Skip to content

Changelog

The full, authoritative changelog lives in CHANGELOG.md in the repository. Highlights below.

0.2.0 — Phase 2: semantic cache

Backward compatible — all Phase 1 exact-cache and decorator APIs are unchanged, and every new capability is off by default.

  • Semantic cache — opt-in embedding-based similarity matching so reworded but equivalent requests reuse a cached result instead of triggering a new LLM call.
  • MemoryCache.lookup / MemoryCache.store — combined exact-first, then semantic flow; an exact hit never computes an embedding.
  • Three embedding providers (local, openai, litellm) behind a common interface, selected by config.
  • In-memory and Redis vector indexes with namespace isolation and expiry safety.
  • extract_answer — return only the best-matching sentence(s) of a string answer on a semantic hit (extractive, no LLM).
  • CacheStats gains exact_hits / semantic_hits.
  • New extras: semantic (API embeddings, no torch) and semantic-local (local embeddings).

0.1.0 — Phase 1: exact caching

First public release: MemoryCache, exact and tool caches, in-memory and Redis backends, multi-scope isolation, LangGraph + LiteLLM integrations, and shipped type hints.