Intent Classification

The Intent Classification Layer is the first gate every request passes through. It determines what the user is trying to accomplish, a quick autocomplete, a code review, a multi-file refactor, or a creative writing task, and tags the request with semantic intent metadata that downstream layers use for routing decisions.

Why It Matters

Not all LLM requests are equal. A “write me a React component” request needs a different provider, model, and budget than a “explain this error” request. Intent classification ensures Layerr doesn’t treat a five-minute coding task the same as a deep reasoning session.

Routing Categories

Every request is assigned exactly one of 8 routing categories. The category taxonomy is defined server-side in routing/prompt-builder.ts (DEFAULT_CATEGORIES), so it cannot be overridden by individual clients.

Category	When It’s Chosen
CODING	Writing, debugging, reviewing, or explaining code. Any request involving programming languages, scripts, algorithms, or software development.
REASONING	Complex analysis, comparisons, multi-step logic, math, science explanations, strategic thinking, or anything requiring deep thought.
CREATIVE	Writing stories, poems, marketing copy, brainstorming, humor, or any open-ended creative task.
VISION	Only when the user has attached an image and wants it analyzed, described, or interpreted (requires an image attachment).
DOCUMENT	Only when the user has attached a document (PDF, text file) and wants it summarized, analyzed, or queried (requires an attachment).
FAST	Pure micro-interactions with no knowledge retrieval, greetings, single-word acknowledgements, or trivial arithmetic needing no explanation.
SECURITY	Security analysis, vulnerability assessment, threat modeling, CTF challenges, penetration testing, malware analysis, or cybersecurity topics.
GENERAL	The default for conversational queries, factual questions, explanations, and summaries that don’t clearly fit a more specific category. When in doubt, GENERAL.

Routing rules bias the classifier toward the more specific category when a request could match several (e.g. CODING over GENERAL for anything code-related, GENERAL over FAST whenever a fact or concept must be retrieved or explained).

Core Concepts

Concept	Description	Key File
Intent Signal	A typed classification of what the request is asking for	routing/types.ts
Classification Confidence	How certain the classifier is about its decision	workflow/classifier.ts
Workflow Style	Whether the task is agentic, interactive, batch, or streaming	workload/signals.ts
Routing Prompt	The assembled prompt sent to the routing engine	routing/prompt-builder.ts

The Classification Pipeline

Raw Request
    |
    v
[1] Extract Text Signals  --> urgency keywords, tool keywords, complexity cues
    |
    v
[2] Classify Workflow    --> classifyWorkflow() in workflow/classifier.ts
    |
    v
[3] Build Routing Prompt  --> routing/prompt-builder.ts
    |
    v
[4] Enrich Decision       --> enrichRoutingDecision() in routing/enrichment.ts

Signal Extraction

The system extracts signals from the request text to infer intent. Key signal types:

Urgency Signals

• High urgency: “fix now”, “urgent”, “broken”, “error”, “crash”
• Low urgency: “explain”, “review”, “when you have time”, “draft”

Tool Usage Prediction

• Agentic style: “create a file”, “run tests”, “deploy”, “execute”
• Interactive style: “what is”, “how do”, “explain why”
• Batch style: “refactor all”, “migrate every”, “batch process”

Complexity Signals

• Simple: “hello”, “quick question”, “one-liner”
• Medium: “component”, “function”, “small script”
• Complex: “architecture”, “system design”, “multi-file”, “review entire codebase”

File Reference

File	What It Does
workflow/classifier.ts	Main classification entrypoint. Parses request text, returns intent classification with confidence scores
routing/prompt-builder.ts	Assembles the routing prompt from classified intent + workload context
routing/types.ts	Type definitions for intent signals, routing decisions, and strategy selectors
workload/signals.ts	Signal extraction helpers, urgency, cost sensitivity, latency sensitivity

Integration with Downstream Layers

The intent classification feeds directly into:

1. Workload Analysis, the classified intent is combined with file-context signals to build a workload profile
2. Strategy Engine, intent determines which strategy (cost, speed, quality, balanced) is most appropriate
3. Provider Scoring, intent signals weight the scoring dimensions (e.g., coding intent prioritizes code-capable models)
4. Explainability, the classified intent appears in routing explanations so users understand why a model was chosen

Guardrails

• Confidence threshold: Classifications below a confidence floor are marked uncertain and trigger conservative routing (balanced strategy, premium provider)
• Ambiguity fallback: Multi-intent requests (e.g., “explain and fix”) are split or routed to the most capable model
• Workspace override: Workspace profiles can pin specific intents to specific strategies (e.g., “always use speed for autocompletes”)