Changelog

​

Versions

• Backend NPM package: 1.48.0

​

What’s New

​

Enterprise Edition Separation

• EE Directory Structure: All enterprise features now organized in dedicated ee/ directory
• License-based Entitlements: New entitlements system with license checking for premium features
• Feature Differentiation: Clear documentation of CE vs EE feature sets
• Moved to EE: Agentic retrieval, BullMQ queues, RBAC, workers, and tokenizer are now enterprise features

​

Advanced Document Processing System

• Multi-format Support: Process PDF, DOCX, DOC, TXT, and MD files
• Python-based PDF Processor: Uses MIT-licensed Docling library for accurate extraction
• VLM Validation: Vision Language Model validates and corrects OCR results
• Smart Chunking: Page-aware markdown chunking with heading hierarchy preservation
• Table Extraction: Automatic table detection and normalization in markdown format
• Hybrid Chunking: Combines semantic understanding with token-based splitting
• Standalone Executable: Build script creates portable PDF processor executable

​

Code Architecture Improvements

• Provider-Agent Separation: Clear distinction between ExuluProvider (LLM configuration) and ExuluAgent (provider + tools + contexts)
• Code-defined Agents: Support for defining agents in code, not just via database/API/UI
• Refactored Load Logic: Agent loading logic moved into ExuluApp as methods
• Shared Utilities: Extracted withRetry utility for reusable retry logic across codebase
• Consolidated Schemas: Updated GraphQL resolvers and schemas for better organization

​

Documentation Updates

• CE vs EE Comparison: New pages documenting feature differences
• Partners Program: Documentation for licensing and partnerships
• Enhanced API Reference: Expanded with EE features
• Migration Guides: Clear instructions for upgrading to new architecture

​

Bug Fixes

• Import Errors: Fixed import path issues after EE separation
• TypeScript Paths: Updated configuration for EE modules
• Queue Structure: Updated validators for new queue organization

​

Dependency Updates

​

New Dependencies

• docling - Python library for document processing
• mammoth - DOCX to markdown conversion
• turndown - HTML to markdown conversion

​

Breaking Changes

​

Migration Guide

1. Update to @exulu/backend@latest
2. Review your usage of enterprise features
3. Update import paths if you’re importing from internal modules
4. Check the CE vs EE documentation to understand feature availability

npm install @exulu/backend@latest

​

What This Means For You

• Clear Licensing: Understand exactly which features are available in each edition
• Better Architecture: Cleaner separation of concerns between providers and agents
• Advanced Processing: Extract structured content from complex documents with high accuracy
• VLM Validation: Leverage vision models to improve OCR quality
• Code-first Agents: Define agents programmatically for better version control and deployment
• Improved Maintainability: Clearer code structure makes customization easier

​

Versions

• Backend NPM package: 1.47.0

​

What’s New

​

Agentic Retrieval System

• Multi-step Retrieval Agent: Implements an intelligent tool loop architecture that dynamically selects the best search strategy
• Smart Search Strategy Selection: Automatically chooses between hybrid, semantic, or keyword search based on query context
• Multilingual Support: Built-in language detection and reasoning support for 8+ languages
• Intelligent Reranking: Improves search relevance by reordering results based on semantic similarity
• Resilient API Handling: Retry logic with exponential backoff ensures reliable operations

​

NLP Query Preprocessing

• Language Detection: Automatic detection using the franc library with support for multiple languages
• Language-Specific Stemming: Implements the natural NLP library for better query understanding
• Query Variations: Generates multilingual query variations for improved search coverage
• PostgreSQL FTS Integration: Maps ISO 639-3 language codes to PostgreSQL full-text search configurations

​

Expanded Default AI Model Support

​

New Cerebras Default Models

• GPT-OSS-120B (128k context)
• LLAMA-3.1-8B (32k context)
• LLAMA-3.3-70B (32k context)

​

Perplexity Internet Search

• Live Web Search: Query the internet in real-time via Perplexity API
• Advanced Filtering: Support for domain filtering and recency filtering
• Automatic Citations: Properly formatted citations in search results
• Rate Limit Handling: Built-in protection against API rate limits

​

Code Quality & Developer Experience

​

Refactored Codebase Architecture

• Modular Organization: Refactored classes.ts into individual modules for better maintainability
• Fixed Import References: Corrected all source import paths for consistency
• Decorator Folder Fix: Corrected typo in registry decorator folder name (decoraters → decorators) and updated all import paths

​

Enhanced Development Tooling

• ESLint Integration: Configured with TypeScript support for consistent code quality
• Prettier Formatting: Automatic code formatting for a clean codebase
• Jest Testing Framework: Set up with ts-jest for robust testing
• Code Duplication Detection: Added jscpd for identifying and managing code duplication
• npm Scripts: New commands for testing, linting, formatting, and validation

npm test           # Run tests
npm run lint       # Check code quality
npm run format     # Format code
npm run validate   # Type check + lint + test
npm run cpd        # Check code duplication

​

Stricter TypeScript Configuration

• Enabled noImplicitAny and strictNullChecks
• Added noUnusedLocals and noUnusedParameters checks
• Updated target from ESNext to ES2022 for better compatibility

​

Bug Fixes & Improvements

​

Security & Stability

• npm Audit Vulnerabilities: Fixed all identified security vulnerabilities
• Removed Debug Code: Cleaned up development artifacts and debug file writes from agentic-retrieval.ts
• Agent ID Conflicts: Fixed duplicate agent IDs in Claude Sonnet 4 and 4.5 agents

​

Code Cleanup

• Linting Errors: Resolved all linting errors across the codebase
• Prettier Formatting: Applied consistent code formatting
• Documentation Organization: Moved file storage documentation to proper folder structure
• Removed Unused Code: Cleaned up commented console.log statements and unnecessary imports
• Math Tools Removed: Removed unused math tools (arithmetic, statistics, trigonometric)

​

Dependency Updates

​

Major Upgrades

• Vercel AI SDK: Upgraded from v5 to v6 for the latest AI capabilities
• AI Provider SDKs: Updated Anthropic, Google Vertex, and OpenAI SDKs

​

New Dependencies

• @ai-sdk/cerebras - Cerebras model support
• @perplexity-ai/perplexity_ai - Web search integration
• franc - Language detection
• natural - NLP and stemming
• wink-nlp - Additional NLP capabilities
• zod-to-json-schema - Schema conversion utilities
• jscpd - Code duplication detection

​

What This Means For You

• Better Code Quality: Enhanced tooling and stricter TypeScript checks catch issues earlier
• Easier Debugging: Cleaner codebase with removed debug artifacts
• More Maintainable: Modular architecture makes it easier to understand and extend
• Smarter Search: Agentic retrieval and NLP preprocessing deliver more relevant results
• Multilingual Support: Your agents can now handle queries in multiple languages
• More AI Options: Choose from expanded model offerings including Cerebras and Perplexity
• Live Web Search: Agents can now access real-time information from the internet
• More Reliable: Bug fixes and security updates ensure stable operations
• Better Performance: Optimized search tools and consolidated architecture
• Enhanced Security: Resolved npm audit vulnerabilities

​

Getting Started

npm install @exulu/backend@latest

​

Versions

• Backend NPM package: 1.45.0
• Frontend Docker image: latest | 1.27.0
• Frontend NPM package: 1.16.1

​

What’s New

​

Enhanced Workflow Automation

• CRON Scheduling: Automate workflow execution with flexible CRON-based scheduling featuring built-in presets (hourly, daily, weekly, monthly) or custom CRON expressions with cron-validator integration for pattern validation
• Simplified Workflow Object: Streamlined workflow schema with direct agent references via UUID, removing unnecessary complexity around ownership, variables, and RBAC fields
• Run History Tracking: Comprehensive history view for every workflow execution with full result, error, and metadata inspection
• Real-time Monitoring: Live status indicators showing active, completed, and failed workflow runs with instant updates
• Visual Feedback: Enhanced UI with status badges, tooltips, and at-a-glance health indicators for better workflow visibility
• Improved UX: Added persistent “Save as Workflow” banner in chat interface for easier workflow template creation, and users can now edit individual chat messages that are part of a workflow
• Workflow Queue Configuration: Workflows are created using a specific Agent instance. It is now possible to define a queue on an ExuluProvider class (in the new workflows property), which if provided enables you to define concurrency limits (in code) and automated schedules (in the UI and API)

​

Chat Experience Overhaul

• Server Components Migration: Chat pages now leverage Next.js server components for dramatically improved initial load performance
• Session Auto-creation: Cleaner flow when starting new conversations with automatic session management
• Improved Session UI: Streamlined the Chat Session UI by hiding metadata such as created date and edit options behind hover states
• Chat Search: New dedicated search page to find specific messages or sessions across your entire chat history
• Save as Workflow: Persistent banner in chat interface making it easy to convert conversations into reusable workflow templates
• Automatic Citations: Agents using context search tools now automatically format inline citations with item metadata, chunk IDs, and context references, these are automatically processed, showing citation badges in the chat including a dialog window showing the original chunk the citation references
• User Personalization: Enhanced system prompts with user firstname/lastname fields for better agent personalization (configurable via privacy settings)

​

Agent & Prompt Management

• Prompt Library: Browse prompts by tags with grouped category views at /prompts/[tag]
• Agent Prompt Templates Browser: Intuitive sheet interface for browsing and assigning prompts to agents with folder organization and tag filtering directly from the agent configuration screen
• Quick Agent Duplication: Copy existing agents with one click to accelerate agent creation using the new <entity>CopyOneById GraphQL mutation
• Increased Agent Tool Execution Steps: Extended default agent step count from 2 to 5 across chat and function calling flows for more complex reasoning tasks
• Enhanced Tool Configuration: Collapsible categories and search functionality in tool selection
• Separated Instructions: Clear distinction between system instructions and custom instructions, now fully exposed in GraphQL schema and shown in the UI if provided

​

Knowledge Base Data Processing Architecture

• Unified Items Filtering: New centralized items-filter.tsx component provides consistent filtering experience across items list view, embedding job trigger creation, and processor job filter creation with enhanced filter options and real-time result preview
• Item-level Processing: Migrated from field-level to item-level processors (ExuluContextFieldProcessor → ExuluContextProcessor) to reduce complexity
• Bulk Operations: Process multiple items at once with new ProcessItems GraphQL mutation supporting filter and limit parameters
• Processor Visibility: New processors view in the knowledge base section showing configuration details, queue status, and execution history
• Timestamp Tracking: Added last_processed_at field on items for better processing visibility
• Queue Management Integration: Direct access to queue controls from processor configuration with configurable timeouts (default: 180s, up to 30 minutes for long operations)
• Enhanced Concurrency: Separate worker and queue-level concurrency settings for granular control over job processing
• Improved Database Pool: Increased connection pool from 20 to 50 connections with statement timeouts to handle concurrent processor jobs
• Automatic JSON Handling: JSON fields now automatically serialize during create and update operations

​

Project Context Integration

• Project-aware Sessions: Chat sessions can be associated with projects for contextual tool access with automatic project tool injection
• Project-scoped Retrieval: New createProjectRetrievalTool enables project-specific information search across multiple contexts with hybrid search and filtering
• Enhanced Metadata: Project cards display creation date, character count, and chunk information
• Automatic Tool Ingestion: Agents automatically gain access to project-specific retrieval tools when sessions have associated projects
• Project Caching: Optimized repeated project queries with intelligent caching mechanisms

​

Vector Search & Context System

• Adaptive Filtering: Automatic threshold filtering (60-70% of best match score) removes low-quality results
• Chunk Expansion: Configurable option to retrieve surrounding chunks (before/after) for better context continuity
• Enhanced Hybrid Search: Improved RRF weighting and scoring with better cosine distance, tsvector, and hybrid cutoff thresholds
• Increased Capacity: Maximum retrieval limit increased from 50 to 250 results
• Improved Text Search: tsvector search now uses OR logic for better partial matching of technical terms
• Archive Filtering: Archived items automatically excluded from all search methods
• Context Metadata: New chunks_count and embeddings_updated_at fields for better visibility
• Chunk Management: Added limit parameters to generateChunks and deleteChunks mutations with proper counting support
• Direct Search API: Added search method directly to ExuluContext class for easier integration

​

Access Control & Security

• Refined RBAC Logic: More explicit and maintainable role-based access control with improved conditionals
• Non-admin Chunk Limits: Non-super-admins can generate/delete chunks with configurable limits for safety
• User-scoped File Access: Proper folder prefixing and user-scoped access control in file upload routes
• Global File Management: Enhanced download access control to support both user-scoped and global file access
• Record Privacy Reset: Copied records automatically reset to private rights_mode with new timestamps
• Improved Authentication: Better handling of unauthenticated requests with proper fallback behavior
• Context-specific Validation: Enhanced role validation for context-specific operations

​

Technical Improvements

​

Performance & Architecture

• Next.js 16 Upgrade: Migrated entire application to Next.js 16 for latest features and performance
• GraphiQL v4: Updated to latest GraphiQL version with improved developer experience
• Server-side Data Fetching: New fetchGraphQLServerSide moves certain fetch operations to server components
• Optimistic Updates: Better perceived performance with optimistic UI updates throughout
• Query Optimization: Full-text search now uses ts_rank instead of ts_rank_cd for better performance
• GraphQL Filter Extensions: Added lte and gte operators for Float filter types
• Eval Worker Scaling: Increased eval runs queue concurrency from 1 to 10 workers for faster evaluation processing

​

Developer Experience

• Toast Notifications: Standardized on Sonner library for consistent notification styling
• Type Safety: Enhanced TypeScript definitions across agent sessions, items, and workflows
• Better Error Handling: Improved API error responses and user-facing error states with proper null handling in job results
• S3 Integration: Simplified bucket handling with slash-based key notation (bucket/key) for better consistency
• Global File Uploads: New global file upload support outside user directories with enhanced access control (used primarily for agent visuals)

​

UI/UX Polish

• Navigation Improvements: Streamlined navigation with better visual hierarchy
• Layout Refinements: Fixed numerous spacing and alignment issues across pages
• Status Indicators: Consistent status badges and tooltips for better information density

​

Bug Fixes & Refinements

​

Minor Fixes

• Resolved duplicate todo list message rendering
• Fixed command item disabled styling and duplicate keys in tool lists
• Corrected search results layout from multi-column to single-column for readability
• Added automatic data refresh after chunk deletion
• Fixed missing spaces in chat session layout
• Improved retry functionality for embeddings and processor jobs
• Fixed hasNextPage logic in pagination queries
• Fixed access control to properly handle unauthenticated requests with super_admin bypass
• Corrected S3 key prefix ordering (user prefix before general prefix) and proper bucket name handling
• Fixed file deletion by properly stripping bucket prefix
• Fixed created_by type handling (string vs number) in access control checks
• Fixed eval run deletion to properly clean up associated job_results entries
• Improved chunk deletion safety by checking table existence before deletion
• Fixed image handling in chat API for proper multimodal support

​

Cleanup & Deprecations

• Removed deprecated jobs-status-area component
• Removed obsolete chat-empty-state component
• Removed deprecated agent-nav component
• Consolidated redundant layout files
• Removed recent-jobs component in favor of unified queue management
• Removed deprecated ExuluDefaultContexts export
• Deleted unused template contexts (code-standards.ts, outputs.ts)
• Removed hardcoded context registrations from ExuluApp
• Removed excessive debug console.log statements from auth, postgres, and redis clients
• Cleaned up redundant code across multiple modules

​

What’s Next

• Advanced workflow conditions and branching
• Enhanced eval capabilities with more flexible test case management
• Real-time collaboration features for shared agents and workflows
• Extended monitoring and observability for production deployments