Classification Light
Changelog for the Classification Light profile. This section covers distilled models, rule-assisted classification, cost-aware routing, and explainability features for lightweight deployments.
What’s Included
- Distilled Models: Compact classifiers with reduced compute requirements
- Rule-Assisted Classification: Hybrid approach combining rules and ML (Machine Learning)
- Cost-Aware Routing: Intelligent model selection based on document complexity
- Explainability: Human-readable explanations for classification decisions
Recent Updates
2024-12-10 — Distilled Classifier v3
Released third-generation distilled classifier with 70% size reduction compared to full model. Maintains 96% of original accuracy on standard document types.
- Impact: Latency
2024-11-26 — Rule-Based Fallback Mode
Added configurable rule-based classification for documents matching known patterns. Reduces model invocations by 40% for homogeneous document streams.
- Impact: Reliability
2024-11-12 — Explainability Endpoint
Introduced /classify/explain endpoint returning top contributing features for each prediction. Supports compliance requirements for automated decision systems.
- Impact: Compliance
2024-10-30 — Cost-Aware Routing Engine
Deployed intelligent routing that selects light vs. full classifier based on document complexity score. Average cost per classification reduced by 28%.
- Impact: Reliability
2024-10-18 — Serverless Function Support
Optimized light classifier for serverless environments with cold start under 800ms. Compatible with AWS Lambda, Google Cloud Functions, and Azure Functions.
- Impact: Latency
2024-10-04 — On-Device Caching
Added local model caching in mobile SDK. Subsequent classifications after initial load complete in under 50ms on modern devices.
- Impact: Latency
2024-09-22 — Threshold Configuration
Introduced per-label confidence thresholds allowing different routing rules for high-stakes vs. routine categories.
- Impact: Accuracy
2024-09-08 — Label Sync Optimization
Reduced Studio label propagation time to light models from 5 minutes to under 60 seconds. Changes reflect immediately in subsequent API calls.
- Impact: UX
Compatibility Notes
- Distilled v3 requires SDK 1.4.0 or later
- Explainability endpoint requires Enterprise tier
- Rule-based fallback configured via Console workflow editor
Roadmap (Next Quarter)
- Browser-based classification via WebAssembly module
- Custom rule language for advanced routing logic
- Confidence threshold auto-tuning based on historical accuracy