/**
 * The **Data Analysis Service** — computes statistics over the grid's live rows
 * so Photon AI can answer analytical questions ("what sells best?", "why is
 * revenue falling?", "show unusual values") with real numbers.
 *
 * ### Privacy is the point
 * Every statistic here is computed **in the browser, from `GridApi`**. Only the
 * resulting compact summary — counts, sums, a handful of top categories — is
 * ever sent to the language model, which does nothing but phrase it. Raw
 * customer rows never leave the page.
 *
 * That choice also makes this work at scale: a million-row dataset costs one
 * linear local pass and a few hundred tokens, where shipping rows to a model
 * would cost far more than any context window allows and still yield wrong
 * totals from a sample.
 *
 * ### Scope
 * These are descriptive statistics, deliberately. Correlation is linear
 * (Pearson) and the trend is a least-squares slope over an ordered measure;
 * neither is a forecast. When a user asks Photon AI to "predict next month",
 * the honest answer is the observed trend plus its direction, and the prompt
 * instructs the model to present it that way rather than inventing a forecast.
 *
 * @packageDocumentation
 */
import type { GridApi } from '../../core/grid-api';
/** Descriptive statistics for one numeric column. */
export interface NumericSummary {
    readonly colId: string;
    readonly header: string;
    readonly count: number;
    readonly sum: number;
    readonly mean: number;
    readonly min: number;
    readonly max: number;
    readonly median: number;
    readonly stdDev: number;
    /** Least-squares slope per row in the current order — positive means rising. */
    readonly trendPerRow: number;
    /** Values further than {@link OUTLIER_Z} standard deviations from the mean. */
    readonly outliers: readonly number[];
}
/** One category's share of a measure. */
export interface CategoryBreakdown {
    readonly value: string;
    readonly count: number;
    readonly total: number;
    readonly share: number;
}
/** A breakdown of one measure across one dimension. */
export interface DimensionSummary {
    readonly dimensionColId: string;
    readonly dimensionHeader: string;
    readonly measureColId: string;
    readonly measureHeader: string;
    readonly distinctCount: number;
    readonly top: readonly CategoryBreakdown[];
    readonly bottom: readonly CategoryBreakdown[];
}
/** A linear relationship between two numeric columns. */
export interface CorrelationPair {
    readonly aColId: string;
    readonly bColId: string;
    /** Pearson coefficient in [-1, 1]. */
    readonly r: number;
}
/** The complete local analysis handed to the model. */
export interface DatasetAnalysis {
    readonly totalRows: number;
    readonly visibleRows: number;
    readonly columnCount: number;
    readonly numeric: readonly NumericSummary[];
    readonly dimensions: readonly DimensionSummary[];
    readonly correlations: readonly CorrelationPair[];
    /** Columns where some rows have no value, as `colId` → missing count. */
    readonly missing: Readonly<Record<string, number>>;
}
/** Computes descriptive statistics over the grid's currently visible rows. */
export declare class DataAnalysisService {
    private readonly api;
    constructor(api: GridApi);
    /**
     * Analyses the rows the user is currently looking at — i.e. after filters.
     * That is almost always what an analytical question means: "what sells best"
     * asked with a region filter applied is a question about that region.
     */
    analyze(): DatasetAnalysis;
    /** Descriptive stats for one numeric column, or `null` when it has no usable values. */
    private summarizeNumeric;
    /**
     * Breaks each numeric measure down by each low-cardinality dimension.
     *
     * This is what answers "best selling product" and "which customers spend the
     * most": the top entries of (product × revenue) and (customer × spend).
     * Capped so a wide grid cannot produce a combinatorial explosion of pairs.
     */
    private buildDimensions;
    /** Pearson correlation for every numeric pair, strongest first. */
    private buildCorrelations;
    /** Counts empty values per column, omitting fully-populated ones. */
    private countMissing;
}
//# sourceMappingURL=data-analysis-service.d.ts.map