Installation¶

Mandol 0.1.0 targets Python >=3.12,<3.13 and Linux.

Published package¶

Install the current release from PyPI. Use the explicit pin when the package version must remain reproducible.

python -m pip install mandol
python -m pip install "mandol==0.1.0"

The package is an early public research release rather than a production service, and its API may evolve during the 0.x series. Exact reproduction of the paper tables requires the frozen paper-repro source checkout and its benchmark-specific instructions.

Source environments¶

Use uv from the repository root so the source environment is resolved from pyproject.toml and uv.lock.

Recommended development environment¶

uv sync --extra dev --extra docs --group spacy-model

For paper reproduction and throughput comparisons, install the full artifact stack when the machine has a compatible CUDA / flash-attention environment:

uv sync --extra dev --extra cuda --group spacy-model

Runtime-only environment¶

uv sync

Optional groups and extras¶

uv sync --group spacy-model
uv sync --extra cuda

spacy-model installs the optional en-core-web-lg wheel. cuda installs the configured flash-attn wheel and should only be used on a matching Linux CUDA environment. Accuracy reproduction can run without cuda, but retrieval and reranking throughput may differ from the paper performance environment.

Verify the install¶

Use a light import check first. It does not instantiate embedding models.

uv run python -c "import mandol; print(mandol.__version__)"

The expected package version in this checkout is 0.1.0.

Important dependency note¶

The default dependency list is intentionally the complete research/runtime stack. It includes Torch, transformers, sentence-transformers, FAISS CPU, RocksDB, graph libraries, LLM clients, retrieval/rerank dependencies and benchmark tooling. This differs from older docs that described many optional pip install mandol[...] extras. In the current pyproject.toml, the maintained extras are dev, docs and cuda.