matheel 0.2.2

Matheel: A CLI and Python package for source-code similarity detection.

Matheel

Matheel is a simple, function-based Python package and CLI for source-code similarity. It combines semantic embeddings, lexical similarity, chunking, preprocessing, and code-aware metrics without forcing a class-heavy API.

Installation

Use Python 3.10 to 3.12.

Base install:

pip install matheel

Optional extras:

pip install "matheel[chunking]"
pip install "matheel[chunking_code]"
pip install "matheel[model2vec]"
pip install "matheel[pylate]"
pip install "matheel[gradio]"
pip install "matheel[all]"
pip install "matheel[dev]"

Tested Python 3.12 constraints:

pip install -c constraints/py312.txt .

matheel[all] installs the currently supported optional backends in one command: Chonkie code chunking, model2vec, PyLate, and the Gradio app dependencies.

Quick Start

The repo includes a small Java archive at sample_pairs.zip for smoke tests.

CLI:

matheel compare sample_pairs.zip \
  --model sentence-transformers/all-MiniLM-L6-v2 \
  --feature-weight semantic=0.7 \
  --feature-weight levenshtein=0.3 \
  --threshold 0.2 \
  --num 10

Python:

from matheel.similarity import get_sim_list

results = get_sim_list(
    "sample_pairs.zip",
    model_name="sentence-transformers/all-MiniLM-L6-v2",
    threshold=0.2,
    number_results=10,
    feature_weights={
        "semantic": 0.7,
        "levenshtein": 0.3,
    },
)
print(results.head())

Supported Languages

• Chunking is language-agnostic by default because it can split any text.
• CodeBLEU-style metrics are intentionally scoped to Java, Python, C, and C++.
• Generic preprocessing works across languages, but code-aware metrics are most defensible in that four-language scope.

Supported Methods

Similarity features:

• semantic
• levenshtein
• jaro_winkler
• code_metric

Code metrics:

• codebleu
• codebleu_ngram
• codebleu_weighted_ngram
• codebleu_syntax
• codebleu_dataflow
• crystalbleu

Chunking methods:

• none
• Chonkie-backed when installed: code, codechunker, chonkie_code, chonkie_token, chonkie_sentence, chonkie_recursive, chonkie_fast

Vector backends:

• auto
• sentence_transformers
• model2vec
• pylate

Single-vector similarity functions:

• cosine
• dot
• euclidean
• manhattan

Sentence Transformers pooling methods:

• mean
• max
• cls
• lasttoken
• mean_sqrt_len_tokens
• weightedmean

auto inspects Hugging Face model metadata and routes to the correct backend when the model exposes a known library.

Core Parts

• Preprocessing: whitespace and comment normalization before any scoring.
• Chunking: Chonkie-backed document splitting with per-method options.
• Vectors: dense single-vector, learned static single-vector, and multivector late interaction.
• Edit distance: normalized Levenshtein and Jaro-Winkler lexical metrics.
• Code metrics: built-in CodeBLEU-style metrics and CrystalBLEU.
• Comparison suite: run multiple configurations, rank them, and optionally write summary/detail artifacts.

CLI

Compare a directory or ZIP archive:

matheel compare codes/ \
  --model sentence-transformers/all-MiniLM-L6-v2 \
  --vector-backend auto \
  --max-token-length 256 \
  --feature-weight semantic=0.6 \
  --feature-weight levenshtein=0.2 \
  --feature-weight jaro_winkler=0.1 \
  --feature-weight code_metric=0.1 \
  --similarity-function dot \
  --pooling-method max \
  --preprocess-mode basic \
  --chunking-method code \
  --chunk-language python \
  --chunker-option include_line_numbers=true \
  --code-metric codebleu \
  --code-language python \
  --threshold 0.5 \
  --num 25

Run multiple configurations:

matheel compare-suite codes/ runs.json \
  --summary-out results/summary.csv \
  --details-dir results/runs

Python API

Pairwise scoring:

from matheel.similarity import calculate_similarity

score = calculate_similarity(
    "def add(a, b):\n    return a + b\n",
    "def add(x, y):\n    return x + y\n",
    model_name="sentence-transformers/all-MiniLM-L6-v2",
    vector_backend="auto",
    max_token_length=256,
    similarity_function="dot",
    pooling_method="max",
    preprocess_mode="basic",
    chunking_method="code",
    chunk_language="python",
    code_metric="codebleu",
    code_language="python",
    feature_weights={"semantic": 0.5, "code_metric": 0.5},
)

Directory or ZIP ranking:

from matheel.similarity import get_sim_list

results = get_sim_list(
    "sample_codes",
    model_name="sentence-transformers/all-MiniLM-L6-v2",
    threshold=0.4,
    number_results=50,
    vector_backend="auto",
    max_token_length=256,
    chunking_method="chonkie_token",
    chunk_size=120,
    chunk_overlap=20,
    similarity_function="cosine",
    pooling_method="mean",
    feature_weights={
        "semantic": 0.7,
        "levenshtein": 0.15,
        "jaro_winkler": 0.15,
    },
)

Hugging Face Routing

Matheel can inspect Hugging Face model metadata and route automatically:

• sentence-transformers models go to the Sentence Transformers path
• model2vec models go to the model2vec static path
• PyLate models go to the multivector late-interaction path

If metadata is unavailable, Matheel falls back to simple name and tag heuristics, then defaults to the Sentence Transformers path.

Docs

• Docs index: docs/index.md
• Quick usage: docs/usage.md
• Preprocessing: docs/preprocessing.md
• Chunking: docs/chunking.md
• Vectors: docs/vectors.md
• Edit distance and feature weights: docs/lexical.md
• Code metrics: docs/code_metrics.md
• Comparison suite: docs/comparison_suite.md

The docs/ folder is already structured well for a later GitHub Pages setup if you decide to publish the docs site from the repository.

Examples

• Quick archive smoke test: examples/sample_pairs_demo.py
• Preprocessing: examples/preprocessing_demo.py
• Chunking: examples/chunking_demo.py
• Vector backends: examples/vectors_demo.py
• Edit distance and feature weights: examples/lexical_demo.py
• Code metrics: examples/code_metrics_demo.py
• Comparison suite: examples/comparison_suite_demo.py

Gradio

The Gradio demo stays in gradio_app/. The core package and CLI can read either a ZIP archive or a directory. The Gradio upload flow remains ZIP-based.

License

This project is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License.