pysublime 0.1.0

Line-level code embeddings with smart segmentation and negative prompting

Sublime 🍋‍🟩

Smart sub-line segmented embeddings that leave you feeling sublime.

Sublime offers two uniquely powerful features:

1. Line-by-line embeddings segmented as-needed to get exactly the code snippets required, nothing more or less
2. Negative prompting - just like image generation, searching over embeddings should include negative prompts

How it works

Instead of chunking code arbitrarily, Sublime embeds every line of code individually. When you search, it uses ML clustering (similar to audio segmentation) to group nearby high-scoring lines into coherent code segments that target exactly what you've searched for.

Quick Start

from embedings import Embeddings

# Index your codebase
emb = Embeddings("./my_project")

# Search with semantic queries
results = emb.search("user authentication and login validation flow", top_n=1)

# Search with negative prompting to exclude unwanted patterns
results = emb.search("database connection and query execution", negative_query="test mocks and unit testing", top_n=1)

# Negative-only search to find problematic code
results = emb.search(negative_query="clean well-documented modern code", top_n=1)

# Print results
for result in results:
    print(f"File: {result.file_path}")
    print(f"Lines {result.start_line}-{result.end_line}")
    print(result)

Output format

Search returns a list of CodeSection objects, globally ranked by similarity:

• file_path: path to the source file
• start_line, end_line: inclusive span of the snippet
• lines: the exact lines in the span
• avg_similarity, max_similarity: scores in [0, 1] (higher is more relevant)

Printing a CodeSection (via print(result)) renders:

File: path/to/file.py (lines 120-134)
  120: def authenticate(user, password):
  121:     # ... code ...
  122:     return is_valid
  123: 
  124: class Session:
  125:     # ...

Usage Notes

• Use from Python by importing Embeddings.
• Control file types with supported_extensions and ignores with an .embedignore file.
• top_n returns the best segments globally across all files.

File watching (optional)

Keep the index up to date while you edit files.

import asyncio
from embedings import EmbeddingsSuite

async def main():
    suite = EmbeddingsSuite("./project", ignore_file=".embedignore")
    await suite.build_initial_index()
    await suite.start_watching()

    # ... use suite.search_code(query, negative_query, top_n) as you work ...

    await suite.stop_watching()

asyncio.run(main())

Notes:

• Respects .embedignore patterns
• Debounced ~1s, then rebuilds index on change
• Safe to run while developing; searches use the latest index

Configuration

# Custom file types and ignore patterns
emb = Embeddings(
    "./project",
    supported_extensions={".py", ".js", ".rs"}, 
    ignore_file=".embedignore"
)

Create .embedignore file:

node_modules
*.log
test_*
__pycache__

Install

pip install -r requirements.txt