What's New in Nuxeo AI (version 5.0.2)
Nuxeo AI 5.0.2 introduces the nuxeo-ai-content-intelligence addon — bringing automatic AI-powered descriptions, summaries, and entity tagging to your images and documents using Hyland's Content Intelligence cloud service.
New Features
Automatic Image and Document Enrichment
Install the new nuxeo-ai-content-intelligence package to enable out-of-the-box AI enrichment:
- Images — Automatically generates a natural-language description and tags detected entities such as people, places, and organizations.
- Documents (PDF, Word, Excel, PowerPoint, plain text) — Automatically summarizes the content and tags named entities found in the text.
Enrichment typically completes in seconds for standard files. Larger files (around 100 MB or more) may take a few minutes.
Broad Image Format Support
All common image formats — including BMP, GIF, and WebP — are automatically handled. The addon transcodes unsupported formats behind the scenes so enrichment works end-to-end without any manual conversion.
Works Alongside Existing AI Providers
If you already use AWS, GCP, or Sightengine enrichment, the new addon coexists cleanly — duplicate tags are prevented automatically, and the AI-generated description is always written independent of your tag configuration.
Improvements
Configurable Enrichment Timeout
A new nuxeo.ai.enrichment.call.timeout.seconds property lets you control how long the system waits for an enrichment provider to respond (default: 60s). This is raised automatically by the Content Intelligence addon to accommodate longer-running document analysis.
Reduced Log Noise for Deduplication Configuration
Missing deduplication configuration warnings now appear only once with a clear diagnostic message, instead of repeating on every event.
Installation
nuxeoctl mp-install nuxeo-ai-content-intelligence
Add the following to your nuxeo.conf:
nuxeo.ai.images.enabled=true
nuxeo.ai.contentintelligence.enabled=true
nuxeo.ai.contentintelligence.documents.enabled=true
nuxeo.enrichment.raiseEvent=true
nuxeo.enrichment.save.facets=true
nuxeo.hyland.cic.auth.baseUrl=<your CIC auth URL>
nuxeo.hyland.cic.contextEnrichment.baseUrl=<your CIC enrichment URL>
nuxeo.hyland.cic.enrichment.clientId=<your client ID>
nuxeo.hyland.cic.enrichment.clientSecret=<your client secret>
Requirements
- Nuxeo Server LTS 2025
- Valid Hyland Content Innovation Cloud credentials
What's New in Nuxeo AI for LTS 2025 (version 5.0.0)
Nuxeo AI 5.0.0 brings faster and more reliable AI-powered content enrichment, broader content safety detection, and richer image analysis — all built on modernized cloud infrastructure designed for long-term stability. This release requires Nuxeo Server LTS 2025.
Released Changes
New Features
Faster, Future-Proof AWS Integration
AWS-powered AI services — including text translation, document analysis, and image recognition — now run on the latest AWS SDK (v2), delivering faster response times and improved error handling. A new abstraction layer also makes future cloud infrastructure upgrades seamless, so your AI workflows remain uninterrupted.
Flexible Search Backend Support
AI-powered search and asset counting are no longer tied to a specific search engine. You can now swap or upgrade your search backend without disrupting AI features, giving you greater flexibility in how you architect your content platform.
Bug Fixes
Image Quality Analysis Now Works Reliably With Sightengine
Previously, image quality checks via Sightengine could fail with errors, preventing image analysis results from being returned. This has been resolved — image quality and safety analysis now completes successfully.
AI Enrichment No Longer Fails Unexpectedly During Retries
In rare cases, AI enrichment processes could crash during automatic retries, leaving documents un-enriched. Retry handling is now more robust, ensuring enrichment completes reliably even when transient errors occur.
Accurate Document Counts in Search and Asset Reports
Search results and asset count reports could sometimes show incorrect totals. Document counts are now accurate across all search and reporting operations.
Reliable Face and Celebrity Detection for All Images
Face and celebrity detection could fail on images that lacked orientation metadata (for example, screenshots or programmatically generated images). These images are now processed successfully.
Improved Memory Management During Image Processing
Image resizing operations could consume excess memory over time. Resource handling has been improved to ensure stable performance during bulk image processing.
Improvements
More Reliable and Efficient Content Enrichment
AI enrichment pipelines are now more resilient to unexpected responses from cloud providers, reducing the chance of enrichment failures. PDF text extraction uses less memory, improving performance for large documents. Image resizing now preserves the requested output format (for examplem PNG, GIF) instead of saving the changes in JPEG format by default, ensuring your assets maintain the quality you expect.
Richer Image Analysis From Google Cloud Vision
You now get more detailed results from GCP Vision-powered analysis:
- Face detection returns full emotion likelihoods (joy, anger, sorrow, surprise, and so on), bounding boxes, and confidence scores.
- Logo and landmark detection includes bounding box geometry and proper classification tags.
- Label detection output is now compatible with downstream classification workflows.
These enhancements give you deeper insight into your visual content for automated tagging and moderation.
Broader Offensive Content Detection
Content safety analysis now detects a wider range of offensive imagery, including hate symbols and offensive gestures. This helps you enforce content policies more comprehensively.
Richer Color Metadata for Images
Image quality analysis now returns HSV (Hue, Saturation, Value) color data, giving you more detailed color information for use in visual search, asset categorization, and design workflows.
Easier Troubleshooting
When AWS permissions are misconfigured, log messages now clearly indicate what's wrong and how to fix it. Search and asset counting issues that previously failed silently now produce visible warnings, making it faster to diagnose and resolve problems.
More Accurate AI Model Training Data
Dataset statistics now correctly reflect multi-class configurations across all fields, ensuring that your custom AI model training starts from accurate data — preventing unexpected behavior caused by stale settings.