tableai 0.2.3

AI toolkit for tabular data — auto EDA, data profiling, anomaly detection, and smart transformations on DataFrames.

tableai

Profile, clean, and query tabular data with one-liners — plus natural-language DataFrame analysis.

tableai is a toolkit for making sense of DataFrames fast. Profile any DataFrame and get column types, null counts, descriptive statistics, correlations, and a data-quality score. Clean it with a single call that imputes missing values, drops duplicates, and clips outliers. Detect anomalies with IQR or Isolation Forest. Get rule-based natural-language insights — or ask questions in plain English and have anyllm generate the pandas code for you.

Built by Viet-Anh Nguyen at NRL.ai.

Why tableai?

• One-liner API — tableai.profile(df) gives you everything in one call
• Plugin architecture — Register custom profilers, cleaners, and anomaly detectors
• Local-first — All core features work without any cloud or LLM call
• Minimal core deps — pandas and numpy; sklearn and anyllm are optional
• Production-ready — Structured dataclass results, JSON export, reproducible

Installation

pip install tableai

For optional features:

pip install tableai[sklearn]   # Isolation Forest + KMeans clustering
pip install tableai[llm]       # NL querying via anyllm
pip install tableai[all]       # everything

Python 3.8+ supported (tested on 3.8, 3.9, 3.10, 3.11, 3.12, 3.13)

Quick Start

import tableai
import pandas as pd

df = pd.read_csv("sales.csv")

# 1. Profile the DataFrame (dtypes, nulls, stats, correlations, quality score)
report = tableai.profile(df)
print(report.quality_score)              # 0.0 - 1.0
print(report.nulls)                      # per-column null counts
print(report.correlations.head())        # top correlated pairs

# 2. Clean the DataFrame (impute, dedupe, clip outliers)
clean = tableai.clean(df, impute=True, dedupe=True, clip_outliers=True)

# 3. Detect anomalies (IQR by default, Isolation Forest if sklearn installed)
anomalies = tableai.anomalies(df, method="iqr")
print(f"{len(anomalies)} anomalous rows")

# 4. Rule-based insights
for insight in tableai.insights(df):
    print("-", insight)

# 5. Natural-language querying (requires tableai[llm] + anyllm)
result = tableai.ask(df, "what is the average revenue by region?")
print(result)

Models & Methods

Profiling

• Dtype detection — numeric / categorical / datetime / text / boolean / ID
• Null analysis — per-column null counts, percentages, and null patterns
• Descriptive statistics — mean, std, min, 25/50/75 percentiles, max, skew, kurtosis
• Cardinality — unique counts and top-K value frequencies
• Correlation matrix — Pearson for numerics, Cramer's V for categoricals
• Duplicate detection — exact and near-duplicate row counts

Cleaning

Configurable pipeline applied in order:

1. Drop constant columns — zero variance
2. Impute — median for numerics, mode for categoricals (configurable)
3. Deduplicate — drop exact-duplicate rows
4. Clip outliers — IQR method ([Q1 - 1.5*IQR, Q3 + 1.5*IQR])
5. Type coercion — auto-convert date-like strings to datetime

Anomaly detection

Method	Algorithm	Notes
iqr (default)	1.5 x IQR per numeric column	Zero deps
zscore	`	z
isolation_forest	sklearn IsolationForest	Needs tableai[sklearn]

Data quality score

Weighted average (0.0 - 1.0) of four sub-scores:

• Completeness — 1 - null_ratio
• Uniqueness — ratio of distinct rows
• Consistency — fraction of columns with a dominant dtype
• Validity — fraction of values inside expected ranges / formats

Insights (rule-based NL)

Pattern-driven natural-language observations, for example:

• "Column 'age' has 23.4% missing values"
• "'price' and 'quantity' are strongly positively correlated (r=0.87)"
• "Column 'id' appears to be a unique identifier"
• "12 rows are exact duplicates"

Natural-language querying (optional)

tableai.ask(df, "…") uses anyllm to generate pandas code for your question, executes it in a sandboxed namespace, and returns the result. Works with any local or cloud LLM that anyllm supports.

Models & Methods

tableai uses pure pandas/numpy for core operations — no ML dependencies required.

Profiling (tableai.profile) — Computes per-column:

• Dtype detection (numeric, categorical, datetime, string)
• Null counts and percentages
• Unique value counts
• Numeric statistics: mean, median, std, min, max, quartiles, skewness, kurtosis
• Top categorical values
• Pearson correlation matrix between numeric columns

Cleaning (tableai.clean) — Configurable strategies:

• Missing values: median (numeric), mode (categorical), drop, or zero
• Duplicate removal
• Outlier handling: IQR-based clipping or removal

Anomaly Detection (tableai.anomalies):

• IQR method (default, no deps) — flags points outside Q1-1.5·IQR / Q3+1.5·IQR
• Isolation Forest (optional via [ml], requires scikit-learn)

Quality Scoring (tableai.quality_score) — Weighted score 0-100:

• Completeness 35% (1 - null_ratio)
• Validity 25% (IQR-based outlier ratio)
• Uniqueness 20% (duplicate detection)
• Consistency 20% (mixed-type detection)

Insights (tableai.insights) — Rule-based natural language insights about missing values, correlations, skewness, cardinality, duplicates, and class imbalance.

Natural Language Querying (tableai.ask, tableai.query) — Optional via [llm] extra. Uses anyllm to generate pandas code from natural language. Falls back to keyword matching when LLM unavailable.

API Reference

Function	Purpose
tableai.profile(df)	Returns ProfileReport dataclass
tableai.clean(df, **opts)	Returns a cleaned DataFrame
tableai.anomalies(df, method="iqr")	Returns rows flagged as anomalous
tableai.quality_score(df)	Returns float 0.0 - 1.0
tableai.insights(df)	Returns list[str] of NL insights
tableai.ask(df, question, model=None)	NL query via LLM
tableai.compare(df1, df2)	Diff two DataFrames (schema + data)

CLI Usage

tableai profile data.csv --out report.json
tableai clean data.csv --out clean.csv
tableai anomalies data.csv --method isolation_forest
tableai ask data.csv "average sales by region"
tableai quality data.csv

Examples

Full profiling report to JSON

import tableai, pandas as pd

df = pd.read_csv("customers.csv")
report = tableai.profile(df)
report.to_json("customers_report.json")
print(f"Quality: {report.quality_score:.2f}")

Custom cleaning pipeline

import tableai

clean = tableai.clean(
    df,
    impute_numeric="median",
    impute_categorical="mode",
    dedupe=True,
    clip_outliers=True,
    drop_constant=True,
)

Ask questions in English (with Ollama)

import tableai

# Uses anyllm; defaults to Ollama if running locally
answer = tableai.ask(df, "which customer spent the most last quarter?",
                     model="llama3.1:8b")
print(answer)

License

MIT (c) Viet-Anh Nguyen