Extraction Light
Changelog for the Extraction Light profile. This section covers CPU-optimized models, edge deployments, cost reduction strategies, and resource-efficient processing options.
What’s Included
- Edge Deployments: On-device and low-compute extraction options
- Cost Optimization: Reduced compute profiles for budget-sensitive workflows
- Fallback Strategies: Graceful degradation when full models are unavailable
- Mobile Support: SDK optimizations for iOS and Android capture apps
Recent Updates
2024-12-08 — CPU-Only Mode Improvements
Optimized CPU-only extraction path with SIMD (Single Instruction, Multiple Data) vectorization. Processing speed improved by 25% on ARM-based edge devices without GPU.
- Impact: Latency
2024-11-22 — Invoice Extractor Light v2
Released invoice-extractor-light-v2 profile with 60% reduced model size. Maintains 94% of full-model accuracy on standard invoice formats.
- Impact: Accuracy
2024-11-08 — Adaptive OCR Fallback
Introduced intelligent fallback to lightweight OCR when document complexity is low. Reduces average compute cost by 35% for simple documents.
- Impact: Reliability
2024-10-28 — Mobile SDK: Bandwidth Optimization
Added client-side JPEG compression (configurable quality 70-95%) before upload. Reduces bandwidth usage by 40-60% for mobile capture applications.
- Impact: Latency
2024-10-14 — Edge Container Updates
Updated lightweight container image with Alpine Linux base. Image size reduced from 1.2GB to 420MB for faster deployment on edge infrastructure.
- Impact: UX
2024-09-30 — Graceful Degradation Mode
Added automatic quality-tier fallback when GPU resources are constrained. Documents processed at reduced quality rather than queued indefinitely.
- Impact: Reliability
2024-09-18 — Field Subset Extraction
Introduced field selection parameter allowing extraction of specific fields only. Reduces processing time for workflows requiring limited data capture.
- Impact: Latency
Compatibility Notes
- Light profiles require SDK version 1.3.0 or later
- Edge containers compatible with ARM64 and x86_64 architectures
- Fallback modes can be disabled via API parameter
strict_mode: true
Roadmap (Next Quarter)
- WebAssembly extraction module for browser-only deployments
- Offline mode with local model caching
- Battery-aware processing for mobile devices