Changelog

Unreleased

Migrated research and deep QA agents to use Pydantic Graph beta API for better graph execution
Automatic semaphore-based concurrency control for parallel sub-question processing
max_concurrency parameter for controlling parallel execution in research and deep QA (default: 1)

BREAKING: Research and Deep QA graphs now use pydantic_graph.beta instead of the class-based graph implementation
Refactored graph common patterns into graph_common module
Sub-questions now process using .map() for true parallel execution
Improved graph structure with cleaner node definitions and flow control

New haiku.rag-slim package with minimal dependencies for users who want to install only what they need
Evaluations package (haiku.rag-evals) for internal benchmarking and testing
Improved search filtering performance by using pandas DataFrames for joins instead of SQL WHERE IN clauses

BREAKING: Restructured project into UV workspace with three packages:
- haiku.rag-slim - Core package with minimal dependencies
- haiku.rag - Full package with all extras (recommended for most users)
- haiku.rag-evals - Internal benchmarking and evaluation tools
Migrated from pydantic-ai to pydantic-ai-slim with extras system
Docling is now an optional dependency (install with haiku.rag-slim[docling])
Package metadata checks now use haiku.rag-slim (always present) instead of haiku.rag
Docker image optimized: removed evaluations package, reducing installed packages from 307 to 259
Improved vector search performance through optimized score normalization

Support for Zero Entropy reranker
Filter parameter to search() for filtering documents before search
Filter parameter to CLI search command
Filter parameter to CLI list command for filtering document listings
Config option to pass custom configuration files to evaluation commands
Document filtering now respects configured include/exclude patterns when using add-src with directories
Max retries to insight_agent when producing structured output