Charts and Visualization¶

KITT can generate charts from stored benchmark results using Matplotlib. This is useful for visualizing quantization quality tradeoffs, comparing engine performance, and tracking metric trends over time.

Installation¶

Charts require the optional charts extra:

poetry install -E charts

This pulls in Matplotlib and its dependencies.

Quantization Curves¶

Generate a chart showing how quantization levels affect quality metrics across model families:

kitt charts quant-curves

Filter by model family:

kitt charts quant-curves --model-family Llama-3

Change the output file:

kitt charts quant-curves --output llama3_quant.svg

Export the underlying data as CSV instead of rendering a chart:

kitt charts quant-curves --csv

Output Formats¶

Charts are saved as SVG by default. The output format is determined by the file extension you provide:

Extension	Format
`.svg`	Scalable Vector Graphics (default)
`.png`	Raster image
`.pdf`	PDF document

Example:

kitt charts quant-curves --output curves.png

Data Sources¶

Chart commands read from the KITT storage backend. KITT tries SQLite first (SQLiteStore) and falls back to the JSON store (JsonStore). Make sure you have stored results (via kitt run or the storage commands) before generating charts.

Chart Types¶

Command	Description
`kitt charts quant-curves`	Quality vs. quantization level curves

Additional chart types may be added through the plugin system. The chart generation logic lives in src/kitt/reporters/quant_curves.py and uses the same result store interface as the rest of KITT.