markit-ai 0.2.0 → 0.3.0

package/dist/converters/pdf/columns.d.ts

@@ -0,0 +1,35 @@
+/**
+ * Multi-column layout detection and text box reordering.
+ *
+ * Many PDFs (legal documents, datasheets, academic papers) use two-column
+ * layouts. Without column detection, text boxes are ordered by Y position
+ * only, interleaving left and right column content.
+ *
+ * Algorithm:
+ *   1. Collect left edges of all text boxes on the page
+ *   2. Find the largest horizontal gap between consecutive left edges
+ *   3. If gap > MIN_GAP_RATIO of the text width and both sides have
+ *      enough boxes → multi-column detected
+ *   4. Assign each text box to a column based on its center X
+ *   5. Return columns in reading order (left-to-right, top-to-bottom)
+ *
+ * This only detects the column structure. The caller is responsible for
+ * processing each column's text boxes independently (table detection,
+ * rendering, etc.).
+ */
+import type { TextBox } from "./types.js";
+export interface ColumnLayout {
+    /** Number of columns detected (1 = single column, 2+ = multi-column). */
+    columnCount: number;
+    /** Text boxes grouped by column, in reading order (left to right). */
+    columns: TextBox[][];
+    /** X positions of column boundaries (between columns). */
+    boundaries: number[];
+}
+/**
+ * Detect column layout and return text boxes grouped by column.
+ *
+ * For single-column pages, returns all boxes in one group.
+ * For multi-column pages, returns boxes split by column in reading order.
+ */
+export declare function detectColumns(textBoxes: TextBox[]): ColumnLayout;

package/dist/converters/pdf/columns.js

@@ -0,0 +1,93 @@
+/**
+ * Multi-column layout detection and text box reordering.
+ *
+ * Many PDFs (legal documents, datasheets, academic papers) use two-column
+ * layouts. Without column detection, text boxes are ordered by Y position
+ * only, interleaving left and right column content.
+ *
+ * Algorithm:
+ *   1. Collect left edges of all text boxes on the page
+ *   2. Find the largest horizontal gap between consecutive left edges
+ *   3. If gap > MIN_GAP_RATIO of the text width and both sides have
+ *      enough boxes → multi-column detected
+ *   4. Assign each text box to a column based on its center X
+ *   5. Return columns in reading order (left-to-right, top-to-bottom)
+ *
+ * This only detects the column structure. The caller is responsible for
+ * processing each column's text boxes independently (table detection,
+ * rendering, etc.).
+ */
+/**
+ * Minimum gap as a fraction of the total text width to consider a column
+ * boundary. A two-column layout typically has ~50% gap; we use a lower
+ * threshold to catch asymmetric columns.
+ */
+const MIN_GAP_RATIO = 0.15;
+/** Minimum number of text boxes on each side of the gap. */
+const MIN_BOXES_PER_COLUMN = 4;
+/** Minimum gap in absolute points to avoid splitting on small whitespace. */
+const MIN_GAP_PTS = 40;
+/**
+ * Detect column layout and return text boxes grouped by column.
+ *
+ * For single-column pages, returns all boxes in one group.
+ * For multi-column pages, returns boxes split by column in reading order.
+ */
+export function detectColumns(textBoxes) {
+    if (textBoxes.length < MIN_BOXES_PER_COLUMN * 2) {
+        return { columnCount: 1, columns: [textBoxes], boundaries: [] };
+    }
+    // Collect unique left edges (rounded to avoid float noise)
+    const lefts = [
+        ...new Set(textBoxes.map((tb) => Math.round(tb.bounds.left))),
+    ].sort((a, b) => a - b);
+    if (lefts.length < 2) {
+        return { columnCount: 1, columns: [textBoxes], boundaries: [] };
+    }
+    const textXMin = lefts[0];
+    const textXMax = Math.max(...textBoxes.map((tb) => Math.round(tb.bounds.right)));
+    const textWidth = textXMax - textXMin;
+    if (textWidth <= 0) {
+        return { columnCount: 1, columns: [textBoxes], boundaries: [] };
+    }
+    // Find the largest gap between consecutive left-edge positions
+    let maxGap = 0;
+    let gapLeft = 0;
+    let gapRight = 0;
+    for (let i = 1; i < lefts.length; i++) {
+        const gap = lefts[i] - lefts[i - 1];
+        if (gap > maxGap) {
+            maxGap = gap;
+            gapLeft = lefts[i - 1];
+            gapRight = lefts[i];
+        }
+    }
+    const gapRatio = maxGap / textWidth;
+    if (gapRatio < MIN_GAP_RATIO || maxGap < MIN_GAP_PTS) {
+        return { columnCount: 1, columns: [textBoxes], boundaries: [] };
+    }
+    // Split point is the midpoint of the gap
+    const splitX = (gapLeft + gapRight) / 2;
+    // Assign boxes to columns based on center X
+    const leftCol = [];
+    const rightCol = [];
+    for (const tb of textBoxes) {
+        const cx = (tb.bounds.left + tb.bounds.right) / 2;
+        if (cx < splitX) {
+            leftCol.push(tb);
+        }
+        else {
+            rightCol.push(tb);
+        }
+    }
+    // Validate both columns have enough content
+    if (leftCol.length < MIN_BOXES_PER_COLUMN ||
+        rightCol.length < MIN_BOXES_PER_COLUMN) {
+        return { columnCount: 1, columns: [textBoxes], boundaries: [] };
+    }
+    return {
+        columnCount: 2,
+        columns: [leftCol, rightCol],
+        boundaries: [splitX],
+    };
+}

package/dist/converters/pdf/extract.d.ts

@@ -0,0 +1,19 @@
+/**
+ * PDF content extraction using mupdf.
+ *
+ * Extracts text boxes (with position, font size, bold) and vector line
+ * segments (table borders) from each page. Uses mupdf's native WASM
+ * engine for fast parsing, and reads raw content streams for vector graphics.
+ *
+ * Coordinate system: PDF native (origin = bottom-left, Y increases upward).
+ */
+import type { ImageRegion, PageContent } from "./types.js";
+/**
+ * Render an image region from a PDF page as a PNG buffer.
+ * Uses mupdf's DrawDevice to render just the cropped area at 2x resolution.
+ */
+export declare function renderImageRegion(input: Uint8Array, region: ImageRegion): Uint8Array;
+/**
+ * Extract text boxes and vector segments from all pages of a PDF buffer.
+ */
+export declare function extractPages(input: Uint8Array): Promise<PageContent[]>;

package/dist/converters/pdf/extract.js

@@ -0,0 +1,513 @@
+/**
+ * PDF content extraction using mupdf.
+ *
+ * Extracts text boxes (with position, font size, bold) and vector line
+ * segments (table borders) from each page. Uses mupdf's native WASM
+ * engine for fast parsing, and reads raw content streams for vector graphics.
+ *
+ * Coordinate system: PDF native (origin = bottom-left, Y increases upward).
+ */
+// ---------------------------------------------------------------------------
+// Text extraction
+// ---------------------------------------------------------------------------
+/** Y tolerance for merging text fragments on the same visual line. */
+const SAME_LINE_Y_TOLERANCE = 2;
+/** Max horizontal gap (pts) to merge adjacent fragments into one text box. */
+const MAX_MERGE_GAP = 14;
+/**
+ * Merge horizontally adjacent raw text items on the same visual line into
+ * word/phrase-level text boxes.
+ */
+function mergeIntoWords(raws) {
+    if (raws.length === 0)
+        return [];
+    // Sort by Y descending (top-first in bottom-left coords), then X ascending
+    const sorted = [...raws].sort((a, b) => {
+        const dy = b.y - a.y;
+        return Math.abs(dy) > SAME_LINE_Y_TOLERANCE ? dy : a.x - b.x;
+    });
+    const merged = [];
+    let cur = { ...sorted[0] };
+    for (let i = 1; i < sorted.length; i++) {
+        const next = sorted[i];
+        const sameY = Math.abs(next.y - cur.y) <= SAME_LINE_Y_TOLERANCE;
+        const close = next.x <= cur.x + cur.width + MAX_MERGE_GAP;
+        if (sameY && close) {
+            const gap = next.x - (cur.x + cur.width);
+            const sep = gap > 1 ? " " : "";
+            cur.text += sep + next.text;
+            cur.width = next.x + next.width - cur.x;
+            cur.height = Math.max(cur.height, next.height);
+            cur.fontSize = Math.max(cur.fontSize, next.fontSize);
+            cur.isBold = cur.isBold || next.isBold;
+        }
+        else {
+            merged.push(cur);
+            cur = { ...next };
+        }
+    }
+    merged.push(cur);
+    return merged;
+}
+/**
+ * Extract text boxes from a mupdf page using structured text output.
+ *
+ * mupdf's structured text JSON uses top-left origin; we convert to
+ * bottom-left (standard PDF coordinates) using the page height.
+ */
+function extractTextBoxes(page, pageNumber, pageHeight, stext) {
+    if (!stext) {
+        stext = JSON.parse(page.toStructuredText("preserve-whitespace").asJSON());
+    }
+    const raws = [];
+    for (const block of stext.blocks) {
+        if (block.type !== "text")
+            continue;
+        for (const line of block.lines) {
+            const text = line.text?.trim();
+            if (!text)
+                continue;
+            const fontSize = line.font?.size ?? 0;
+            const weight = line.font?.weight ?? "normal";
+            const fontName = line.font?.name ?? "";
+            const isBold = weight === "bold" ||
+                /bold/i.test(fontName) ||
+                /Black|Heavy/i.test(fontName);
+            // mupdf bbox: {x, y, w, h} in top-left coords
+            // Convert to bottom-left: pdfY = pageHeight - (bbox.y + bbox.h)
+            const bboxY = line.bbox.y;
+            const bboxH = line.bbox.h;
+            const pdfY = pageHeight - (bboxY + bboxH);
+            raws.push({
+                text,
+                x: line.bbox.x,
+                y: pdfY,
+                width: line.bbox.w,
+                height: bboxH,
+                fontSize,
+                isBold,
+            });
+        }
+    }
+    const words = mergeIntoWords(raws);
+    return words
+        .map((w, i) => ({
+        id: `p${pageNumber}-t${i}`,
+        text: w.text.trim(),
+        pageNumber,
+        fontSize: w.fontSize,
+        isBold: w.isBold,
+        bounds: {
+            left: w.x,
+            right: w.x + w.width,
+            bottom: w.y,
+            top: w.y + w.height,
+        },
+    }))
+        .filter((b) => b.text.length > 0);
+}
+// ---------------------------------------------------------------------------
+// Vector segment extraction from raw content stream
+// ---------------------------------------------------------------------------
+/** Minimum aspect ratio for a filled rect to be considered a line. */
+const LINE_ASPECT_THRESHOLD = 6;
+/** Minimum length (pts) for a segment to count. */
+const MIN_LENGTH = 2;
+/** Maximum thickness (pts) for a border line (filters out filled areas). */
+const MAX_THICKNESS = 3;
+/**
+ * Convert a thin filled rectangle to a horizontal or vertical segment.
+ * Returns null if the rect doesn't look like a border line.
+ */
+function thinRectToSegment(id, x, y, w, h) {
+    const aw = Math.abs(w);
+    const ah = Math.abs(h);
+    if (aw > ah * LINE_ASPECT_THRESHOLD &&
+        aw >= MIN_LENGTH &&
+        ah <= MAX_THICKNESS) {
+        // Horizontal line
+        const cy = y + ah / 2;
+        return { id, x1: x, y1: cy, x2: x + aw, y2: cy };
+    }
+    if (ah > aw * LINE_ASPECT_THRESHOLD &&
+        ah >= MIN_LENGTH &&
+        aw <= MAX_THICKNESS) {
+        // Vertical line
+        const cx = x + aw / 2;
+        return { id, x1: cx, y1: y, x2: cx, y2: y + ah };
+    }
+    return null;
+}
+/**
+ * Emit 4 edge segments from a stroked rectangle.
+ */
+function pushStrokedRectEdges(segments, id, x, y, w, h) {
+    const aw = Math.abs(w);
+    const ah = Math.abs(h);
+    const base = id;
+    if (aw >= MIN_LENGTH) {
+        segments.push({ id: `${base}-b`, x1: x, y1: y, x2: x + aw, y2: y });
+        segments.push({
+            id: `${base}-t`,
+            x1: x,
+            y1: y + ah,
+            x2: x + aw,
+            y2: y + ah,
+        });
+    }
+    if (ah >= MIN_LENGTH) {
+        segments.push({ id: `${base}-l`, x1: x, y1: y, x2: x, y2: y + ah });
+        segments.push({
+            id: `${base}-r`,
+            x1: x + aw,
+            y1: y,
+            x2: x + aw,
+            y2: y + ah,
+        });
+    }
+}
+const CTM_IDENTITY = [1, 0, 0, 1, 0, 0];
+/** Concatenate two affine matrices: result = parent × child. */
+function ctmConcat(p, c) {
+    return [
+        p[0] * c[0] + p[2] * c[1],
+        p[1] * c[0] + p[3] * c[1],
+        p[0] * c[2] + p[2] * c[3],
+        p[1] * c[2] + p[3] * c[3],
+        p[0] * c[4] + p[2] * c[5] + p[4],
+        p[1] * c[4] + p[3] * c[5] + p[5],
+    ];
+}
+function ctmApply(m, x, y) {
+    return [m[0] * x + m[2] * y + m[4], m[1] * x + m[3] * y + m[5]];
+}
+// ---------------------------------------------------------------------------
+// Content stream parsing
+// ---------------------------------------------------------------------------
+/**
+ * Parse a PDF content stream and extract line segments from thin filled
+ * rectangles (re+f), stroked rectangles (re+S), and explicit lines (m/l+S).
+ * Tracks the CTM via q/Q/cm operators so coordinates are in page space.
+ */
+function extractSegmentsFromContentStream(raw, pageNumber) {
+    const segments = [];
+    const tokens = tokenizeContentStream(raw);
+    let idx = 0;
+    let strokeWidth = 1.0;
+    // Graphics state stack (q/Q): saves CTM + strokeWidth
+    let ctm = [...CTM_IDENTITY];
+    const stateStack = [];
+    // State for path building (in user coordinates, pre-CTM)
+    let curX = 0;
+    let curY = 0;
+    let pathStartX = 0;
+    let pathStartY = 0;
+    const pendingRects = [];
+    const pendingLines = [];
+    function flushPath(mode) {
+        const sid = () => `p${pageNumber}-s${segments.length}`;
+        if (mode === "fill") {
+            for (const r of pendingRects) {
+                // Transform the rect corners through CTM, then check if it's a thin line
+                const [x0, y0] = ctmApply(ctm, r.x, r.y);
+                const [x1, y1] = ctmApply(ctm, r.x + r.w, r.y + r.h);
+                const seg = thinRectToSegment(sid(), Math.min(x0, x1), Math.min(y0, y1), Math.abs(x1 - x0), Math.abs(y1 - y0));
+                if (seg)
+                    segments.push(seg);
+            }
+        }
+        else if (mode === "stroke" && strokeWidth <= MAX_THICKNESS) {
+            for (const r of pendingRects) {
+                const [x0, y0] = ctmApply(ctm, r.x, r.y);
+                const [x1, y1] = ctmApply(ctm, r.x + r.w, r.y + r.h);
+                pushStrokedRectEdges(segments, sid(), Math.min(x0, x1), Math.min(y0, y1), Math.abs(x1 - x0), Math.abs(y1 - y0));
+            }
+            for (const l of pendingLines) {
+                const [lx1, ly1] = ctmApply(ctm, l.x1, l.y1);
+                const [lx2, ly2] = ctmApply(ctm, l.x2, l.y2);
+                const dx = Math.abs(lx2 - lx1);
+                const dy = Math.abs(ly2 - ly1);
+                // Only keep H/V lines
+                if ((dx >= MIN_LENGTH && dy < 1) || (dy >= MIN_LENGTH && dx < 1)) {
+                    segments.push({ id: sid(), x1: lx1, y1: ly1, x2: lx2, y2: ly2 });
+                }
+            }
+        }
+        pendingRects.length = 0;
+        pendingLines.length = 0;
+    }
+    while (idx < tokens.length) {
+        const t = tokens[idx];
+        if (t === "q") {
+            stateStack.push({ ctm: [...ctm], strokeWidth });
+        }
+        else if (t === "Q") {
+            const saved = stateStack.pop();
+            if (saved) {
+                ctm = saved.ctm;
+                strokeWidth = saved.strokeWidth;
+            }
+        }
+        else if (t === "cm" && idx >= 6) {
+            const a = Number(tokens[idx - 6]);
+            const b = Number(tokens[idx - 5]);
+            const c = Number(tokens[idx - 4]);
+            const d = Number(tokens[idx - 3]);
+            const e = Number(tokens[idx - 2]);
+            const f = Number(tokens[idx - 1]);
+            ctm = ctmConcat(ctm, [a, b, c, d, e, f]);
+        }
+        else if (t === "w" && idx >= 1) {
+            strokeWidth = Number(tokens[idx - 1]) || strokeWidth;
+        }
+        else if (t === "re" && idx >= 4) {
+            const x = Number(tokens[idx - 4]);
+            const y = Number(tokens[idx - 3]);
+            const w = Number(tokens[idx - 2]);
+            const h = Number(tokens[idx - 1]);
+            if (Number.isFinite(x + y + w + h)) {
+                pendingRects.push({ x, y, w, h });
+            }
+        }
+        else if (t === "m" && idx >= 2) {
+            curX = Number(tokens[idx - 2]);
+            curY = Number(tokens[idx - 1]);
+            pathStartX = curX;
+            pathStartY = curY;
+        }
+        else if (t === "l" && idx >= 2) {
+            const x2 = Number(tokens[idx - 2]);
+            const y2 = Number(tokens[idx - 1]);
+            pendingLines.push({ x1: curX, y1: curY, x2, y2 });
+            curX = x2;
+            curY = y2;
+        }
+        else if (t === "h") {
+            // closePath: line back to start
+            if (curX !== pathStartX || curY !== pathStartY) {
+                pendingLines.push({
+                    x1: curX,
+                    y1: curY,
+                    x2: pathStartX,
+                    y2: pathStartY,
+                });
+            }
+            curX = pathStartX;
+            curY = pathStartY;
+        }
+        else if (t === "f" || t === "F" || t === "f*") {
+            flushPath("fill");
+        }
+        else if (t === "S" || t === "s") {
+            if (t === "s") {
+                // closeStroke: implicit closePath
+                if (curX !== pathStartX || curY !== pathStartY) {
+                    pendingLines.push({
+                        x1: curX,
+                        y1: curY,
+                        x2: pathStartX,
+                        y2: pathStartY,
+                    });
+                }
+            }
+            flushPath("stroke");
+        }
+        else if (t === "B" || t === "B*" || t === "b" || t === "b*") {
+            // fill + stroke combined
+            flushPath("fill");
+            flushPath("stroke");
+        }
+        else if (t === "n") {
+            // end path without painting — discard
+            pendingRects.length = 0;
+            pendingLines.length = 0;
+        }
+        idx++;
+    }
+    return segments;
+}
+/**
+ * Fast tokenizer for PDF content streams.
+ * Splits on whitespace, skipping comments and string literals.
+ */
+function tokenizeContentStream(raw) {
+    const tokens = [];
+    const len = raw.length;
+    let i = 0;
+    while (i < len) {
+        const ch = raw.charCodeAt(i);
+        // Skip whitespace
+        if (ch <= 32) {
+            i++;
+            continue;
+        }
+        // Skip comments
+        if (ch === 37 /* % */) {
+            while (i < len && raw.charCodeAt(i) !== 10)
+                i++;
+            continue;
+        }
+        // Skip string literals (...)
+        if (ch === 40 /* ( */) {
+            let depth = 1;
+            i++;
+            while (i < len && depth > 0) {
+                const c = raw.charCodeAt(i);
+                if (c === 92 /* \ */) {
+                    i++;
+                }
+                else if (c === 40) {
+                    depth++;
+                }
+                else if (c === 41) {
+                    depth--;
+                }
+                i++;
+            }
+            continue;
+        }
+        // Skip hex strings <...>
+        if (ch === 60 /* < */ && i + 1 < len && raw.charCodeAt(i + 1) !== 60) {
+            i++;
+            while (i < len && raw.charCodeAt(i) !== 62)
+                i++;
+            i++; // skip >
+            continue;
+        }
+        // Skip dict delimiters << >>
+        if (ch === 60 && i + 1 < len && raw.charCodeAt(i + 1) === 60) {
+            i += 2;
+            continue;
+        }
+        if (ch === 62 && i + 1 < len && raw.charCodeAt(i + 1) === 62) {
+            i += 2;
+            continue;
+        }
+        // Regular token: read until whitespace or delimiter
+        const start = i;
+        while (i < len) {
+            const c = raw.charCodeAt(i);
+            if (c <= 32 || c === 40 || c === 41 || c === 60 || c === 62 || c === 37)
+                break;
+            i++;
+        }
+        if (i > start) {
+            tokens.push(raw.substring(start, i));
+        }
+    }
+    return tokens;
+}
+// ---------------------------------------------------------------------------
+// Image region detection
+// ---------------------------------------------------------------------------
+/** Minimum area (pts²) for an image to be considered a diagram, not an icon. */
+const MIN_IMAGE_AREA = 5000;
+function extractImageRegions(stext, pageNumber, pageHeight) {
+    const regions = [];
+    for (const block of stext.blocks) {
+        if (block.type !== "image")
+            continue;
+        const { x, y, w, h } = block.bbox;
+        if (w * h < MIN_IMAGE_AREA)
+            continue; // skip tiny icons
+        // Convert Y from mupdf (top-left) to PDF (bottom-left) for ordering
+        const pdfTopY = pageHeight - y;
+        regions.push({
+            id: `p${pageNumber}-img${regions.length}`,
+            pageNumber,
+            bbox: { x, y, w, h },
+            topY: pdfTopY,
+        });
+    }
+    return regions;
+}
+// ---------------------------------------------------------------------------
+// Public API
+// ---------------------------------------------------------------------------
+/**
+ * Render an image region from a PDF page as a PNG buffer.
+ * Uses mupdf's DrawDevice to render just the cropped area at 2x resolution.
+ */
+export function renderImageRegion(input, region) {
+    const mupdf = require("mupdf");
+    const doc = mupdf.Document.openDocument(input, "application/pdf");
+    const page = doc.loadPage(region.pageNumber - 1);
+    const pad = 10;
+    const bx = region.bbox.x - pad;
+    const by = region.bbox.y - pad;
+    const bw = region.bbox.w + 2 * pad;
+    const bh = region.bbox.h + 2 * pad;
+    const scale = 2;
+    const pw = Math.round(bw * scale);
+    const ph = Math.round(bh * scale);
+    const pix = new mupdf.Pixmap(mupdf.ColorSpace.DeviceRGB, [0, 0, pw, ph], false);
+    pix.clear(255);
+    const matrix = [scale, 0, 0, scale, -bx * scale, -by * scale];
+    const dl = page.toDisplayList();
+    const dev = new mupdf.DrawDevice(matrix, pix);
+    dl.run(dev, mupdf.Matrix.identity);
+    dev.close();
+    return pix.asPNG();
+}
+/**
+ * Extract text boxes and vector segments from all pages of a PDF buffer.
+ */
+export async function extractPages(input) {
+    let mupdf;
+    try {
+        mupdf = await import("mupdf");
+    }
+    catch {
+        throw new Error("PDF support requires 'mupdf'. Install it: npm install mupdf");
+    }
+    const doc = mupdf.Document.openDocument(input, "application/pdf");
+    const pages = [];
+    for (let i = 0; i < doc.countPages(); i++) {
+        const pageNumber = i + 1;
+        const page = doc.loadPage(i);
+        const bounds = page.getBounds();
+        const pageHeight = bounds[3] - bounds[1];
+        // Single structured text pass with both flags
+        const stext = JSON.parse(page.toStructuredText("preserve-whitespace,preserve-images").asJSON());
+        // Extract text boxes and image regions from the same parse
+        const textBoxes = extractTextBoxes(page, pageNumber, pageHeight, stext);
+        const images = extractImageRegions(stext, pageNumber, pageHeight);
+        // Extract vector segments from raw content stream
+        let segments = [];
+        try {
+            const pageObj = page.getObject();
+            const contents = pageObj.get("Contents");
+            if (contents) {
+                let rawBytes;
+                if (contents.isArray()) {
+                    // Multiple content streams — concatenate
+                    const parts = [];
+                    const len = contents.length ?? 0;
+                    for (let j = 0; j < len; j++) {
+                        const stream = contents.get(j);
+                        if (stream?.readStream) {
+                            parts.push(stream.readStream().asUint8Array());
+                        }
+                    }
+                    const totalLen = parts.reduce((s, p) => s + p.length, 0);
+                    rawBytes = new Uint8Array(totalLen);
+                    let offset = 0;
+                    for (const part of parts) {
+                        rawBytes.set(part, offset);
+                        offset += part.length;
+                    }
+                }
+                else {
+                    rawBytes = contents.readStream().asUint8Array();
+                }
+                const raw = new TextDecoder().decode(rawBytes);
+                segments = extractSegmentsFromContentStream(raw, pageNumber);
+            }
+        }
+        catch {
+            // Content stream extraction failed — proceed with text only
+        }
+        pages.push({ pageNumber, textBoxes, segments, images });
+    }
+    return pages;
+}

package/dist/converters/pdf/grid.d.ts

@@ -0,0 +1,25 @@
+/**
+ * Table grid detection from vector segments and text boxes.
+ *
+ * Ported from @oharato/pdf2md-ts with TypeScript types and without
+ * CJK-specific borderless table heuristics. The core algorithm:
+ *
+ * 1. Classify segments as horizontal or vertical lines
+ * 2. Group horizontal Y-lines into table groups (split by vertical gaps)
+ * 3. For each group:
+ *    a. Full grid (H+V lines): build cells from grid intersections,
+ *       place text via raycasting
+ *    b. H-line only (no V lines): infer columns from text X positions
+ * 4. Prune empty rows/cols
+ *
+ * Coordinate system: PDF native (bottom-left origin, Y increases upward).
+ */
+import type { Segment, TableGrid, TextBox } from "./types.js";
+export interface GridResult {
+    grids: TableGrid[];
+    consumedIds: string[];
+}
+/**
+ * Detect all table grids on a single page from its text boxes and segments.
+ */
+export declare function resolveTableGrids(pageNumber: number, textBoxes: TextBox[], segments: Segment[]): GridResult;