AIDB 7.5.0 release notes
========================

Released: 29 Jul 2026

AIDB 7.5.0 includes the following enhancements and bug fixes:

.. csv-table::
  :header: Type,Description
  :widths: 10,30
  :align: left
  :class: longtable

  Enhancement,"**Model validation on creation:**   `aidb.create_model()` now validates a model by default, running a small probe inference — downloading and loading local models, or sending a test request to external providers — so configuration and credential problems surface at creation time instead of on first use. Pass `validate => false` to skip it, or use the new `aidb.validate_model()` to check an existing model. See  :ref:`Integrating models<Integrating models>` ."
  Enhancement,"**Local LLM inference with llama.cpp:**   AIDB can now run GGUF-format models locally through the new `llamacpp_generate` and `llamacpp_embeddings` providers, alongside the existing Candle-based local models. Ships with 7 pre-registered default models (5 for embeddings, 2 for generation). See  :ref:`Local models<Local models>` ."
  Enhancement,"**Requeuing failed pipeline records:**   The new `aidb.requeue_pipeline_errors()` function marks record-level error log entries dirty for reprocessing on the pipeline's next run, without needing to touch source data. See  :ref:`Requeuing failed records<Requeuing failed records>` ."
  Enhancement,"**Live-tunable model thread limit:**   The new `aidb.max_threads` parameter caps CPU threads used for local model inference (Candle and llama.cpp), and can be changed live without a restart. See  :ref:`Threading limits<Threading limits>` ."
  Enhancement,"**I/O thread pool control:**   The new `aidb.max_io_threads` parameter caps the async I/O thread pool used for remote-model HTTP calls and volume/object-store access, live-tunable without a restart. See  :ref:`Threading limits<Threading limits>` ."
  Enhancement,"**llama.cpp native log control:**   The new `aidb.enable_llamacpp_logs` parameter turns on llama.cpp's own diagnostic logging in the server log, for troubleshooting model loading and inference issues. Off by default; requires a restart to take effect. See  :ref:`AI pipelines <Local models>` ."
  Enhancement,"**Clearer error for pipeline column/step type mismatches:**   `aidb.create_pipeline()` now rejects a source column whose SQL type doesn't match its configured step (for example, a `text` column feeding a `KnowledgeBase(Image)` step) with a specific error at creation time, instead of failing later with an opaque internal error."
  Change,"**Metrics collector is now opt-in:**   The background worker that materializes pipeline metrics into `aidb.pipeline_metrics_collection` is disabled by default. Enable it with the new `aidb.enable_background_metrics_collector` setting, instance-wide or per database. See  :ref:`Materialized metrics<Materialized metrics>` ."
  Change,"**`topk` must be a positive integer:**   `aidb.retrieve_key()` and `aidb.retrieve_text()` now reject `topk` values less than 1 with a clear error, instead of silently clamping to 1."
  Bug fix,"**Error log functions and Empty-source pipelines:**   `aidb.get_error_logs()`, `aidb.get_error_log_summary()`, and `aidb.get_all_error_summaries()` no longer raise an error for pipelines with an Empty source (such as Semantic KB) — they return empty results instead, and no longer abort results for other pipelines in the same call."