omnicache-ai v0.2.0

What's New in v0.2.0

Bug Fixes

• FAISSBackend.delete(): switched IndexFlatIP → IndexIDMap2 — deleted vectors now actually removed (previously caused stale semantic matches)
• EvictionPolicy: was exported but never used; now wired into CacheManager and InMemoryBackend
• LangChainCacheAdapter.clear(): was a silent no-op; now calls manager.clear()

New Features

• CacheMetrics: hit/miss/eviction/set counters + hit_rate — access via manager.metrics
• Pluggable Serializer: PickleSerializer (default) + JsonSerializer; all layers accept serializer=
• StampedeShield: per-key lock in ResponseCache.get_or_generate() — prevents duplicate LLM calls under concurrency
• TieredBackend: L1 memory + L2 Redis/disk with automatic promotion on read
• AsyncCacheBackend + AsyncInMemoryBackend: native async backends for FastAPI / async LangGraph
• OpenAICacheAdapter: wraps client.chat.completions.create (sync + async)
• AnthropicCacheAdapter: wraps client.messages.create (sync + async)
• GzipCompressor: optional compression wired into CacheManager
• StreamingResponseCache: buffers streaming LLM output, replays from cache as generator
• CLI: python -m omnicache_ai stats|flush|inspect <key>
• pre-commit: ruff + mypy + file checks on every commit
• Dockerfile: slim Python 3.12 image

Install

pip install omnicache-ai==0.2.0

See CHANGELOG for full details.