v1.2.0

A2rchi v1.2.0 Release Notes

Overview

This major release includes approximately 200+ commits since v1.1.0, bringing significant architectural improvements, new features, enhanced configurability, and better developer experience to A2RCHI.

Major Features

Multi-Configuration Support

• Multiple Prompt Configurations: Users can now run A2rchi with multiple configuration files simultaneously, allowing different prompt strategies for different use cases
• Dynamic Configuration Switching: New UI dropdown allows switching between different prompt configurations on the fly
• API Endpoint for Prompts: Added /prompts endpoint to retrieve available prompt configurations

Enhanced Data Sources & Scrapers

• Git Repository Scraper: New scraper for ingesting documentation directly from Git repositories, including support for MkDocs sites
  • Authentication support with username and personal access tokens
  • Automatic detection and handling of MkDocs repositories
  • Configurable via git_username and git_token secrets
• Sources Registry System: Implemented centralized sources registry for better management of data ingestion pipelines
• SSO Scraper Improvements: Enhanced SSO scraping with better recursion handling and URL tracking in vector database

Retrieval Enhancements

• Hybrid Search with BM25: Introduced hybrid retrieval combining semantic search with BM25 keyword matching for improved accuracy
  • Configurable hybrid retriever settings
  • Better handling of keyword-based queries
• Document Stemming: Optional stemming of documents before embedding creation for improved retrieval accuracy
  • Configurable stemming for both documents and queries
  • Particularly useful for technical documentation
• Embedding Options: Enhanced embedding configuration including:
  • Distance metric selection
  • Custom embedding instructions
  • Embedding model selection improvements

New LLM Integrations

• Ollama Support: Full integration with Ollama for local model inference
  • Configurable GPU allocation (num_gpu: -1 for all available GPUs)
  • Support for various Ollama models
• vLLM Improvements: Enhanced vLLM integration for faster inference
• Improved HuggingFace Support: Better handling of HuggingFace models for both inference and evaluation

Improvements

Architecture & Code Organization

• Chain/Pipeline Abstraction: Major refactoring of the chain system
  • Introduced BasePipeline for generalized LLM and prompt initialization
  • Support for multiple pipelines running simultaneously
  • Cleaner separation between chains, workflows, and wrappers
  • Created dedicated chains.py module
• Configuration Structure Overhaul: Completely restructured configuration file format
  • More intuitive hierarchy
  • Better validation and error handling
  • Support for pipelines (plural) in configuration

CLI Enhancements

• New CLI Implementation: Rebuilt CLI with improved functionality
  • --config flag to specify configuration files
  • -d/--dry-run mode for testing configurations without deployment
  • -f/--force flag for a2rchi create to automatically delete existing deployment first
  • --print-config option to display the loaded configuration
  • -p flag added to a2rchi delete for proper cleanup
• Better Secret Handling: Improved management of API keys and passwords across services

Container & Deployment

• Base Images: Created optimized base Docker images
  • pytorch-base and python-base images now available on DockerHub
  • Significantly faster deployment times
  • GPU and non-GPU variants for different use cases
• Slimmer Images: Optional lightweight images without GPU dependencies when running API-based models
• Requirements Reorganization: Split requirements into multiple files for better dependency management
• CUDA Version Update: Fixed CUDA mismatch issues, now running CUDA 12.4
• OpenShift/OKD Support: Added proper permissions and configurations for Kubernetes deployments
• Health Check Probes: Implemented health check endpoints for container orchestration

Network & API

• Host Mode Support: Fixed and improved host networking mode
  • Grafana works correctly in host mode
  • ChromaDB respects chromadb_external_port configuration in host mode
• Same-Origin API Calls: Optimized API calls from the frontend to avoid unnecessary host/port additions
• HTTPS Support: Frontend can now communicate with HTTPS APIs
• ChromaDB API Endpoints: Added REST API for ChromaDB operations
  • Document listing endpoint
  • Document search endpoint
  • Configurable enable/disable option
  • Comprehensive API documentation in user guide

Developer Experience

• GitHub Actions CI/CD:
  • Added smoke tests for PR validation
  • PR preview environment automation
  • Automated testing pipeline
• Logging Improvements:
  • Better structured logging across all containers

Interface Improvements

• Redmine Enhancements:
  • Bug fixes for Redmine mailer
  • Ticket client improvements
  • Better Postgres integration
• Grafana Monitoring:
  • Added retrieval scores to Grafana dashboards
  • Better visualization of context and history
  • Timeout and batch size configurations

Benchmarking & Evaluation

• Benchmarking Framework: New comprehensive benchmarking functionality
  • Support for multiple evaluation LLMs and providers
  • HuggingFace evaluation model support
  • Better default evaluation models
  • Configurable timeout, embedding model, verbosity, and batch_size
  • Queries configuration via queries.json
  • Plotting dependencies and visualization tools
  • Dedicated benchmarking documentation

Documentation

• Complete Documentation Overhaul: Comprehensive rewrite of user and developer guides
• New User Guide Sections:
  • ChromaDB API endpoints documentation
  • Hybrid search documentation
  • Stemming and Ollama interface documentation
  • Git scraper setup and usage
  • Benchmarking guide
  • Vector store configuration
• Developer Guide Updates: Enhanced developer documentation with architectural diagrams
• README Improvements: Updated README with new logo, clearer instructions, and current examples
• API Documentation: Complete API endpoint documentation
• Configuration Examples: Added example configurations for common use cases

Infrastructure Changes

• Directory Restructure: Major reorganization of src/data_manager/ and related directories
• Requirements Management: Split into multiple organized requirements files
• Automated Image Publishing: Script to push new Docker images to registry
• MkDocs Material: Integration with MkDocs Material for enhanced documentation
• Firefox GPU Support: Fixed Firefox compatibility for GPU-accelerated instances
• .gitignore Updates: Proper handling of .env files and .github workflows

Configuration Changes

New Configuration Options

• chromadb_external_port: Configure external port for ChromaDB in host mode
• enable_chromadb_api: Toggle ChromaDB API endpoints
• hybrid_search: Enable/disable hybrid retrieval
• stemming: Configure document and query stemming
• distance_metric: Choose embedding distance metric
• embedding_instructions: Custom instructions for embedding models
• ollama_num_gpu: GPU allocation for Ollama models
• timeout, batch_size, verbosity: Benchmarking configurations
• pipelines: Support for multiple pipeline configurations (replaces single chain)

Breaking Changes

• Configuration structure has been significantly refactored
• chain configuration section renamed to pipelines (plural)
• Some configuration keys have been reorganized into new hierarchies
• Old configuration files will need to be migrated to new format

Testing & Quality

• Added smoke tests for core functionality
• Implemented GitHub Actions for automated testing
• PR preview environments for testing changes
• Improved error handling throughout codebase
• Better validation of configuration files

Resources

• [Documentation](https://mit-submit.github.io/A2rchi/)
• [User Guide](https://mit-submit.github.io/A2rchi/user_guide/)
• [GitHub Repository](https://github.com/mit-submit/A2rchi)
• [Issue Tracker](https://github.com/mit-submit/A2rchi/issues)

---

Full Changelog: v1.1.0...v1.2.0