Version 7.12.0

Minor release: Ollama support for agents. Agents built in BESSER and the Web Modeling Editor can now use locally hosted Ollama models — both as the LLM provider and as the RAG embedding provider — end to end, from the editor’s agent-configuration UI through the BAF generator and the agent converters. Additive only: no metamodel or API contract changes beyond the new provider.

Highlights

  • Ollama as an agent LLM provider: a new LLMOllama wrapper joins the supported LLM suite (alongside OpenAI, Hugging Face and Replicate). In the Web Modeling Editor you can now pick Ollama (local) for an agent LLM and set its base_url (default http://localhost:11434) and model. The BAF generator imports LLMOllama and emits an nlp.ollama.base_url entry in the generated agent config, and the agent template instantiates it automatically.

  • Ollama as a RAG embedding provider: RAG elements can now choose OpenAI or Ollama for embeddings, with dedicated base-URL and model fields in the editor. The generated agent wires up OllamaEmbeddings for Ollama-backed vector stores (importing only the embedding classes it actually uses). The choice round-trips through the agent converters (json_to_buml ↔ buml_to_json), covered by a new round-trip test.

  • Docs: the agent-personalization reference now documents the ollama provider, its ollamaModels recommendations, and the local base_url / model configuration.

Also in this release

  • B-UML export round-trip (backend): the multi-object exporter now writes a single numbered object-model banner, so re-importing a project that contains several object models yields exactly those models instead of extra empty ones (web editor issue #161 follow-up).

  • BPMN pools and lanes (frontend, via the submodule bump): elements inside a BPMN lane/pool now move with their parent, and pools/lanes are preserved when the assistant injects a BPMN diagram (WME #159).