v0.5.0

What's New

Hybrid Search & Smart Injection

• SmartInjector (new default) — two-tier injection: pinned items always injected as "CORE RULES", normal items hybrid-searched per query
  as "RELEVANT KNOWLEDGE"
• Hybrid search — FTS5 keyword matching + embedding cosine similarity merged via Reciprocal Rank Fusion (RRF)
• Knowledge pinning — engine.knowledge.pin(id) / .unpin(id) to mark critical rules that are always injected

Async & Flexible Tracing

• @engine.trace now supports async functions natively
• input_extractor / output_extractor params for multi-param and structured-output agents
• auto_inject=True mode — skip manual get_knowledge() calls
• Streaming support via extractors

Evaluation & Validation

• Batch evaluation — engine.evaluate() with CSV import, run comparison, and regression detection
• Custom scorers — per-agent or per-eval-case scoring functions for domain-specific grading
• Statistical validation — A/B testing with 5%+ improvement required, zero regressions, 0.8 confidence
• engine.set_agent_runner() for fully automated learn cycles

Other

• 220 tests passing
• SQLite WAL mode + FTS5 full-text search on traces and knowledge
• Plugin system: all components swappable via Python Protocols

Install

pip install agentlearn