Install

This covers running piramid serve, configuring it, and the HTTP API. config.example.yaml, attached to each release, lists every setting at its default.

BinaryCopy Anchor Link

Linux and macOS, x86_64 or arm64:

curl -fsSL https://piramiddb.com/install.sh | sh

The script installs piramid to ~/.local/bin. PIRAMID_INSTALL_DIR changes the directory and PIRAMID_VERSION picks a release tag. The binaries run models on the CPU. Windows builds and every archive are on the releases page.

CargoCopy Anchor Link

Rust 1.87 or newer. inference-candle runs models; add gpu-cuda for CUDA, which needs the CUDA toolkit at build time:

cargo install piramid --locked --features inference-candle
cargo install piramid --locked --features inference-candle,gpu-cuda

Without inference-candle the server stores and searches collections but cannot generate.

ModelCopy Anchor Link

Piramid runs Qwen2 and Qwen3 checkpoints (Qwen2.5 included). The model directory holds config.json, tokenizer.json, tokenizer_config.json and the .safetensors weights:

hf download Qwen/Qwen2.5-0.5B-Instruct --local-dir ./models/Qwen2.5-0.5B-Instruct

RunCopy Anchor Link

startup:
  embedding:
    provider: openai
    model: text-embedding-3-small

runtime:
  inference:
    enabled: true
    model_path: ./models/Qwen2.5-0.5B-Instruct
    kv_cache:
      max_bytes: 2147483648

export OPENAI_API_KEY=...
piramid serve --config piramid.yaml

The server listens on 127.0.0.1:6333. Configuration covers GPU settings and every other block; HTTP API and Python client show how to use it.