Table · LSD API

Table is the data that flows between steps — an immutable, zero-copy view over one pyarrow.Table. Every verb returns a new Table (or a Column), and each compiles to a single Arrow compute kernel, so chains stay zero-copy and fast. The point is that a geologist never types pyarrow.compute and an LLM gets a small, promptable verb list — while performance stays identical to raw pyarrow.

Usage#

Build one by calling the Table(...) constructor, then chain verbs:

Python

from lsdtools import Table

t = Table({"name": ["ann", "bob", "cy"], "score": [12, 30, 7]})
top = t.filter(t["score"] > 10).sort("score", descending=True)
top.print()
# → name:["bob","ann"] score:[30,12]

Inside a step you receive a Table as an input port and return one:

Python

@tool.shape
def keep_top(t: Table, cutoff: int = 10) -> Table:
    return t.filter(t["score"] >= cutoff)

Where to start#

Learn — Tables & data flow · Shape steps

Examples — Clean a CSV

API — Column · Source · Context · Engine

Methods

drop removes the named columns, keeping the rest

filter keeps the rows where a boolean mask column is true

group_by groups by one or more columns; chain .agg to reduce

head takes the first n rows

join joins to another table on a key column

null_report summarises null counts per column as a small table

print prints a compact preview and returns self for chaining

rename renames columns from an {old: new} mapping

select keeps only the named columns, in the given order

sort sorts by one or more columns, ascending or descending

to_pylist materialises to a list of row dicts (preview/small tables)

with_column adds or replaces a column from a column expression or array

write_csv writes the table to a csv file

write_parquet writes the table to a parquet file

Usage#

Where to start#

Properties

Methods