Recallm

Semantic caching for LLMs. Ask once, recall forever.

Exact-match caching is useless for LLMs — two users asking the same question in slightly different words both pay the full API cost. Recallm uses sentence embeddings to find near-matches and return cached responses instantly. The result: lower API costs, reduced latency, and no changes to your existing LLM client code.

Install

pip install recallm
pip install "recallm[redis]"   # persistent cache, shared across workers
pip install "recallm[torch]"   # sentence-transformers embedder (700MB, PyTorch)

Once installed, import directly from the recallm package:

from recallm import SemanticCache, CacheConfig, InMemoryStorage

Quickstart

pip install recallm

from recallm import CacheConfig, InMemoryStorage, SemanticCache

storage = InMemoryStorage()
cache = SemanticCache(storage=storage, config=CacheConfig(threshold="balanced"))

def fake_llm(**kwargs):
    return {"id": "resp-1", "choices": [{"message": {"content": "Paris"}}]}

cached = cache.wrap(fake_llm, mode="sync")

request = {
    "model": "gpt-4o-mini",
    "messages": [{"role": "user", "content": "What is the capital of France?"}],
    "cache_context": {"user_id": "u-1", "document_id": "geo-v1"},
}

first = cached(**request)   # miss: calls fake_llm and stores response
second = cached(**request)  # hit: returns cached response
print(first["choices"][0]["message"]["content"], second["choices"][0]["message"]["content"])

Debugging

Inspect cache behaviour during development with cache.stats():

stats = cache.stats()
print(stats.hit_rate)         # fraction of requests served from cache
print(stats.hits, stats.misses)
print(stats.avg_similarity)   # mean cosine similarity of cache hits
print(stats.namespace_sizes)  # entry counts per namespace

stats() returns a CacheStats dataclass and is intended for development and debugging. Use the Prometheus metrics for production observability.

Deployment note: SemanticCache(...) loads the embedding model synchronously. In async frameworks (FastAPI, etc.), use await cache.async_warmup() during startup instead of relying on the constructor — see getting started.

Known limitations

• stream=True bypasses the cache entirely
• Redis backend is not suitable for namespaces > 5,000 entries without partitioning
• Sync callers using RedisStorage have no timeout protection (v0.1.0)

Full docs · Contributing · MIT License