Source: Redis LangCache docs (overview and API/SDK examples)
Redis's managed semantic-caching service: a search-then-store REST API with automatic embeddings, similarity thresholds, attribute scoping, exact/semantic search strategies and attribute-based deletion; in preview as of late 2026.
Contents
- What: official product documentation (“Redis LangCache”, “Use the LangCache API and SDK”)
- Publisher: Redis
- Date: accessed 2026-11-14 (living docs)
- Original: LangCache overview, API and SDK examples
Summary
LangCache is a hosted semantic cache. The application calls a search endpoint before the LLM; on a miss it calls the LLM and then calls a store endpoint. LangCache generates the embeddings itself, so the application never manages an embedding model or vector index.
Key claims
- Two calls:
POST /v1/caches/{cacheId}/entries/search(prompt in, cached response or empty out) andPOST /v1/caches/{cacheId}/entries(prompt + response, optionalattributes). Auth is a Bearer API key; Python (langcache) and JS (@redis-ai/langcache) SDKs exist. - Search takes
similarityThreshold(examples use 0.9, a similarity, so higher is stricter), optionalattributesthat must match for a hit, andsearchStrategiesofexact(case-insensitive string match) and/orsemantic; semantic only is the default. - Invalidation: delete one entry by ID, delete all entries matching a set of attributes, or flush. A delete-by-attributes call with no attributes deletes the whole cache.
- Configurable thresholds, TTLs and eviction; hit-rate monitoring in the Redis Cloud console.
- Savings model offered by Redis: monthly output-token spend times hit rate, on the assumption that saved input tokens roughly pay for embedding and storage. Worked example: $120 of output spend at a 50% hit rate saves $60/month.
- Positioned for chatbots, RAG, caching intermediate agent steps, and as a shared cache behind an AI gateway.
- Status (as of 2026-11): public preview on Redis Cloud; self-managed Kubernetes deployment and a fully managed tier are in private preview. Redis states it does not use cached data for training.