Dependency shape (pyarrow-free by default)¶

pip install laterite pulls polars + duckdb only — no pandas, no pyarrow. A base install stays lean.

pip install laterite                  # polars + duckdb only
pip install laterite[compat]          # + pandas<3 (the python-ags4 drop-in) — pyarrow-free
pip install laterite[compat,pyarrow]  # + pyarrow accelerator (faster pandas hop + string dtype)
pip install laterite[pyarrow]         # pyarrow for the explicit Arrow backend
pip install laterite[all]             # pandas + pyarrow

DuckDB is the pyarrow-free dataframe bridge. The compat drop-in's AGS4_to_dataframe reads a Rust-built all-Utf8 Arrow table (no per-cell Python boxing) and hands pandas an object-dtype frame through DuckDB's NumPy .df() materialiser — the same trick the core uses — so pandas works without pyarrow, and already runs ~2× faster than python-ags4. The polars path ingests the same Arrow via the C-stream capsule, also pyarrow-free.

pyarrow is an optional accelerator¶

Add pyarrow ([compat,pyarrow] or [all]) and compat auto-detects it: the pandas hop swaps to pyarrow's to_pandas (a touch faster), and you unlock the string_dtype="string" output — pandas' Arrow-backed str dtype, which is what python-ags4 itself returns once it runs on pandas 3. Without pyarrow, the default object dtype path is used; string_dtype="string" raises an actionable error rather than downgrading.

from laterite import compat as AGS4

# object dtype (numpy) — today's python-ags4 baseline, the default
tables, _ = AGS4.AGS4_to_dataframe("delivery.ags")

# string dtype (pandas' Arrow-backed str) — needs [compat,pyarrow]
tables, _ = AGS4.AGS4_to_dataframe("delivery.ags", string_dtype="string")
AGS4.set_string_dtype("string")            # process-wide (or LATERITE_COMPAT_STRING_DTYPE)

Which extra do I want?

Most callers want none. Reach for [compat] to use the laterite.compat python-ags4 shim (pandas-backed by default); it adds pandas<3 and nothing else. Add [compat,pyarrow] for the faster pandas hop and the Arrow-backed string dtype. Reach for [pyarrow] when you explicitly want the Arrow backend — e.g. handing native pyarrow.Table objects to another Arrow-native library.

So a base user gets polars + duckdb and an Arrow-capable bridge without dragging in two heavyweight dataframe stacks. Importing laterite.compat without the extra is safe — the lazy materialiser only raises (with a pip install laterite[compat] hint) when a pandas-backed call is actually made.