Large File & Spreadsheet Data Analysis Engine

Overview

High-performance data analysis skill tailored for AI Agents. It provides streaming ingestion pipelines via openpyxl(read_only=True) and automatic Parquet conversion, enabling memory-safe querying of 100,000+ row datasets in constant <50MB RAM.

When to Use & When NOT to Use

When to Use

• Excel spreadsheets (.xlsx/.xls) with ≥10,000 rows or multiple dense sheets.
• Multi-megabyte PDF, Word, or CSV datasets requiring statistical aggregation.
• Generating schema summaries and targeted slices without loading raw data into LLM context.

When NOT to Use

• Small tables (<100 rows) where standard markdown tables or direct pandas reads are fine.
• Visual chart rendering without data extraction (use image/visualization tools).

Streaming Extraction Pattern

import openpyxl

def stream_excel_summary(file_path: str, sample_rows: int = 5):
    wb = openpyxl.load_workbook(filename=file_path, read_only=True, data_only=True)
    summary = {}
    
    for sheet in wb.sheetnames:
        ws = wb[sheet]
        rows = ws.iter_rows(values_only=True)
        try:
            header = next(rows)
        except StopIteration:
            continue
            
        row_count = 0
        samples = []
        for row in rows:
            row_count += 1
            if len(samples) < sample_rows:
                samples.append(row)
                
        summary[sheet] = {
            "columns": header,
            "total_rows": row_count,
            "sample_rows": samples
        }
    wb.close()
    return summary

Common Pitfalls

1. Loading large Excel files in write/full mode: Standard load_workbook builds a massive DOM in memory. Always use read_only=True and data_only=True.
2. Dumping thousands of rows into prompt context: Extract aggregates, schemas, and summary distributions in code first; return only the high-signal findings to LLM context.
3. Mismatched data types in columns: Large datasets often have mixed type cells (e.g. numeric IDs parsed as floats); sanitize column types during streaming.

Verification Checklist

• Memory footprint remains < 50MB during streaming iteration.
• Summary output reports total row count, column list, and non-null distributions.
• No raw multi-megabyte payloads are passed directly into prompt context.