---

name: load-data description: Ultra-fast data discovery and loading skill for industrial finance datasets. Handles CSV, JSON, DuckDB, Parquet, and Feather with memory-efficient 2-step discovery and loading. license: MIT metadata: version: "1.0" capabilities: ["CSV", "JSON", "DuckDB", "Parquet", "Feather", "Bulk Loading"]

Load Data Skill

This skill provides a robust framework for importing data from disparate sources into a clean, analysis-ready format. It is designed to be standalone and high-performance.

Instructions for Agents (Discovery & Inference)

Mandatory Discovery Step: If the user provides a data file but does not specify its structure, schema, or contents, you MUST first use the peek_file capability to extract the first 5 rows. Send this sample to the LLM context to infer the data schema, column types, and general structure before attempting any full load or complex analysis.

Capabilities

• Bulk Multi-Format Collation: Efficiently merge multiple CSV, Parquet, or Feather files from a directory.
• DuckDB Integration: Native support for searching and reading DuckDB databases in read-only mode.
• JSON Normalization: Flatten nested JSON structures for tabular analysis.
• Automated Type Inference: Intelligent parsing of dates and numeric types with epoch correction for modern data datasets (post-2000).
• Industrial Memory Safety: Forced chunked loading for files > 100MB and strict size-based collation guards.
• Automatic Column Cleaning: Standardizes column names by default (lowercase, strip non-printable, spaces to underscores, valid Python identifiers for attribute access).

Usage

1. Multi-File Collation

Merge multiple files into a single DataFrame. Supports CSV, Parquet, and Feather. Operation is only performed if total size is < 100MB.

Copyfrom fastbt.loaders import collate_data

# Merge all small historical CSVs
df = collate_data(directory="./history", pattern="*.csv")

# Merge parquet partitions
df_p = collate_data(directory="./partitions", pattern="*.parquet")

2. Normalized JSON Loading

Convert complex JSON logs into a flat structure.

Copyfrom fastbt.loaders import normalize_json

data = [
    {"id": 1, "meta": {"time": "2023-01-01", "val": 10}},
    {"id": 2, "meta": {"time": "2023-01-02", "val": 20}}
]
df = normalize_json(data)
# Resulting columns: id, meta_time, meta_val

Flexible Data Workflow

This skill supports both data discovery and full data loading. These can be used independently, or combined for a complete end-to-end pipeline.

1. Data Discovery

Use this when you want to understand the structure, schema, or contents of a data file without committing to a full load. This is especially useful for large files or unknown datasets.

• Sample Inspection: Uses high-speed CLI tools (where possible) or specialized readers to extract a small sample.
• DuckDB Discovery: When peeking at a .duckdb or .db file, it lists all available tables and returns a sample from the first one.
• Metadata Awareness: If accompanying metadata is available (e.g., schemas, column descriptions), the discovery step can incorporate this information.

Copyfrom fastbt.loaders import peek_file

# Rapid discovery (head-based for text files)
df_sample = peek_file("potential_dataset.csv")

# DuckDB discovery: lists tables and peeks at the first one
df_db_sample = peek_file("market_data.duckdb")

2. Full Data Loading

Use this when you are ready to load data into memory or a persistent store. This step includes industrial-grade memory safety and performance optimizations.

• Memory Efficiency: Automatically switches to chunked loading for files > 100MB.
• Read-Only Safety: Database connections (DuckDB) are strictly read-only to prevent accidental mutation of source data.
• Aggregation: Merge multiple partitions or files into a single dataset.

Copyfrom fastbt.loaders import efficient_load, collate_data

# Direct full load with safety guards
df = efficient_load("source_data.parquet")

# DuckDB connection: returns a read-only DuckDB connection object
con = efficient_load("analytics.db")
df_query = con.execute("SELECT * FROM trades").df()

# Process multiple files with an optional transform
df_merged = collate_data(directory="./history", pattern="*.csv")

3. Combined Pipeline (Discovery + Loading)

In most cases, you will use discovery to infer the schema and then perform a full load with specific parameters (like date column names or data types) identified during discovery.

Configuration & Overrides

All loading functions apply default column cleaning and datetime conversion. You can override these behaviors by passing keyword arguments:

• lower (default: True): Convert columns to lowercase.
• strip_non_printable (default: True): Remove non-printable characters and extra whitespace.
• replace_spaces (default: True): Replace spaces with underscores.
• ensure_identifiers (default: True): Ensure column names are valid Python identifiers (attribute access).

Copy# Load without modifying column case or replacing spaces
df = efficient_load("data.csv", lower=False, replace_spaces=False)

Included Scripts

• fastbt.loaders: The core module for the loading operations.

Reference Materials

• references/DATA_FORMATS.md: Guide on supported data types and optimization techniques.