@theoremjs/react 0.3.0 → 0.3.1

package/client/embedding-input.js

@@ -0,0 +1,194 @@
+/**
+ * The rules behind an embedding console: the inputs as the JSON a request
+ * sends, text split into documents, documents packed into the requests the
+ * profile allows, all checked the way the kernel checks them, and documents
+ * ranked by how alike they point to a query.
+ *
+ * @module
+ */
+import { base64Bytes, embeddingMediaIssues, } from '@theoremjs/agents';
+function isPart(value) {
+    if (typeof value !== 'object' || value === null)
+        return false;
+    const part = value;
+    if (part.type === 'text')
+        return typeof part.text === 'string';
+    if (typeof part.mimeType !== 'string')
+        return false;
+    if (part.type === 'media')
+        return typeof part.data === 'string';
+    return part.type === 'media-ref' && typeof part.uri === 'string';
+}
+function isInput(value) {
+    if (typeof value === 'string')
+        return true;
+    if (typeof value !== 'object' || value === null)
+        return false;
+    const parts = value.parts;
+    return Array.isArray(parts) && parts.every(isPart);
+}
+/** The request's inputs, read from its JSON; `undefined` when it is not a list of inputs. */
+export function readEmbeddingInputs(text) {
+    let value;
+    try {
+        value = JSON.parse(text);
+    }
+    catch {
+        return undefined;
+    }
+    return Array.isArray(value) && value.every(isInput) ? value : undefined;
+}
+function partsOf(input) {
+    return typeof input === 'string' ? [{ type: 'text', text: input }] : input.parts;
+}
+/** One input, read from its JSON; `undefined` when it is not one. */
+export function readEmbeddingInput(text) {
+    let value;
+    try {
+        value = JSON.parse(text);
+    }
+    catch {
+        return undefined;
+    }
+    return isInput(value) ? value : undefined;
+}
+/** The input as fields, or `null` when it holds more than one text part and fields would lose it. */
+export function inputFields(input) {
+    const parts = partsOf(input);
+    const texts = parts.filter((part) => part.type === 'text');
+    if (texts.length > 1)
+        return null;
+    return {
+        text: texts[0]?.text ?? '',
+        media: parts.filter((part) => part.type !== 'text'),
+    };
+}
+/** Fields as the input they send: plain text alone, else its parts with the text first. */
+export function fieldsInput(fields) {
+    if (fields.media.length === 0)
+        return fields.text;
+    return {
+        parts: [
+            ...(fields.text ? [{ type: 'text', text: fields.text }] : []),
+            ...fields.media,
+        ],
+    };
+}
+function mediaFacts(part) {
+    return part.type === 'media'
+        ? { mimeType: part.mimeType, sizeBytes: base64Bytes(part.data) }
+        : { mimeType: part.mimeType };
+}
+/** The inputs, held to the profile's rules and its model's limits as the kernel holds them in one request. */
+export function checkInputs(inputs, rules) {
+    const list = [...inputs];
+    if (list.length === 0)
+        return { ok: false, reason: 'no_inputs' };
+    if (rules.maxInputs !== undefined && list.length > rules.maxInputs)
+        return { ok: false, reason: 'too_many_inputs', max: rules.maxInputs, inputs: list };
+    const media = [];
+    for (const [index, input] of list.entries()) {
+        const parts = partsOf(input);
+        if (parts.length === 0 || parts.every((part) => part.type === 'text' && !part.text.trim()))
+            return { ok: false, reason: 'empty_input', index, inputs: list };
+        if (!rules.text && parts.some((part) => part.type === 'text' && part.text.trim()))
+            return { ok: false, reason: 'text_off', index, inputs: list };
+        const files = [];
+        for (const part of parts) {
+            if (part.type === 'text')
+                continue;
+            try {
+                files.push(mediaFacts(part));
+            }
+            catch {
+                return { ok: false, reason: 'invalid_json' };
+            }
+        }
+        media.push(files);
+    }
+    const issues = embeddingMediaIssues(rules, media);
+    return issues.length > 0
+        ? { ok: false, reason: 'attachments', issues, inputs: list }
+        : { ok: true, inputs: list };
+}
+/** The request's JSON as the inputs it sends, held to the profile's rules. */
+export function checkEmbeddingInputs(text, rules) {
+    const inputs = readEmbeddingInputs(text);
+    return inputs === undefined ? { ok: false, reason: 'invalid_json' } : checkInputs(inputs, rules);
+}
+/** The text as documents, each trimmed, none empty. */
+export function splitText(text, split) {
+    const pieces = split === 'whole'
+        ? [text]
+        : split === 'lines'
+            ? text.split(/\r?\n/)
+            : split === 'paragraphs'
+                ? text.split(/\r?\n[^\S\r\n]*\r?\n/)
+                : split === 'commas'
+                    ? text.split(',')
+                    : split.separator
+                        ? text.split(split.separator)
+                        : [text];
+    return pieces.map((piece) => piece.trim()).filter((piece) => piece !== '');
+}
+/** A file read as words and split, not sent as media: plain text, Markdown, CSV, and JSON. */
+export function readsAsText(file) {
+    return (file.type.startsWith('text/') ||
+        file.type === 'application/json' ||
+        /\.(txt|md|markdown|csv|tsv|json)$/i.test(file.name));
+}
+// why: OpenRouter caps no batch (it took 101 inputs); 100 is Gemini's cap, a size every provider checked carries.
+/** The most inputs the console sends in one request when the model sets no cap. */
+export const MAX_BATCH_INPUTS = 100;
+/**
+ * The inputs, in order, as the fewest requests the profile and its model
+ * allow: each holds at most `most` inputs and passes {@linkcode checkInputs}.
+ * An input refused on its own goes alone, for the kernel to refuse.
+ */
+export function embeddingBatches(inputs, rules, most = MAX_BATCH_INPUTS) {
+    const batches = [];
+    let current = [];
+    for (const input of inputs) {
+        const next = [...current, input];
+        if (current.length > 0 && (next.length > most || !checkInputs(next, rules).ok)) {
+            batches.push(current);
+            current = [input];
+        }
+        else
+            current = next;
+    }
+    if (current.length > 0)
+        batches.push(current);
+    return batches;
+}
+/** Two calls' usage as one: token counts and costs added, estimated when either was. */
+export function addUsage(a, b) {
+    if (!a || !b)
+        return a ?? b;
+    const sum = (x, y) => x === undefined && y === undefined ? undefined : (x ?? 0) + (y ?? 0);
+    const inputTokens = sum(a.inputTokens, b.inputTokens);
+    const costUsd = sum(a.costUsd, b.costUsd);
+    return {
+        ...(inputTokens === undefined ? {} : { inputTokens }),
+        ...(costUsd === undefined ? {} : { costUsd }),
+        ...(a.estimated || b.estimated ? { estimated: true } : {}),
+    };
+}
+/** How alike two vectors point, from -1 to 1. */
+export function cosineSimilarity(a, b) {
+    let dot = 0;
+    let aa = 0;
+    let bb = 0;
+    for (let i = 0; i < Math.min(a.length, b.length); i += 1) {
+        dot += a[i] * b[i];
+        aa += a[i] * a[i];
+        bb += b[i] * b[i];
+    }
+    return aa === 0 || bb === 0 ? 0 : dot / Math.sqrt(aa * bb);
+}
+/** The documents, most alike to the query first; a vector of another width is left out. */
+export function rankMatches(query, documents) {
+    return documents
+        .flatMap((values, index) => values.length === query.length ? [{ index, score: cosineSimilarity(query, values) }] : [])
+        .sort((a, b) => b.score - a.score);
+}

package/client/embedding-space.d.ts

@@ -0,0 +1,82 @@
+/**
+ * The index as a space: each document a point in a ball, placed by the
+ * three directions its vectors differ most along, each measured in its own
+ * spread so the space is as deep as it is wide. Its direction from the
+ * centre is its direction along those three; its distance is how far it sits
+ * from the index's mean along them, as `r / (r + median)`, scaled so the
+ * farthest document is on the rim: the order of distances is kept, and no
+ * one far document squeezes the rest into the centre. A pair is alike when
+ * its cosine similarity stands out against every other pair in the index:
+ * more than {@link OUTLIER} robust deviations above the median pair. Most
+ * pairs in an index are unrelated, so the median and its deviation say what
+ * unrelated looks like for this model and these documents. Each document is
+ * linked to its nearest neighbour, the one its cosine similarity is highest
+ * with, when that pair is alike. A cluster is documents alike pairs connect,
+ * directly or through each other. Positions are a projection and only
+ * approximate; links and clusters are the vectors' own cosine similarity.
+ *
+ * @module
+ */
+/** A point in the space, inside the unit ball. */
+export type SpacePoint = readonly [number, number, number];
+/** One document in the space. */
+export type SpaceDocument = {
+    /** Its index among the documents. */
+    readonly index: number;
+    /** Where it sits in the unit ball. */
+    readonly position: SpacePoint;
+    /** The cluster it belongs to. */
+    readonly cluster: number;
+};
+/** Documents alike pairs connect, directly or through each other. */
+export type SpaceCluster = {
+    /** Its documents' indexes, in order. */
+    readonly members: readonly number[];
+    /** Its hue, from 0 to 1 around the colour wheel: where its members' mean vector lands, as an angle around the space's two widest directions, so neighbouring clusters share a family. */
+    readonly tone: number;
+};
+/** A document and its nearest neighbour, whose cosine similarity is above the space's threshold. */
+export type SpaceLink = {
+    readonly a: number;
+    readonly b: number;
+    /** Their cosine similarity. */
+    readonly similarity: number;
+};
+/** The three directions the space is drawn along, kept so a query lands consistently. */
+export type SpaceBasis = {
+    readonly mean: Float64Array;
+    /** Each a unit direction in the vectors' own space; fewer than three when the documents span fewer. */
+    readonly axes: readonly Float64Array[];
+    /** Each axis's spread over the documents; a point is measured along each in its own spread, so the space is as deep as it is wide. */
+    readonly spreads: readonly number[];
+    /** The documents' median distance from the mean, so measured. */
+    readonly scale: number;
+    /** The farthest document's `r / (r + scale)`; dividing by it puts that document on the rim. */
+    readonly reach: number;
+};
+export type EmbeddingSpace = {
+    readonly documents: readonly SpaceDocument[];
+    readonly clusters: readonly SpaceCluster[];
+    /** Each document's nearest neighbour, where that pair is above the threshold; a pair that is each other's nearest appears once. */
+    readonly links: readonly SpaceLink[];
+    /** The cosine similarity a pair must pass to be linked; `Infinity` when the index has too few pairs, or too alike, to say what unrelated looks like. */
+    readonly threshold: number;
+    readonly basis: SpaceBasis;
+    /** How wide the vectors are. */
+    readonly dimensions: number;
+};
+/**
+ * The similarity one must pass to stand out among `similarities`: their
+ * median plus {@link OUTLIER} robust deviations. `Infinity` when they have no
+ * spread, since then nothing stands out.
+ */
+export declare function outlierThreshold(similarities: ArrayLike<number>): number;
+/** The space for a set of document vectors, all at once. */
+export declare function layoutSpace(vectors: readonly (readonly number[])[]): EmbeddingSpace;
+/**
+ * The same space, laid out in slices between which the page goes on, so a
+ * large index never freezes it. Rejects with the signal's reason once aborted.
+ */
+export declare function layoutSpaceSoon(vectors: readonly (readonly number[])[], signal?: AbortSignal): Promise<EmbeddingSpace>;
+/** Where a query lands in the space: projected along the same directions as the documents. `null` when its width differs. */
+export declare function placeQuery(space: EmbeddingSpace, query: readonly number[]): SpacePoint | null;

package/client/embedding-space.js

@@ -0,0 +1,389 @@
+/**
+ * The index as a space: each document a point in a ball, placed by the
+ * three directions its vectors differ most along, each measured in its own
+ * spread so the space is as deep as it is wide. Its direction from the
+ * centre is its direction along those three; its distance is how far it sits
+ * from the index's mean along them, as `r / (r + median)`, scaled so the
+ * farthest document is on the rim: the order of distances is kept, and no
+ * one far document squeezes the rest into the centre. A pair is alike when
+ * its cosine similarity stands out against every other pair in the index:
+ * more than {@link OUTLIER} robust deviations above the median pair. Most
+ * pairs in an index are unrelated, so the median and its deviation say what
+ * unrelated looks like for this model and these documents. Each document is
+ * linked to its nearest neighbour, the one its cosine similarity is highest
+ * with, when that pair is alike. A cluster is documents alike pairs connect,
+ * directly or through each other. Positions are a projection and only
+ * approximate; links and clusters are the vectors' own cosine similarity.
+ *
+ * @module
+ */
+/** Power-iteration rounds per axis; the Gram matrix is small, so this converges well inside it. */
+const ROUNDS = 200;
+/** Where a point lands when the projection says nothing about it. */
+const CENTRE = [0, 0, 0];
+/** How many robust deviations above the median pair a pair must be to count as alike: the usual cut for an outlier. */
+const OUTLIER = 3;
+/** Scales the median absolute deviation to a standard deviation, for normally spread values. */
+const MAD_SCALE = 1.4826;
+function dot(a, b, length) {
+    let sum = 0;
+    for (let i = 0; i < length; i += 1)
+        sum += a[i] * b[i];
+    return sum;
+}
+/** A point's offset from the mean along each axis, in that axis's own spread; 0 along an axis with none. */
+function along(axes, mean, spreads, vector) {
+    const offset = Array.from(vector, (v, d) => v - mean[d]);
+    return [0, 1, 2].map((k) => axes[k] && spreads[k] > 1e-12 ? dot(axes[k], offset, offset.length) / spreads[k] : 0);
+}
+/** Each axis's spread: the root mean square of the documents' offsets along it. */
+function spreadsOf(axes, mean, vectors) {
+    return [0, 1, 2].map((k) => {
+        const axis = axes[k];
+        if (!axis || vectors.length === 0)
+            return 0;
+        const squares = vectors.reduce((sum, v) => {
+            const at = dot(axis, Array.from(v, (x, d) => x - mean[d]), v.length);
+            return sum + at * at;
+        }, 0);
+        return Math.sqrt(squares / vectors.length);
+    });
+}
+/** Coordinates in the axes, drawn into the ball: same direction, distance `r / (r + scale) / reach`. */
+function inBall(point, scale, reach) {
+    const length = Math.hypot(point[0], point[1], point[2]);
+    if (length < 1e-9 || scale <= 0 || reach <= 0)
+        return CENTRE;
+    const radius = length / (length + scale) / reach / length;
+    return [point[0] * radius, point[1] * radius, point[2] * radius];
+}
+/** The documents packed row after row, so the loops over them stay on one typed array. */
+function packed(vectors, dimensions) {
+    const rows = new Float64Array(vectors.length * dimensions);
+    vectors.forEach((vector, i) => {
+        rows.set(vector, i * dimensions);
+    });
+    return rows;
+}
+/** Every document against every other: the raw dot products, once. Yields after each row, so a large index can share the thread. */
+function* gramOf(rows, n, dimensions) {
+    const gram = new Float64Array(n * n);
+    const tail = dimensions - (dimensions % 4);
+    for (let i = 0; i < n; i += 1) {
+        const a = i * dimensions;
+        for (let j = i; j < n; j += 1) {
+            const b = j * dimensions;
+            let s0 = 0;
+            let s1 = 0;
+            let s2 = 0;
+            let s3 = 0;
+            // why: four running sums let the engine pipeline the multiplies; this loop is most of the layout's time.
+            for (let d = 0; d < tail; d += 4) {
+                s0 += rows[a + d] * rows[b + d];
+                s1 += rows[a + d + 1] * rows[b + d + 1];
+                s2 += rows[a + d + 2] * rows[b + d + 2];
+                s3 += rows[a + d + 3] * rows[b + d + 3];
+            }
+            for (let d = tail; d < dimensions; d += 1)
+                s0 += rows[a + d] * rows[b + d];
+            const sum = s0 + s1 + s2 + s3;
+            gram[i * n + j] = sum;
+            gram[j * n + i] = sum;
+        }
+        yield;
+    }
+    return gram;
+}
+/** The same products about the documents' mean, from the raw ones. */
+function centred(gram, n) {
+    const rows = new Float64Array(n);
+    let all = 0;
+    for (let i = 0; i < n; i += 1) {
+        for (let j = 0; j < n; j += 1)
+            rows[i] += gram[i * n + j];
+        rows[i] /= n;
+        all += rows[i];
+    }
+    all /= n;
+    const out = new Float64Array(n * n);
+    for (let i = 0; i < n; i += 1)
+        for (let j = 0; j < n; j += 1)
+            out[i * n + j] = gram[i * n + j] - rows[i] - rows[j] + all;
+    return out;
+}
+/** One power-iteration round: `matrix · vector`, normalised, with its length and how far it moved. */
+function powerRound(matrix, n, vector) {
+    const next = new Float64Array(n);
+    let length = 0;
+    for (let i = 0; i < n; i += 1) {
+        let sum = 0;
+        for (let j = 0; j < n; j += 1)
+            sum += matrix[i * n + j] * vector[j];
+        next[i] = sum;
+        length += sum * sum;
+    }
+    length = Math.sqrt(length);
+    if (length < 1e-12)
+        return { next: vector, length: 0, moved: 0 };
+    let moved = 0;
+    for (let i = 0; i < n; i += 1) {
+        next[i] /= length;
+        moved = Math.max(moved, Math.abs(next[i] - vector[i]));
+    }
+    return { next, length, moved };
+}
+/** Flips a vector so its largest entry is positive. */
+function signed(vector) {
+    let lead = 0;
+    for (let i = 1; i < vector.length; i += 1)
+        if (Math.abs(vector[i]) > Math.abs(vector[lead]))
+            lead = i;
+    return vector[lead] < 0 ? vector.map((v) => -v) : vector;
+}
+/** The largest eigenvector left in a matrix, by power iteration. */
+function* largestEigen(matrix, n, k) {
+    // invariant: a fixed start, so the same documents always draw the same space.
+    let vector = Float64Array.from({ length: n }, (_, i) => 1 + ((i * 7919 + k * 104729) % 97) / 97);
+    let value = 0;
+    for (let round = 0; round < ROUNDS; round += 1) {
+        const step = powerRound(matrix, n, vector);
+        if (step.length === 0)
+            break;
+        value = step.length;
+        vector = step.next;
+        if (step.moved < 1e-12)
+            break;
+        yield;
+    }
+    return { vector: signed(vector), value };
+}
+/** The largest eigenvectors of a symmetric matrix, by power iteration with deflation, each signed so its largest entry is positive. */
+function* topEigen(matrix, n, count) {
+    const work = Float64Array.from(matrix);
+    const found = [];
+    const scale = Math.max(...Array.from({ length: n }, (_, i) => Math.abs(work[i * n + i])), 0);
+    for (let k = 0; k < Math.min(count, n); k += 1) {
+        const { vector, value } = yield* largestEigen(work, n, k);
+        if (value <= scale * 1e-9)
+            break;
+        found.push({ vector, value });
+        for (let i = 0; i < n; i += 1)
+            for (let j = 0; j < n; j += 1)
+                work[i * n + j] -= value * vector[i] * vector[j];
+    }
+    return found;
+}
+function meanOf(rows, n, dimensions) {
+    const mean = new Float64Array(dimensions);
+    for (let i = 0; i < n; i += 1)
+        for (let d = 0; d < dimensions; d += 1)
+            mean[d] += rows[i * dimensions + d];
+    for (let d = 0; d < dimensions; d += 1)
+        mean[d] /= Math.max(n, 1);
+    return mean;
+}
+/** One axis in the vectors' own space: the documents weighted by an eigenvector of their Gram matrix. */
+function axisOf(rows, mean, weights) {
+    const dimensions = mean.length;
+    const axis = new Float64Array(dimensions);
+    for (let i = 0; i < weights.length; i += 1)
+        for (let d = 0; d < dimensions; d += 1)
+            axis[d] += weights[i] * (rows[i * dimensions + d] - mean[d]);
+    let length = 0;
+    for (let d = 0; d < dimensions; d += 1)
+        length += axis[d] * axis[d];
+    length = Math.sqrt(length);
+    if (length > 0)
+        for (let d = 0; d < dimensions; d += 1)
+            axis[d] /= length;
+    return axis;
+}
+/** Every pair's cosine similarity, from the raw products. */
+function cosineOf(gram, n) {
+    const norms = Array.from({ length: n }, (_, i) => Math.sqrt(Math.max(gram[i * n + i], 0)));
+    const cosine = new Float64Array(n * n);
+    for (let i = 0; i < n; i += 1)
+        for (let j = 0; j < n; j += 1)
+            cosine[i * n + j] =
+                norms[i] === 0 || norms[j] === 0 ? 0 : gram[i * n + j] / (norms[i] * norms[j]);
+    return cosine;
+}
+/** The median of some values; it sorts them. */
+function medianOf(values) {
+    values.sort();
+    const middle = values.length >> 1;
+    return values.length % 2 ? values[middle] : (values[middle - 1] + values[middle]) / 2;
+}
+/**
+ * The similarity one must pass to stand out among `similarities`: their
+ * median plus {@link OUTLIER} robust deviations. `Infinity` when they have no
+ * spread, since then nothing stands out.
+ */
+export function outlierThreshold(similarities) {
+    const values = Float64Array.from(similarities);
+    if (values.length === 0)
+        return Infinity;
+    const median = medianOf(values.slice());
+    const deviation = MAD_SCALE * medianOf(values.map((v) => Math.abs(v - median)));
+    return deviation > 0 ? median + OUTLIER * deviation : Infinity;
+}
+/** Every pair's similarity, once each. */
+function pairsOf(cosine, n) {
+    const pairs = new Float64Array((n * (n - 1)) / 2);
+    let at = 0;
+    for (let i = 0; i < n; i += 1)
+        for (let j = i + 1; j < n; j += 1)
+            pairs[at++] = cosine[i * n + j];
+    return pairs;
+}
+/** Every pair above the threshold, a row at a time: what clusters are made of. */
+function* linksOf(cosine, n, threshold) {
+    const links = [];
+    for (let i = 0; i < n; i += 1) {
+        for (let j = i + 1; j < n; j += 1)
+            if (cosine[i * n + j] > threshold)
+                links.push({ a: i, b: j, similarity: cosine[i * n + j] });
+        yield;
+    }
+    return links;
+}
+/** The document whose cosine similarity with `i` is highest, the first on a tie; -1 when there is no other. */
+function nearestTo(cosine, n, i) {
+    let best = -1;
+    for (let j = 0; j < n; j += 1)
+        if (j !== i && (best < 0 || cosine[i * n + j] > cosine[i * n + best]))
+            best = j;
+    return best;
+}
+/** Each document's nearest neighbour where that pair passes the threshold, each pair once, a row at a time. */
+function* nearestOf(cosine, n, threshold) {
+    const seen = new Set();
+    const links = [];
+    for (let i = 0; i < n; i += 1) {
+        const best = nearestTo(cosine, n, i);
+        const [a, b] = i < best ? [i, best] : [best, i];
+        if (best >= 0 && cosine[i * n + best] > threshold && !seen.has(a * n + b)) {
+            seen.add(a * n + b);
+            links.push({ a, b, similarity: cosine[i * n + best] });
+        }
+        yield;
+    }
+    return links.sort((x, y) => x.a - y.a || x.b - y.b);
+}
+/** The groups the links connect, each in document order. */
+function groupsOf(n, links) {
+    const root = Array.from({ length: n }, (_, i) => i);
+    const find = (i) => {
+        while (root[i] !== i) {
+            root[i] = root[root[i]];
+            i = root[i];
+        }
+        return i;
+    };
+    for (const link of links)
+        root[find(link.a)] = find(link.b);
+    const groups = new Map();
+    for (let i = 0; i < n; i += 1) {
+        const top = find(i);
+        const group = groups.get(top);
+        if (group)
+            group.push(i);
+        else
+            groups.set(top, [i]);
+    }
+    return [...groups.values()];
+}
+/** The mean of some rows. */
+function meanOfRows(rows, members, dimensions) {
+    const mean = new Array(dimensions).fill(0);
+    for (const i of members)
+        for (let d = 0; d < dimensions; d += 1)
+            mean[d] += rows[i * dimensions + d];
+    return mean.map((v) => v / Math.max(members.length, 1));
+}
+/** A position's hue: its angle around the space's two widest directions, from 0 to 1. */
+function toneOf(position) {
+    return (Math.atan2(position[1], position[0]) / (2 * Math.PI) + 1) % 1;
+}
+/**
+ * The layout, a slice at a time. Vectors whose width differs from the first
+ * are placed at the centre, each a cluster of its own.
+ */
+function* layoutSteps(vectors) {
+    const dimensions = vectors[0]?.length ?? 0;
+    const original = vectors.flatMap((vector, i) => (vector.length === dimensions ? [i] : []));
+    const n = original.length;
+    const rows = packed(original.map((i) => vectors[i]), dimensions);
+    const mean = meanOf(rows, n, dimensions);
+    const gram = yield* gramOf(rows, n, dimensions);
+    const axes = (yield* topEigen(centred(gram, n), n, 3)).map(({ vector }) => axisOf(rows, mean, vector));
+    const kept = original.map((i) => vectors[i]);
+    const spreads = spreadsOf(axes, mean, kept);
+    const coordinates = kept.map((vector) => along(axes, mean, spreads, vector));
+    const lengths = Float64Array.from(coordinates, (p) => Math.hypot(p[0], p[1], p[2]));
+    // why: when more than half sit at the mean, the farthest sets the scale, so the rest still spread out.
+    const median = n > 0 ? medianOf(lengths.slice()) : 0;
+    const scale = median > 1e-9 ? median : lengths.reduce((most, v) => Math.max(most, v), 0);
+    const reach = lengths.reduce((most, r) => Math.max(most, scale > 0 ? r / (r + scale) : 0), 0);
+    const basis = { mean, axes, spreads, scale, reach };
+    const cosine = cosineOf(gram, n);
+    const threshold = outlierThreshold(pairsOf(cosine, n));
+    const found = yield* linksOf(cosine, n, threshold);
+    const groups = groupsOf(n, found);
+    const nearest = yield* nearestOf(cosine, n, threshold);
+    const clusters = groups.map((group) => ({
+        members: group.map((i) => original[i]),
+        tone: toneOf(project(basis, meanOfRows(rows, group, dimensions))),
+    }));
+    for (const [i, vector] of vectors.entries())
+        if (vector.length !== dimensions)
+            clusters.push({ members: [i], tone: toneOf(CENTRE) });
+    const clusterOf = new Map(clusters.flatMap((c, k) => c.members.map((i) => [i, k])));
+    const documents = vectors.map((vector, index) => ({
+        index,
+        // invariant: documents and queries go through the one projection, so a query lands on its own document.
+        position: vector.length === dimensions ? project(basis, vector) : CENTRE,
+        cluster: clusterOf.get(index) ?? 0,
+    }));
+    const links = nearest.map((link) => ({ ...link, a: original[link.a], b: original[link.b] }));
+    return { documents, clusters, links, threshold, basis, dimensions };
+}
+/** The space for a set of document vectors, all at once. */
+export function layoutSpace(vectors) {
+    const steps = layoutSteps(vectors);
+    for (;;) {
+        const step = steps.next();
+        if (step.done)
+            return step.value;
+    }
+}
+/** Time the layout works before it lets the page go on, in milliseconds. */
+const SLICE_MS = 8;
+/**
+ * The same space, laid out in slices between which the page goes on, so a
+ * large index never freezes it. Rejects with the signal's reason once aborted.
+ */
+export async function layoutSpaceSoon(vectors, signal) {
+    const steps = layoutSteps(vectors);
+    for (;;) {
+        const until = performance.now() + SLICE_MS;
+        for (;;) {
+            const step = steps.next();
+            if (step.done)
+                return step.value;
+            if (performance.now() >= until)
+                break;
+        }
+        await new Promise((resolve) => setTimeout(resolve, 0));
+        signal?.throwIfAborted();
+    }
+}
+/** Where a query lands in the space: projected along the same directions as the documents. `null` when its width differs. */
+export function placeQuery(space, query) {
+    if (query.length !== space.dimensions || space.dimensions === 0)
+        return null;
+    return project(space.basis, query);
+}
+function project(basis, vector) {
+    return inBall(along(basis.axes, basis.mean, basis.spreads, vector), basis.scale, basis.reach);
+}