The technology behind thalabus.
For IT architects and AI specialists: how thalabus connects search, knowledge, agents and language models to deliver reliable results. All modules are developed in Switzerland for thalabus and aligned with a common architecture. The core is open source. The core is open source.
Multi-Agenten, Knowledge-Graph-RAG, lokale LLMs.
thalabus orchestrates AI solutions within your IT landscape, on-premises or in your private cloud. A task planner forms the core: it breaks complex tasks down into sub-steps that specialized agents solve reliably, even with small language models.
- Operation
- On-premises, private cloud, local cloud provider or your own machine. Cloud models available upon approval.
- Orchestration
- thalaplan: goal extraction, decomposition into sub-steps, prompt optimization, parallel agents and tools. Deep reasoning with template-based thinking patterns.
- Knowledge
- Complex RAG: content and relationships in ontologies and graph databases, with exact source verification.
- Data access
- Protected data remains in your source systems and is queried directly as needed, respecting the source system permissions.
- Models
- Small LLMs for German, French and Italian, with fine-tuning on demand.
- Quality
- Automatische Prüfung aller Zwischenergebnisse, 99+% correct results. Clean code, 100% testable. Multilingual regression test framework.
- Interfaces
RESTPython-SDKMCP, expandable agent and skill library, extension points.
Three layers, one end-to-end process.
Your business process remains in control. thaladeep breaks a process step down into stages and works through them in cycles. thalaplan assigns the tasks of each stage to agents working with skills and tools.
- Configuration
- Each use case is configured with structured plain-text documents for a clearly defined purpose.
- Workflow
- Each stage passes through a status cycle from open to in progress, summary, review and completed, across multiple cycles if needed.
- Workspace
- Dossier, incoming documents, notepad and output documents for every case.
- Quality
- Automatic prevention of hallucinations, with every statement traceable back to its source.
- User Interface
- Optionally a dedicated user interface for each use case, along with PDF reports.
- Integration
- Optionally connected to external BPM applications and your backends.
thalaplan, the planning engine at the core of thalabus.
thalaplan works with customizable plan templates: it generates plans autonomously, configures template-based plans and executes customer-specific plans. It breaks down every goal into precise sub-steps that even small local models handle reliably. Specialized agents work in parallel, using tools and skills ranging from search and OCR to database queries, APIs and MCP. Every intermediate result is automatically checked for accuracy and completeness, and the process ends with a sourced result that your specialist approves.
Multiple search methods, one verified dataset.
thalabus combines semantic, category, multilingual full-text and knowledge graph search, re-ranking all hits together. The result is a clean dataset that users can save and reuse across chats and tools.
- Search
- Semantic search using embeddings, deterministic category indexes, multilingual full-text search (BM25), knowledge graph search (GraphRAG, PathRAG, multi-turn RAG) and query expressions that combine them all.
- Result
- Fast re-ranking across all search methods, relevance controls for users, datasets ready to save and reuse.
- Ontology
- Fully customizable with entities, relationships, attributes, categories and vocabularies. Import and export in OWL, Turtle and JSON-LD.
- Knowledge Graph
- Automatic extraction from documents, Confluence and intranet sources. Graph visualization with direct links to sources.
- Consistency
- Source consolidation and automated detection of contradictions.
The right model for every requirement.
thalabus is optimized for small, local open-weight models running directly in your data center. Where scale matters and data protection policies allow, it works just as effectively with models from hyperscalers.
- Local models
- Open-weight models in your local data center for maximum confidentiality, featuring models optimized for German, French and Italian.
- Cloud models
- Frontier and cost-effective cloud models for maximum scalability, enabled upon your approval.
- Knowledge distillation
- Automated training of smaller models based on thalabus instructions, domain instructions and user feedback.
- Isolation
- Every chat and tool session runs isolated inside its own private container.
- Access control
- Access control lists for information, knowledge graphs, dossiers and documents: private, shared with groups or published to all. Anonymous users only see public content.
Small models, continuously monitored and corrected.
Proprietary tools get the most out of small on-premises models. The de²generator sits between agents and the language model, monitoring the token stream and resolving errors as they occur.
- de²generator
- Validates the token stream between agents and inference, correcting errors on the fly or re-querying the model.
- Tests
- Automated, multilingual regression testing. Lean code, 100% testable, 99+% correct results.
- Model Trainer
- Automated knowledge distillation for fine-tuning smaller models.
- Prompt Optimizer
- Automatically writes and refines agent prompts.
- Log Analyser
- Automated error analysis with AI-assisted resolution.
An open core, a consistent architecture.
The core of thalabus is open source. All modules, from RAG, planning and deep reasoning to the knowledge graph, are developed in Switzerland for thalabus and aligned with a shared architecture and roadmap.
- Interfaces
RESTPython-SDKfor custom applications using AI services from thalabus.- Extensions
- Custom agents, skills and external
MCP-servers, widgets for chat and tools, additional tools for users. - Integration
- Integration with case management and BPM systems, databases, document repositories and intranet environments.
- Open Source
- The core is open source. Specific tools and use case templates are part of the solutions from t2b.
- In-house development
- RAG, planning, orchestration, deep reasoning, ontology and knowledge graph, vendor-independent inference, prompt management, container orchestration with auto-scaling, model tester and trainer, regression testing.
Let us find your best AI use case.
In a use case workshop, we collaborate with your business team and IT department to build a solid project proposal for your management.
