Changelog 

Features 

Retriever steps can now contribute multiple, separately named retrieval groups to one pipeline run. The retriever-step title names its retrieval group and is visible in the LLM prompt, so instructions can distinguish between sources. Retriever-step titles must therefore be unique within a pipeline. Every retriever appends its named group, preventing later retrievers from implicitly discarding previous results. Chunk and character limits apply to each retrieval group independently.

Vector store connections can define a default collection and further allowed collections. Assistant profiles provide a default vector store connection. Every retriever step can override that connection and select one of the effective connection's collections explicitly. Without overrides, the profile connection and its default collection are used.

The backend Diagnostics area can validate complete assistant profiles. It checks pipeline configuration, the AI connection and every effective vector store and collection used by the default vector retriever. Remote collection checks are read-only and do not execute the chat pipeline.

Breaking changes 

The collection field on assistant profiles has been removed. Move a profile's existing value to the vector store connection's default_collection or to the relevant retriever step's retrieval_collection before updating the database schema.

The retrieval state in Madj2k\\AiCore is group-based only. The former flat result, raw-result and processor-identifier accessors have been removed. Stored retrieval session snapshots from earlier development versions are not reused. Normal conversation messages are unaffected.

PipelineStepConfigurationInterface now includes vector store connection and collection accessors. The existing step title is the prompt-visible retrieval name. Retrieval processors always append their named group. VectorStoreConnectionConfigurationInterface now includes getCollectionList(). Internal implementations have been updated to these contracts.

Database migration 

The pipeline-step fields retrieval_vector_store_connection and retrieval_collection and the vector-store-connection field collections have been added.

The former top_k field was renamed to max_retrieval_results. Existing values must be copied before TYPO3 removes the old field:

UPDATE tx_aiassistant_assistant_pipeline_step
SET max_retrieval_results = top_k;

The separate frontend_source_fields setting was intentionally consolidated into prompt_metadata_fields. The latter now controls both prompt metadata and frontend source output, so no value migration is required for frontend_source_fields.

Existing retriever steps must receive meaningful, unique titles as part of the internal pipeline configuration.

Breaking changes 

The temporary Madj2k\\AiAssistant aliases for framework-independent assistant, connection and indexing classes have been removed. Extensions must use the corresponding Madj2k\\AiCore classes and interfaces directly. TYPO3-specific models, repositories, controllers and integration services remain in the Madj2k\\AiAssistant namespace.