carloforte 0.2.2

Extract structured data from Excel files with minimal token usage

carloforte

Extract structured data from Excel files with minimal token usage.

carloforte uses an island-detection algorithm to convert Excel sheets into a compact intermediate representation (CSV, Markdown, or JSON), making it efficient to pass spreadsheet data to LLMs.

Installation

uv add carloforte

Usage

import carloforte

# Extract all sheets as CSV (default)
text = carloforte.extract("data.xlsx")

# Extract specific sheets as Markdown
text = carloforte.extract("data.xlsx", sheets=["Revenue", "Costs"], fmt="markdown")

# Extract as JSON
text = carloforte.extract("data.xlsx", fmt="json")

Formats

Format	Best for
csv	Compact, low token count
markdown	Readable, good for LLM prompts
json	Structured output, programmatic use

CLI

carloforte data.xlsx --fmt markdown
carloforte data.xlsx --sheets Revenue Costs --fmt json

How it works

Excel sheets often contain multiple disconnected tables, empty rows, and metadata scattered around. carloforte detects each contiguous block of data ("island") independently and serialises only what matters — reducing token usage by 60–75% compared to passing raw Excel content to an LLM.

Performance

Benchmarked against raw CSV export (worst case baseline). Run examples/scripts/benchmark_tokens.py to reproduce.

Characters:

File	Raw CSV	Clean CSV	CSV islands	Markdown	JSON
large_sparse — 1 sheet, 5 islands spread over 165 rows	4,470	↓ 1,480 (-67%)	↓ 960 (-79%)	↓ 1,453 (-68%)	↓ 1,337 (-70%)
fragmented — 4 sheets, 4% fill, tiny tables far apart	10,664	↓ 5,334 (-50%)	↓ 4,811 (-55%)	↓ 6,690 (-37%)	↓ 6,221 (-42%)
stray_cells — 1 sheet, 7 islands + scattered stray cells	2,143	↓ 1,668 (-22%)	↓ 1,458 (-32%)	↑ 2,175 (+2%)	↓ 1,994 (-7%)
multisheet — 4 sheets, mixed structure	1,304	↓ 1,084 (-17%)	↓ 1,156 (-11%)	↑ 1,879 (+44%)	↑ 1,755 (+35%)
invoice — single invoice sheet	4,180	↓ 3,745 (-10%)	↓ 3,855 (-8%)	↑ 5,097 (+22%)	↑ 5,020 (+20%)
minimal — 1 sheet, 3 small islands	248	↓ 216 (-13%)	↓ 244 (-2%)	↑ 417 (+68%)	↑ 380 (+53%)
enterprise — 9 sheets, dense real-world complexity	23,983	↓ 22,533 (-6%)	↓ 22,639 (-6%)	↑ 29,357 (+22%)	↑ 28,741 (+20%)

Tokens (cl100k_base):

File	Raw CSV	Clean CSV	CSV islands	Markdown	JSON
large_sparse	1,033	↓ 513 (-50%)	↓ 443 (-57%)	↓ 616 (-40%)	↓ 605 (-41%)
fragmented	2,556	↓ 1,504 (-41%)	↓ 1,696 (-34%)	↓ 2,375 (-7%)	↓ 2,283 (-11%)
stray_cells	630	↓ 555 (-12%)	↓ 575 (-9%)	↑ 824 (+31%)	↑ 811 (+29%)
multisheet	499	↓ 430 (-14%)	↓ 484 (-3%)	↑ 737 (+48%)	↑ 721 (+45%)
invoice	1,317	↓ 1,230 (-7%)	↑ 1,323 (+0%)	↑ 1,934 (+47%)	↑ 1,839 (+40%)
minimal	83	↓ 71 (-15%)	↑ 89 (+7%)	↑ 143 (+72%)	↑ 139 (+67%)
enterprise	9,606	↓ 9,311 (-3%)	↑ 9,627 (+0%)	↑ 11,723 (+22%)	↑ 12,386 (+29%)

Optional Rust backend

carloforte ships a Rust implementation of the island-detection algorithm via PyO3. It is not enabled by default — the pure-Python implementation is used unless you build and install the extension manually.

The Rust backend (carloforte-rs) re-implements find_islands in Rust and exposes it as a native Python extension (_islands_rs). The BFS logic is identical to the Python version; the speedup comes from Rust's memory layout and the absence of Python object overhead on large grids.

Build from source:

cd carloforte-rs
maturin develop --release

Once built, the .so / .pyd extension will be importable as carloforte._islands_rs. Integration into the default pipeline is planned for a future release.

Architecture

carloforte/
├── __init__.py        public API: re-exports extract()
├── _extract.py        extract() — orchestrates the pipeline
├── _cli.py            main() — CLI entry point
├── _reader.py         load_workbook_sheets() — openpyxl → Grid
├── _islands.py        find_islands() — BFS island detection
└── _serialiser.py     serialise() — Grid → csv / markdown / json

carloforte.extract(path, sheets=None, fmt="csv")
│
│  1. _reader.load_workbook_sheets(path, sheets)
│     openpyxl → dict[sheet_name, Grid]
│
│  2. _islands.find_islands(grid)   ← per sheet
│     BFS over non-empty cells → list[Island]
│     each Island: bounding box + header row + data rows
│
│  3. _serialiser.serialise(sheet_islands, fmt)
│     "csv"      → one block per island, blank-line separated
│     "markdown" → ## heading per sheet, fenced table per island
│     "json"     → {"sheets": {"name": {"tables": [...]}}}
│
└─ returns str

License

MIT