v2.0.0 - Major Upgrade

BESSER Modeling Agent v2.0.0

Major upgrade with 2 new diagram types, voice input, file conversion, code generation triggering, multi-step orchestration, deterministic layout engine, and a complete architecture overhaul.

New Diagram Types

• GUI No-Code Diagram: Metric cards, action buttons, charts, navigation headers, sample data generation
• Quantum Circuit Diagram: Gate symbols, qubit management, circuit operations

New Capabilities

• Voice input via OpenAI Whisper speech-to-text
• File conversion: PlantUML, RDF/OWL/TTL, images, PDFs, and generic text converted to diagrams
• Code generation triggering for 15+ BESSER generators (Django, Python, Java, SQL, WebApp, Qiskit, etc.) plus export and deploy
• Multi-step orchestration: compound requests decomposed across diagram types and generators
• Deterministic layout engine (2,134 lines) with collision avoidance, replaces random positioning
• Conversation memory with sliding-window and optional LLM summarization
• Context-aware suggestions after each action based on project state
• Confirmation flows before replacing existing diagrams (replace/keep/new tab)
• "Describe my diagram" intent for explaining canvas contents
• Token and cost tracking per session and globally
• Domain pattern library with pre-built templates for common system types

Enhanced Existing Capabilities

• Model modification expanded from Class-only to all 6 diagram types
• Intent recognition overhauled with pattern-based cross-validation (9 intents, up from 6)
• Pydantic structured outputs for all diagram types via beta.chat.completions.parse
• Architecture restructured from flat files into modular src/ package

Tech Stack Changes

Component	v1.0.0	v2.0.0
Python	3.11+	3.10+
BAF	v4.2.0	v4.3.1
LLM	GPT-4o-mini (temp 0.4, 3K tokens)	gpt-4.1-mini (temp 0.2, 8K tokens)
Structured output	Manual JSON prompting	Pydantic via beta.chat.completions.parse
Speech	--	OpenAI Whisper
Layout	Random offset	Deterministic engine with collision avoidance

Testing

16 test files (up from 3) covering all diagram handlers, schemas, memory, tracking, file conversion, generation, protocol, orchestrator, suggestions, and confirmation flows.

Breaking Changes

• Architecture: flat files replaced by src/ package (all imports changed)
• WebSocket protocol upgraded to v2.0 (AssistantRequest dataclass, workspace/widget modes)
• LLM model changed from GPT-4o-mini to gpt-4.1-mini
• BAF upgraded from v4.2.0 to v4.3.1
• config.ini replaced by config.yaml

v1.0.0 - Initial Release

BESSER Modeling Agent v1.0.0

Initial release of the BESSER Modeling Agent -- the conversational AI backend for the BESSER Web Modeling Editor.

Features

• 4 diagram types: Class, Object, State Machine, and Agent
• Natural-language diagram creation: single elements or complete systems from a description
• Model modification for Class diagrams (add, remove, rename elements)
• UML specification Q&A via RAG (ChromaDB + LangChain over official UML 2.5 spec)
• 6 intents: greeting, create single element, create complete system, modify model, modeling help, UML spec queries
• WebSocket real-time communication (port 8765)

Architecture

• Built on BESSER Agentic Framework (BAF) v4.2.0
• LLM: OpenAI GPT-4o-mini (temperature 0.4, max_tokens 3000)
• 7 agent states with unified transitions supporting text and JSON events
• 4 diagram handlers: ClassDiagramHandler, AgentDiagramHandler, ObjectDiagramHandler, StateMachineHandler
• Docker deployment with health check

Testing

• 3 test files with full BESSER framework stubs (runs without OpenAI keys)