Stat Chat

Data Cleaning · Normalization · Statistical Analysis · AI-Powered Iterative Adjustment · PDF Reports

A Python tool with both a GUI (Tkinter) and CLI interface for analysing tabular data from CSV or Excel files. Powered by Claude API or a local LM Studio model for natural-language dataset adjustment and annotated-report parsing.

---

Installation

pip install statchat-app

Tkinter ships with standard Python on Windows/macOS. On Linux: sudo apt install python3-tk

---

Launch

statchat              # GUI
statchat --cli --help # CLI

---

GUI Walkthrough

1. Load CSV / Excel — opens a file picker; the Data Preview tab fills instantly.
2. Data Cleaning — tick any combination of:
   • Drop duplicate rows
   • Drop rows with null values
   • Fill nulls (mean / median / mode / zero)
3. Normalization — choose one method:
   • Z-score — standardise with custom target mean & std (default 0, 1)
   • Min-Max — scale to [0, 1]
   • Robust — median/IQR scaling (outlier-resistant)
4. Analysis Metrics — tick any combination:
   • Measures of Central Tendency (mean, median, mode)
   • Measures of Dispersion (std dev, variance, IQR, range, CV)
   • Shape stats (skewness, kurtosis)
   • Percentile Statistics (P5–P95)
   • Normality Tests (Shapiro-Wilk)
   • Correlation Matrix (heatmap in PDF)
   • ROC-AUC (enter the binary target column name)
5. Run Analysis — results appear in the Analysis Results tab.
6. Save Cleaned Data — exports as CSV, Excel, or JSON.
7. Export PDF Report — full styled report with tables and charts.
8. ✦ Adjust Data — iteratively modify the dataset using plain English or annotated images (see below).

---

AI-Powered Data Adjustment

The ✦ Adjust Data tab lets you modify the dataset conversationally after analysis, without writing any code.

Text instructions

Type any natural-language instruction and press Send:

"Add $1000 to each value in spend"
"Multiply income by 1.1"
"Clip spend between 0 and 5000"
"Remove rows where age < 18"
"Log-transform spend"
"Rename 'score' to 'risk_score'"
"Fill nulls in income with 0"

Stat Chat sends the instruction to the configured LLM, parses the response into typed operations, shows you a diff preview (columns affected, mean before/after), and waits for you to Accept or Discard before applying the change.

Annotated report images (vision)

Print or screenshot your PDF report, annotate it with handwritten or typed notes (e.g. circle a column and write "× 1.05"), then click 📷 Image to upload the photo. A vision-capable model reads the annotations and proposes the matching operations — same Accept/Discard flow.

Version history

Every accepted change is saved as a numbered version in the sidebar. You can:

• Revert to any prior version
• Save any version to file (in the original file format)
• Generate a PDF report for any specific version

The PDF report includes a full Iterative Adjustment History section showing every version, what changed, and a mean-evolution table across versions.

---

LLM Backend Configuration

Click ⚙ Settings in the top-right to choose your backend.

Provider	Use case	Setup
Claude API (default)	Cloud, no local setup	API key injected automatically
LM Studio — Text	Local, privacy-first	Load any text model in LM Studio
LM Studio — Vision	Local + image annotations	Load a multimodal model (LLaVA, Moondream, etc.)

LM Studio setup:

1. Download LM Studio
2. Load a text model (e.g. Llama 3, Mistral) for chat adjustments
3. Load a vision model (e.g. LLaVA, BakLLaVA) for image annotation parsing
4. Developer tab → Enable Local Server → Start Server
5. In Stat Chat → ⚙ Settings → select LM Studio → click ↻ Fetch loaded models → Test Connection → Save

---

CLI Reference

statchat --cli --input FILE [options]

File I/O:
  --input,  -i PATH         Input CSV or Excel file (required)
  --output, -o PATH         Save processed data here
  --output-format           csv | xlsx | json  (default: csv)
  --report, -r PATH         Save PDF report here

Cleaning:
  --drop-duplicates         Remove duplicate rows
  --drop-nulls              Drop rows containing any null
  --fill-nulls STRATEGY     mean | median | mode | zero

Normalization:
  --normalize METHOD        zscore | minmax | robust
  --norm-mean FLOAT         Z-score target mean  (default 0.0)
  --norm-std  FLOAT         Z-score target std   (default 1.0)

Analysis:
  --central-tendency        Mean, median, mode
  --dispersion              Std dev, variance, IQR, range
  --shape                   Skewness & kurtosis
  --percentiles             P5–P95
  --correlation             Pearson correlation matrix
  --roc-auc TARGET_COL      ROC-AUC vs a binary target column
  --all-metrics             Enable all metrics above

Adjustment (requires LLM backend):
  --adjust INSTRUCTION      Natural-language adjustment (repeatable)
  --adjust-image IMAGE_PATH Annotated report image (PNG/JPG)

LLM Backend:
  --backend                 claude | lmstudio | lmstudio_vision  (default: claude)
  --lmstudio-url URL        LM Studio server URL (default: http://localhost:1234)
  --lmstudio-model ID       LM Studio text model ID
  --lmstudio-vision-model ID LM Studio vision model ID

Examples

Full pipeline with text adjustments:

statchat --cli \
  --input sales.csv \
  --drop-duplicates --fill-nulls mean \
  --normalize zscore \
  --central-tendency --dispersion --correlation \
  --adjust "Add 1000 to spend" \
  --adjust "Multiply income by 1.1" \
  --output adjusted_sales.xlsx \
  --output-format xlsx \
  --report report.pdf

Parse an annotated report image with a local vision model:

statchat --cli \
  --input data.csv \
  --adjust-image annotated_report.png \
  --backend lmstudio_vision \
  --lmstudio-url http://localhost:1234 \
  --lmstudio-vision-model llava-1.5-7b \
  --output adjusted.csv \
  --report report.pdf

---

Supported Metrics

Category	Metrics
Central Tendency	Mean, Median, Mode, Count
Dispersion	Std Dev, Variance, Range, IQR, Min, Max, CV
Shape	Skewness, Kurtosis
Percentiles	P5, P10, P25, P50, P75, P90, P95
Normality	Shapiro-Wilk statistic & p-value
Correlation	Pearson correlation matrix + heatmap
ROC-AUC	Per-feature AUC score + ROC curve plot

---

Project Structure

statchat/
├── __main__.py           Entry point (GUI or CLI)
├── icon.ico
├── assets/
├── core/
│   ├── loader.py         File I/O (CSV, Excel, JSON)
│   ├── cleaner.py        Cleaning & normalization
│   ├── analyzer.py       Statistical metrics
│   ├── adjuster.py       NL instruction parser & executor
│   ├── llm_backend.py    Claude API + LM Studio abstraction
│   └── reporter.py       PDF report generation
├── gui/
│   ├── app.py            Main Tkinter GUI
│   ├── chat_panel.py     Adjust Data tab (chat + version history)
│   └── settings_dialog.py LLM backend settings
└── cli/
    └── runner.py         CLI runner

---

Requirements

pandas >= 2.0
numpy >= 1.24
scipy >= 1.10
scikit-learn >= 1.3
openpyxl >= 3.1
reportlab >= 4.0
matplotlib >= 3.7
Pillow >= 9.0
requests >= 2.28

Tkinter is required for the GUI and ships with standard Python. Linux users may need sudo apt install python3-tk.