@juspay/neurolink 12.30.1 → 12.31.0

package/dist/utils/htmlToMarkdown.js

@@ -0,0 +1,442 @@
+/**
+ * HTML to Markdown conversion for document processors.
+ *
+ * Scoped deliberately to the small, well-formed HTML subset that document
+ * converters emit (mammoth for DOCX): headings, paragraphs, emphasis, links,
+ * images, lists, block quotes, preformatted text and tables. It is not a
+ * general-purpose web-page converter, which is why the repository gains no
+ * dependency for it.
+ *
+ * @module utils/htmlToMarkdown
+ */
+/** Tags that never have a closing tag and so never open a scope. */
+const VOID_TAGS = new Set([
+    "area",
+    "base",
+    "br",
+    "col",
+    "embed",
+    "hr",
+    "img",
+    "input",
+    "link",
+    "meta",
+    "source",
+    "track",
+    "wbr",
+]);
+const NAMED_ENTITIES = {
+    amp: "&",
+    lt: "<",
+    gt: ">",
+    quot: '"',
+    apos: "'",
+    nbsp: " ",
+};
+/**
+ * Decode the numeric and named character references a document converter can
+ * emit. Unknown references are left verbatim rather than dropped, so no text
+ * is silently lost.
+ */
+function decodeEntities(value) {
+    return value.replace(/&(#x?[0-9a-fA-F]+|[a-zA-Z]+);/g, (match, body) => {
+        if (body.startsWith("#")) {
+            const codePoint = body.startsWith("#x") || body.startsWith("#X")
+                ? Number.parseInt(body.slice(2), 16)
+                : Number.parseInt(body.slice(1), 10);
+            if (Number.isNaN(codePoint) || codePoint < 0 || codePoint > 0x10ffff) {
+                return match;
+            }
+            try {
+                return String.fromCodePoint(codePoint);
+            }
+            catch {
+                return match;
+            }
+        }
+        const named = NAMED_ENTITIES[body.toLowerCase()];
+        return named ?? match;
+    });
+}
+/**
+ * Parse the HTML subset into a node tree.
+ *
+ * Unbalanced closing tags are ignored and unclosed elements are closed at the
+ * end of input, so malformed markup degrades to partial structure instead of
+ * throwing.
+ */
+function parseHtml(html) {
+    const root = {
+        kind: "element",
+        tag: "#root",
+        attrs: {},
+        children: [],
+    };
+    const stack = [root];
+    const tagPattern = /<\/?([a-zA-Z][a-zA-Z0-9-]*)((?:[^>"']|"[^"]*"|'[^']*')*)>/g;
+    let cursor = 0;
+    let match;
+    const pushText = (raw) => {
+        if (!raw) {
+            return;
+        }
+        const current = stack[stack.length - 1];
+        current.children.push({ kind: "text", value: decodeEntities(raw) });
+    };
+    while ((match = tagPattern.exec(html)) !== null) {
+        pushText(html.slice(cursor, match.index));
+        cursor = match.index + match[0].length;
+        const tag = match[1].toLowerCase();
+        const isClosing = match[0].startsWith("</");
+        const isSelfClosing = match[2].trimEnd().endsWith("/");
+        if (isClosing) {
+            // Only unwind if the tag is actually open, so a stray `</div>` is
+            // inert. Search from the top of the stack down so a closing tag
+            // matches the innermost (most recently opened) element with that
+            // name — `findIndex` would match the OUTERMOST one instead, and for
+            // nested same-tag elements (`<div><div>A</div>B</div>`) that pops both
+            // levels at once, reattaching everything after the real inner close to
+            // the wrong ancestor.
+            let openIndex = -1;
+            for (let i = stack.length - 1; i >= 0; i--) {
+                if (stack[i].tag === tag) {
+                    openIndex = i;
+                    break;
+                }
+            }
+            if (openIndex > 0) {
+                stack.length = openIndex;
+            }
+            continue;
+        }
+        const element = {
+            kind: "element",
+            tag,
+            attrs: parseAttributes(match[2]),
+            children: [],
+        };
+        stack[stack.length - 1].children.push(element);
+        if (!isSelfClosing && !VOID_TAGS.has(tag)) {
+            stack.push(element);
+        }
+    }
+    pushText(html.slice(cursor));
+    return root.children;
+}
+function parseAttributes(raw) {
+    const attrs = {};
+    const pattern = /([a-zA-Z_:][a-zA-Z0-9_.:-]*)\s*=\s*("([^"]*)"|'([^']*)'|([^\s"'>]+))/g;
+    let match;
+    while ((match = pattern.exec(raw)) !== null) {
+        const value = match[3] ?? match[4] ?? match[5] ?? "";
+        attrs[match[1].toLowerCase()] = decodeEntities(value);
+    }
+    return attrs;
+}
+function isElement(node) {
+    return node.kind === "element";
+}
+/** Collapse runs of whitespace the way an HTML renderer would. */
+function collapseWhitespace(value) {
+    return value
+        .replace(/\s+/g, " ")
+        .replace(/&/g, "&amp;")
+        .replace(/</g, "&lt;")
+        .replace(/>/g, "&gt;");
+}
+/**
+ * Format a link/image destination per CommonMark's angle-bracket form.
+ *
+ * A bare `(dest)` destination ends at the first whitespace or unbalanced
+ * `)`, so a value containing either — an `href`/`src` with a space, or one
+ * with unbalanced parentheses — would silently truncate or corrupt the
+ * surrounding `[text](dest)` / `![alt](dest)` markup. Wrapping in `<...>`
+ * makes the destination unambiguous; CommonMark requires any literal `<`
+ * or `>` inside that form to be escaped. Ordinary destinations (no
+ * whitespace or parens) are returned byte-identical to before.
+ */
+function formatLinkDestination(dest) {
+    if (!/[\s()]/.test(dest)) {
+        return dest;
+    }
+    // Backslash first: escaping only `<`/`>` turns `\>` into `\\>`, which
+    // CommonMark reads as a literal backslash followed by a bare `>`.
+    const escaped = dest
+        .replace(/\\/g, "\\\\")
+        .replace(/</g, "\\<")
+        .replace(/>/g, "\\>");
+    return `<${escaped}>`;
+}
+/** Render inline content (emphasis, links, images, line breaks) to markdown. */
+function renderInline(nodes) {
+    let out = "";
+    for (const node of nodes) {
+        if (node.kind === "text") {
+            out += collapseWhitespace(node.value);
+            continue;
+        }
+        switch (node.tag) {
+            case "script":
+            case "style":
+                break;
+            case "strong":
+            case "b": {
+                const inner = renderInline(node.children).trim();
+                out += inner ? `**${inner}**` : "";
+                break;
+            }
+            case "em":
+            case "i": {
+                const inner = renderInline(node.children).trim();
+                out += inner ? `*${inner}*` : "";
+                break;
+            }
+            case "code": {
+                const inner = renderInline(node.children).trim();
+                out += inner ? `\`${inner}\`` : "";
+                break;
+            }
+            case "a": {
+                const inner = renderInline(node.children).trim();
+                const href = node.attrs.href;
+                out += href ? `[${inner}](${formatLinkDestination(href)})` : inner;
+                break;
+            }
+            case "img": {
+                const alt = node.attrs.alt ?? "";
+                const src = node.attrs.src ?? "";
+                // A DOCX image is inlined as a multi-megabyte data URI; keep the
+                // alt text and drop the payload rather than flooding the prompt.
+                out += src.startsWith("data:")
+                    ? `![${alt}](embedded-image)`
+                    : `![${alt}](${formatLinkDestination(src)})`;
+                break;
+            }
+            case "br":
+                out += "\n";
+                break;
+            default:
+                out += renderInline(node.children);
+                break;
+        }
+    }
+    return out;
+}
+/**
+ * Extract a `<pre>` block's text content verbatim.
+ *
+ * `renderInline()` funnels every text node through `collapseWhitespace()`,
+ * which does `.replace(/\s+/g, " ")` — right for prose, destructive for
+ * code: it collapses every newline, tab and indentation run onto one line.
+ * This walks the same tree but keeps each text node's raw value, so a
+ * multi-line, indented code sample survives into the fenced block intact.
+ * A nested element (e.g. `<pre><code>…</code></pre>`) still contributes its
+ * text, but ignores inline markdown semantics like `**`/`*`/`` ` `` —
+ * `<pre>` content is literal code, not prose, and injecting markdown markers
+ * into it would corrupt the code rather than format it.
+ */
+function renderPreformattedText(nodes) {
+    let out = "";
+    for (const node of nodes) {
+        if (node.kind === "text") {
+            out += node.value;
+            continue;
+        }
+        if (node.tag === "script" || node.tag === "style") {
+            continue;
+        }
+        if (node.tag === "br") {
+            out += "\n";
+            continue;
+        }
+        out += renderPreformattedText(node.children);
+    }
+    return out;
+}
+/** Collect the `<tr>` rows under a table, seeing through thead/tbody/tfoot. */
+function collectRows(node) {
+    const rows = [];
+    for (const child of node.children.filter(isElement)) {
+        if (child.tag === "tr") {
+            rows.push(child);
+        }
+        else if (["thead", "tbody", "tfoot"].includes(child.tag)) {
+            rows.push(...collectRows(child));
+        }
+    }
+    return rows;
+}
+/** A pipe inside a cell would otherwise split it into two columns. */
+function escapeCell(value) {
+    return value
+        .replace(/\\/g, "\\\\")
+        .replace(/\|/g, "\\|")
+        .replace(/\n/g, " ")
+        .trim();
+}
+function renderTable(node) {
+    const rows = collectRows(node);
+    if (rows.length === 0) {
+        return "";
+    }
+    const cellsOf = (row) => row.children
+        .filter(isElement)
+        .filter((cell) => cell.tag === "td" || cell.tag === "th")
+        .map((cell) => escapeCell(renderBlocks(cell.children).replace(/\n+/g, " ")));
+    // Markdown has no headerless table, so when the source marked no header row
+    // the first row is promoted — that is what a reader sees anyway.
+    const header = cellsOf(rows[0]);
+    const bodyRows = rows.slice(1).map(cellsOf);
+    const columnCount = Math.max(header.length, ...bodyRows.map((r) => r.length), 1);
+    const pad = (cells) => Array.from({ length: columnCount }, (_, i) => cells[i] ?? "");
+    const lines = [
+        `| ${pad(header).join(" | ")} |`,
+        `| ${Array.from({ length: columnCount }, () => "---").join(" | ")} |`,
+        ...bodyRows.map((cells) => `| ${pad(cells).join(" | ")} |`),
+    ];
+    return lines.join("\n");
+}
+function renderList(node, depth) {
+    const ordered = node.tag === "ol";
+    const items = node.children
+        .filter(isElement)
+        .filter((child) => child.tag === "li");
+    const indent = "  ".repeat(depth);
+    return items
+        .map((item, index) => {
+        const marker = ordered ? `${index + 1}.` : "-";
+        // Nested lists render themselves at the next depth; everything else on
+        // the item is inline content belonging to this bullet.
+        const nested = item.children
+            .filter(isElement)
+            .filter((child) => child.tag === "ul" || child.tag === "ol");
+        const own = item.children.filter((child) => !(isElement(child) && (child.tag === "ul" || child.tag === "ol")));
+        const label = renderBlocks(own).replace(/\n+/g, " ").trim();
+        const nestedText = nested
+            .map((child) => renderList(child, depth + 1))
+            .filter(Boolean)
+            .join("\n");
+        return [`${indent}${marker} ${label}`.trimEnd(), nestedText]
+            .filter(Boolean)
+            .join("\n");
+    })
+        .join("\n");
+}
+/** Render block-level content, separating blocks with a blank line. */
+function renderBlocks(nodes) {
+    const blocks = [];
+    let inlineBuffer = "";
+    const flushInline = () => {
+        const text = inlineBuffer.trim();
+        if (text) {
+            blocks.push(text);
+        }
+        inlineBuffer = "";
+    };
+    for (const node of nodes) {
+        if (node.kind === "text") {
+            inlineBuffer += collapseWhitespace(node.value);
+            continue;
+        }
+        const headingMatch = /^h([1-6])$/.exec(node.tag);
+        if (headingMatch) {
+            flushInline();
+            const level = Number.parseInt(headingMatch[1], 10);
+            const text = renderInline(node.children).replace(/\n+/g, " ").trim();
+            if (text) {
+                blocks.push(`${"#".repeat(level)} ${text}`);
+            }
+            continue;
+        }
+        switch (node.tag) {
+            case "script":
+            case "style":
+                flushInline();
+                break;
+            case "p": {
+                flushInline();
+                const text = renderBlocks(node.children).trim();
+                if (text) {
+                    blocks.push(text);
+                }
+                break;
+            }
+            case "ul":
+            case "ol": {
+                flushInline();
+                const text = renderList(node, 0);
+                if (text) {
+                    blocks.push(text);
+                }
+                break;
+            }
+            case "table": {
+                flushInline();
+                const text = renderTable(node);
+                if (text) {
+                    blocks.push(text);
+                }
+                break;
+            }
+            case "blockquote": {
+                flushInline();
+                const inner = renderBlocks(node.children).trim();
+                if (inner) {
+                    blocks.push(inner
+                        .split("\n")
+                        .map((line) => (line ? `> ${line}` : ">"))
+                        .join("\n"));
+                }
+                break;
+            }
+            case "pre": {
+                flushInline();
+                const inner = renderPreformattedText(node.children).replace(/^\n+|\n+$/g, "");
+                if (inner) {
+                    blocks.push(`\`\`\`\n${inner}\n\`\`\``);
+                }
+                break;
+            }
+            case "hr": {
+                flushInline();
+                blocks.push("---");
+                break;
+            }
+            case "div":
+            case "section":
+            case "article": {
+                flushInline();
+                const inner = renderBlocks(node.children).trim();
+                if (inner) {
+                    blocks.push(inner);
+                }
+                break;
+            }
+            default:
+                inlineBuffer += renderInline([node]);
+                break;
+        }
+    }
+    flushInline();
+    return blocks.join("\n\n");
+}
+/**
+ * Convert an HTML fragment to Markdown.
+ *
+ * @param html - HTML produced by a document converter
+ * @returns Markdown text, or an empty string when the fragment has no content
+ *
+ * @example
+ * ```typescript
+ * htmlToMarkdown("<h1>Title</h1><p>Body</p>"); // "# Title\n\nBody"
+ * ```
+ */
+export function htmlToMarkdown(html) {
+    if (!html || !html.trim()) {
+        return "";
+    }
+    return renderBlocks(parseHtml(html))
+        .replace(/[ \t]+$/gm, "")
+        .replace(/\n{3,}/g, "\n\n")
+        .trim();
+}

package/dist/utils/messageBuilder.js

@@ -895,6 +895,9 @@ async function detectFileForUnifiedArray(file, options, provider, genericFileMax
             "unknown",
         ],
         csvOptions: options.csvOptions,
+        // Office options travel the same unified `files` path, so this is where a
+        // caller's sheet selection / format choice reaches the XLSX processor.
+        officeOptions: options.officeOptions,
         // #478: videos arrive through this unified `files` path, so this is
         // where the CLI's frame/quality/format request has to be handed on.
         videoOptions: options.videoOptions

package/dist/utils/multimodalOptionsBuilder.d.ts

@@ -16,6 +16,7 @@ import type { StreamOptions } from "../types/index.js";
  *   - input.videoFiles: Video files (Buffer | string paths)
  *   - csvOptions: CSV parsing options
  *   - videoOptions: Video frame/quality/format/transcription options
+ *   - officeOptions: DOCX/XLSX sheet selection and output-format options
  *   - systemPrompt: System-level instructions
  *   - conversationMessages: Chat history
  *   - temperature: Model temperature (0-1)
@@ -28,6 +29,7 @@ import type { StreamOptions } from "../types/index.js";
  * @returns {object} Normalized options object with:
  *   - input: { text, images, content, files, csvFiles, pdfFiles, audioFiles, videoFiles }
  *   - csvOptions: CSV processing options
+ *   - officeOptions: Office (DOCX/XLSX) processing options
  *   - systemPrompt: System prompt string
  *   - conversationHistory: Message history array
  *   - provider: Provider name
@@ -66,6 +68,7 @@ export declare function buildMultimodalOptions(options: StreamOptions, providerN
         outputFormat?: import("../index.js").VisionImageOutputFormat;
     } | undefined;
     videoOptions: import("../index.js").VideoProcessorOptions | undefined;
+    officeOptions: import("../index.js").OfficeProcessorOptions | undefined;
     systemPrompt: string | undefined;
     conversationHistory: import("../index.js").ChatMessage[] | undefined;
     provider: string;

package/dist/utils/multimodalOptionsBuilder.js

@@ -15,6 +15,7 @@
  *   - input.videoFiles: Video files (Buffer | string paths)
  *   - csvOptions: CSV parsing options
  *   - videoOptions: Video frame/quality/format/transcription options
+ *   - officeOptions: DOCX/XLSX sheet selection and output-format options
  *   - systemPrompt: System-level instructions
  *   - conversationMessages: Chat history
  *   - temperature: Model temperature (0-1)
@@ -27,6 +28,7 @@
  * @returns {object} Normalized options object with:
  *   - input: { text, images, content, files, csvFiles, pdfFiles, audioFiles, videoFiles }
  *   - csvOptions: CSV processing options
+ *   - officeOptions: Office (DOCX/XLSX) processing options
  *   - systemPrompt: System prompt string
  *   - conversationHistory: Message history array
  *   - provider: Provider name
@@ -67,6 +69,10 @@ export function buildMultimodalOptions(options, providerName, modelName) {
         // `StreamOptions` never reach VideoProcessor on the Bedrock branch, and
         // it silently falls back to the duration-tier default.
         videoOptions: options.videoOptions,
+        // Same #1259 whitelist-drop bug class as audioFiles/videoFiles above:
+        // omitted here, Bedrock's multimodal generate()/stream() path always got
+        // sheetName/formatStyle undefined regardless of what the caller passed.
+        officeOptions: options.officeOptions,
         systemPrompt: options.systemPrompt,
         conversationHistory: options.conversationMessages,
         provider: providerName,