Skip to content

v1.9.0

v1.9.0 — hierarchical Supervisor + docstring styles + memory scoring
@UpendraBhandari UpendraBhandari tagged this 10 May 00:34 
Three small, additive polish items:

- Supervisor(validate_outputs=True): manager LLM reviews each worker
  output before accepting; on rejection the worker re-runs once with
  feedback. Logs guardrail_events row on exhausted retries.
- Tool docstring parsing: NumPy and Sphinx (reST) styles in addition
  to Google. Detection order Google -> NumPy -> Sphinx.
- VectorBlock + PersistentFactBlock memory scoring: optional
  recency_weight + importance_weight kwargs. Defaults preserve today's
  similarity-only retrieval. PersistentFactBlock importance sourced
  from learned_memory.confidence.

Breaking changes: None.
Tests: 1488 passed, 3 skipped (+31 new).

PyPI: https://pypi.org/project/fastaiagent/1.9.0/
PR:   https://github.com/fastaifoundry/fastaiagent-sdk/pull/69
Assets 2
Loading