Skip to main content

Akabot Vision — Release Notes

v3.2.0 — April 2026​

New features

  • Generative extraction — Use a large language model to extract fields from documents without a predefined template — just describe the fields in plain text.
  • Handwritten text support — OCR engine now handles cursive and mixed print/handwritten content with a new HTR (Handwritten Text Recognition) model.
  • Batch processing queue — Submit hundreds of documents in a single API call; Vision processes them asynchronously and posts results to a webhook.

Improvements

  • Table detection accuracy on complex invoices with nested headers improved from 87% to 94%.
  • Validation station UI redesigned: keyboard-only navigation, bulk-accept for high-confidence pages.
  • Model training now requires 30% fewer labeled samples to reach production-ready accuracy.

Bug fixes

  • Fixed confidence scores displaying as 0% for multi-page PDFs after the 10th page.
  • Resolved an issue where rotating a document 180° in preprocessing broke the field coordinate mapping.
  • Fixed extraction template editor not saving changes when switching document types without explicit save.

v3.1.0 — January 2026​

New features

  • Document splitting — Automatically detect and split multi-document PDF bundles into individual files before classification.
  • Signature detection — New field type that detects the presence or absence of a signature in a defined region.
  • REST API v2 — Synchronous extraction endpoint for documents under 5 MB; no polling required.

Improvements

  • Vietnamese language model accuracy improved by 18% after retraining on expanded corpus.
  • Processing throughput increased to 500 pages/minute per tenant (up from 200).
  • Extraction results now include bounding box coordinates for each extracted value.

Bug fixes

  • Fixed an intermittent failure when processing TIFF files with CCITT Group 4 compression.
  • Corrected field extraction on documents with right-to-left (RTL) text mixed with LTR content.

v3.0.0 — September 2025​

New features

  • Akabot Vision v3 launch — Rebuilt on a transformer-based OCR and extraction pipeline.
  • No-code template editor — Draw field zones on a document image and label them — no configuration files required.
  • Human-in-the-loop station — Embedded web UI for reviewers to validate and correct uncertain extractions before they flow downstream.

Breaking changes

  • Vision v2 template format (.vtemplate) is not compatible with v3. Re-create templates using the new editor.
  • The v2 Python SDK for Vision is deprecated; use the REST API or the new Studio activity pack.

Bug fixes

  • Resolved a memory exhaustion issue when processing PDF files larger than 100 MB.