Installation¶

LongCache requires Python 3.10 or newer.

Installing via uv (Recommended)¶

uv provides instant, reproducible installations:

# Core package (SQLite backend + FAISS + BGE-small local embedder)
uv pip install longcache

# With LangChain & LangGraph
uv pip install "longcache[langchain]"

# With Redis for multi-worker production
uv pip install "longcache[redis]"

# With LlamaIndex
uv pip install "longcache[llamaindex]"

# With OpenAI official client wrapper
uv pip install "longcache[openai]"

# With FastAPI / Starlette middleware
uv pip install "longcache[server]"

# Full installation (all integrations)
uv pip install "longcache[all]"

Installing via pip¶

# Standard installation
pip install longcache

# Full bundle
pip install "longcache[all]"

CPU PyTorch Optimization

LongCache uses PyTorch purely for local embedding inference with sentence-transformers. In production environments without GPUs, you can install the lightweight CPU-only wheel of PyTorch (~180MB instead of ~3GB):

pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install longcache

Offline Usage & First Run¶

On the very first run, LongCache automatically downloads the compact BAAI/bge-small-en-v1.5 model weights (~130MB) into your local Hugging Face cache (~/.cache/huggingface/hub).

Once downloaded, LongCache works 100% offline without any outgoing network connections required for embedding queries.

To pre-warm the model in container builds (e.g. Dockerfile):

python -c "from longcache.core.embedder import LocalEmbedder; LocalEmbedder().embed('warmup')"