Construct an :class:`LLMProvider`. Falls back to :class:`Settings` defaults for any value not supplied, and resolves the API key from setting

Compute summary statistics for ``df``. ``sample_size`` controls how many example values we capture per column — useful for showing the LLM wh

Load a CSV from a path, bytes, or file-like object. Raises :class:`DataError` on failure with a readable message. Never silently swallows dec

Return a namespaced logger under ``csvai.<name>``.

Return suggested models for ``provider``. Custom names are allowed.

Return the list of registered provider names.

Return a markdown string describing ``df`` for prompt injection.

Return a single, fully-formed response.

Return (provider, model, temperature, max_tokens, api_key, base_url).

Render an expandable preview with key metrics + the head of the frame.

Return the loaded dataset or render an info message and ``None``.

Return up to ``n`` rows, stratified when possible.

Render an assistant message that updates as tokens stream in. Returns the assembled text so the caller can persist it to history.

Wrap a user question with a light instruction reinforcement. Keeps the user message focused and consistent across providers.

Return a cached :class:`Settings` instance. On Streamlit Community Cloud, ``st.secrets`` provides API keys. We try to merge them in (best-eff

Try each candidate encoding until one decodes the first 4 KiB cleanly.

Group-by stratified sample, robust across pandas versions.

Group-by stratified sample, robust across pandas versions.

Inject the dashboard CSS. Idempotent within a session.

First + last rows for compact prompts.

Initialize every session key the app reads.

Render the sidebar and return the currently selected page name.

Render a tabbed view: schema, missing, numeric, correlation, charts.

Token-incremental answer. The caller is responsible for assembly.

One-shot answer (non-streaming).

Estimate the number of tokens for ``text``. Uses tiktoken for OpenAI-family models; otherwise approximates one token per four characters (a w

Return head + tail rows for compact prompts.

A tiny but realistic CSV with mixed dtypes and a missing value.

Return at most ``n`` rows from ``df`` using stratification if possible.

Yield incremental text chunks as the model generates.