Tokenize an NL question. Collapses whitespace + punctuation, preserves quoted strings as single tokens (for future "where name = 'alice'" support; cur

Returns true iff `e` is an AND-tree of (ColumnRef OP Constant) comparisons on numeric columns; fills `out`. On `false`, `out` may be partially filled

Union one group's per-thread distinct-int sets and return the distinct count. For W>=2 the union is split into W uid-hash buckets so a single huge gro

Decompress a Parquet ZSTD page using the vendored libzstd. Sized from the page header's uncompressed_size; ZSTD_decompressDCtx reuses a thread-local c

Find the SlothDB shared library.

Load the SlothDB C library.

Build list-of-row-tuples from C buffers. Slow for large results; prefer fetchnumpy() / fetchdf() in that case.

Open the database connection using the C API.

Assign a parsed value to a row slot based on the column's declared type.

Walk the question for referenced files, peek each schema that isn't already in the catalog, and splice the results into `schema`. Keeps the Qwen promp

Build a NL-to-SQL schema snapshot using the catalog-introspection C API. Straight ports the catalog contents into the engine's Schema struct; each str

Free the underlying C result. Called automatically on GC.

Execute a SQL statement (no result expected).

Find every filename-looking token in a question. Quoted ('x.csv') and bare (x.csv) paths both count. Used to auto-peek schemas the model needs for que

Return columns as a dict of numpy arrays. Numeric columns (INTEGER / BIGINT / FLOAT / DOUBLE) are zero-copy wraps over the C buffers

Find first byte equal to any of c1, c2, c3. Used to scan for the end of an unquoted CSV field in one pass - delim / '\n' / '\r'.