v0.3.6

What's Changed (since v0.3.0)

Features

• OpenAI SDK backend: Added direct OpenAI SDK support (openai_client.py) with AsyncOpenAI/AsyncAzureOpenAI clients (d3ea4e8)
• Microsoft Fabric support: Auto-detect Fabric platform and use synapse.ml.fabric.credentials for authentication (95f115a)
• Azure OpenAI support: Detect Azure endpoints via �pi_version, env vars, or URL and use AsyncAzureOpenAI with proper config (f41cb0b)
• GitHub Copilot as default backend: Changed default LLM backend from openai to copilot across all components (3c263b6)
• Demo video generator: Combined video pipeline with HTML slides + notebook walkthrough + narrations (1bf5d1e)
• Backend parameter: Added �ackend parameter to FeatureExplainer and FeatureCodeGenerator (defb252)

Bug Fixes

• Fabric credential auth: Use get_openai_httpx_async_client() to avoid placeholder API key 401 errors on Fabric (95f115a)
• Azure api_version: Set default �pi_version=2024-10-21 for Fabric Azure client (bf2fe04)
• Duplicate logging: Prevent duplicate log output by disabling propagation to root logger (c97f11a)
• Backend propagation: Pass �ackend, �pi_key, �pi_base, �pi_version from llm_config through AutoFeatureEngineer to SemanticEngine (6ba724e)
• Token parameter: Use max_completion_tokens instead of deprecated max_tokens for newer models (9c56697)
• sklearn compatibility: Fix sklearn error in AutoFeatureEngineer (c0b0145)

Chores

• Bump version to 0.3.6
• Update .gitignore for generated video files (eed8420)
• Update gitignore (12e3269)

Full Changelog: v0.3.0...v0.3.6

v0.3.0

Highlights

• Expanded benchmark coverage with multi-framework support (FLAML, H2O, AutoGluon) and feature caching.
• Added a unified category-based dataset API and reusable natural-language transform rules.
• Refreshed benchmark docs and examples with corrected metrics and new results.

Benchmarks & datasets

• Added simple_models LLM benchmark results (52 datasets) and expanded engine coverage.
• Included forecasting datasets and standardized report naming/formatting.
• Shared feature cache across benchmarks with consistent cache versioning.

Fixes & maintenance

• Hardened benchmark tool handling and fixed cache key issues.
• Updated benchmark dependencies (numpy pin, autogluon).
• Removed the FLAML 90s benchmark.

Docs & examples

• Added FLAML Spotify example.
• Updated benchmark highlights in docs.

Full Changelog: v0.2.0...v0.3.0

v0.2.0

New features in this release:

• LiteLLM integration supporting 100+ LLM providers
• Feast feature store integration for production deployment
• Comprehensive demo notebook with visualizations
• HTML presentation slides for demos
• Improved async handling for Jupyter notebooks
• Enhanced documentation with e2e example

Full Changelog: v0.1.0...v0.2.0

v0.1.0

Features

• Multi-engine architecture: tabular, time series, relational, and text feature engines
• LLM-powered intelligence via GitHub Copilot SDK
• Intelligent feature selection with statistical testing
• Sklearn-compatible transformers for ML pipelines

Full Changelog: https://github.com/thinkall/featcopilot/commits/v0.1.0