; err toward the lighter bucket)

Output

Return a single JSON object on one line:

{
  "complexity": "short-answer|medium|deep-research|full-document|agentic",
  "estimated_tokens_out": <integer>,
  "needs_tools": ["rag", "web", "calc", "legal-data-hunter"],
  "confidence": 0.0-1.0,
  "grading_reason": "<one sentence>"
}

If the request is ambiguous between two buckets, default to the lighter bucket and include a note in grading_reason. Over-grading (escalating a short-answer to deep-research) wastes compute and harms latency more than under-grading.

Why This Matters

Latency is the single most reported user-experience complaint in legal AI products. The grader prevents two failure modes:

1. Over-processing: a greeting triggers RAG retrieval and web search → 8-second response to "Hi"
2. Under-processing: a multi-jurisdiction regulatory comparison is handled with no tools → hallucinated answer citing non-existent statutes

The grader is a cheap operation (runs in < 100ms) and pays for itself in reduced compute waste and improved latency across all other requests.

Related Skills

• [[router-intent-detection]]
• [[router-tool-selector]]
• [[router-confidence-scorer]]
• [[router-platform-aware]]
• [[router-jurisdiction-detector]]
• [[router-skip-rag-when-not-needed]]