Datasets:

frupniew
/

macaulay2-rag-chunks

task_categories:
  - text-retrieval
language:
  - en
  - code
tags:
  - macaulay2
  - rag
  - documentation
  - vector-database
pretty_name: Macaulay2 RAG Chunks
license: gpl-2.0

Macaulay2 RAG Chunks

A comprehensive knowledge base extracted directly from the official Macaulay2 source code and documentation packages (M2/Macaulay2/packages). Designed to be ingested into Vector Databases (e.g., ChromaDB, FAISS) for Retrieval-Augmented Generation (RAG).

📊 Dataset Statistics

Total Chunks: 15,711
Size: ~45MB
Code Density: ~85% of chunks contain executable Macaulay2 examples.
Source: Parsed from 2,208 .m2 files using a custom AST/Regex-aware parser (https://gitlab.com/frupniew/llm_macaulay/src/extract_m2_docs_v2.py) that groups documentation per mathematical symbol.

🏷️ Metadata & Filtering

Each chunk is enriched with metadata to allow for highly targeted retrieval:

{
  "id": "chunk_8492",
  "symbol": "primaryDecomposition",
  "package": "PrimaryDecomposition",
  "has_code": true,
  "headline": "compute the primary decomposition of an ideal",
  "usage": "primaryDecomposition I",
  "example_code": "R = QQ[x,y,z]; I = ideal(x^2, x*y); primaryDecomposition I"
}

This structure allows RAG pipelines to filter out purely theoretical text when the user explicitly asks for code implementation, or filter by specific algebraic packages (e.g., Schubert2).
🧩 Recommended Embedding Model
Tested and optimized with sentence-transformers/all-MiniLM-L6-v2 for a balance of semantic mathematical understanding and low-latency retrieval.