bunnyquery 1.5.6 → 1.5.7

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "bunnyquery",
-  "version": "1.5.6",
+  "version": "1.5.7",
   "description": "Embeddable BunnyQuery AI chat widget + its framework-agnostic chat engine",
   "main": "bunnyquery.js",
   "exports": {

package/src/engine/office.ts

@@ -101,6 +101,30 @@ export function isServerExtractable(name?: string, mime?: string): boolean {
 /** @deprecated renamed to {@link isServerExtractable} (now also covers text files). */
 export const isOfficeFile = isServerExtractable;
 
+// Extensions that are best read WINDOW-BY-WINDOW via the readFileContent tool instead of
+// inlined once: spreadsheets (grid rows + embedded photos, and can be huge) and PDFs
+// (scanned page images). For these the indexing agent is told to page through the whole
+// file with readFileContent rather than reading a capped inline dump or web_fetching a URL.
+const PAGED_READ_EXTENSIONS = new Set(['xls', 'xlsx', 'xlsm', 'ods', 'pdf']);
+
+/**
+ * True when a file should be indexed by PAGING through readFileContent (spreadsheets and
+ * PDFs), rather than inline extraction or a web_fetch URL. This is what lets a huge sheet
+ * be read row-window by row-window and a scanned PDF be read page-image by page-image,
+ * with embedded photos delivered to the vision model.
+ */
+export function isPagedReadFile(name?: string, mime?: string): boolean {
+	const ext = (name || '').split('.').pop()?.toLowerCase() || '';
+	if (PAGED_READ_EXTENSIONS.has(ext)) return true;
+	const m = (mime || '').toLowerCase();
+	return (
+		m === 'application/pdf' ||
+		m === 'application/vnd.ms-excel' ||
+		m.includes('spreadsheetml') ||
+		m.includes('opendocument.spreadsheet')
+	);
+}
+
 // Monotonic counter so placeholders are unique even for same-named files in one
 // request. Token shape must match the worker's _EXTRACT_PLACEHOLDER_RE:
 // {{SKAPI_FILE_CONTENT::<id>}}.

package/src/engine/prompts/indexing_user_message.ts

@@ -36,6 +36,12 @@ export type BuildIndexingUserMessageOptions = {
 	 * web_fetch for this file. Takes precedence over `inlineContentPlaceholder`.
 	 */
 	inlineContent?: string;
+	/**
+	 * Spreadsheet or PDF: read by PAGING through the readFileContent tool (grid rows +
+	 * embedded photos / rendered scanned pages), not inline and not by web_fetch. The
+	 * message instructs the agent to page through EVERY window and datafy each.
+	 */
+	pagedRead?: boolean;
 };
 
 export function buildIndexingUserMessage(
@@ -78,5 +84,19 @@ export function buildIndexingUserMessage(
 		);
 	}
 
+	if (options?.pagedRead) {
+		// Spreadsheet / PDF: force the paging path. The agent MUST read this with the
+		// readFileContent tool (which returns the file window by window, with grid rows,
+		// embedded photos, and rendered scanned pages), NOT by fetching the URL.
+		return (
+			head +
+			`\nRead this file with the readFileContent tool, using the storage path above - do NOT fetch a URL and do NOT rely on a single sample. ` +
+			`readFileContent returns the file ONE WINDOW at a time: spreadsheets as coordinate-tagged grid rows (e.g. 'R4 A:E&I NUMBER | B:E1007'), scanned/large PDFs as rendered PAGE IMAGES, and windows may include embedded photos - LOOK at any images and datafy what they show. ` +
+			`Page through EVERY window: for each window SAVE records for its rows/items/pages (postRecords, one record per row/item), THEN if the window says MORE REMAINS call readFileContent again with the cursor it gives you. Repeat until it says END OF FILE, so the WHOLE file is indexed. ` +
+			`Do NOT stop after the first window and do NOT just write a summary. Use the storage path above for the "src::" unique_id.` +
+			(attachment.url ? `\n(A temporary URL is provided ONLY as a fallback if readFileContent fails: ${attachment.url})` : '')
+		);
+	}
+
 	return head + `- temporary URL (fetch this to read the file contents): ${attachment.url}`;
 }

package/src/engine/requests.ts

@@ -11,7 +11,7 @@
  *                                    only the BgTaskEntry TYPE lives here)
  */
 import { buildIndexingSystemPrompt, buildIndexingUserMessage } from './prompts';
-import { isServerExtractable, makeExtractPlaceholder, type ExtractDirective, type FileUrlDirective } from './office';
+import { isServerExtractable, isPagedReadFile, makeExtractPlaceholder, type ExtractDirective, type FileUrlDirective } from './office';
 import { chatEngineConfig, pollOpt } from './config';
 
 export const ANTHROPIC_MESSAGES_API_URL = 'https://api.anthropic.com/v1/messages';
@@ -450,8 +450,13 @@ export type AttachmentSaveInfo = {
 export async function notifyAgentSaveAttachment(info: AttachmentSaveInfo) {
 	const { platform, service, owner, attachment, parsedContent } = info;
 
+	// Spreadsheets and PDFs are read by PAGING through readFileContent (grid rows +
+	// embedded photos / scanned page images), NOT inlined once - so they skip the inline
+	// server-extract and the agent is told to page the whole file (see pagedRead below).
+	const pagedRead = !parsedContent && isPagedReadFile(attachment.name, attachment.mime);
+
 	// Client-parsed content wins over server-side extraction.
-	const serverExtract = !parsedContent && isServerExtractable(attachment.name, attachment.mime);
+	const serverExtract = !parsedContent && !pagedRead && isServerExtractable(attachment.name, attachment.mime);
 	const placeholder = serverExtract ? makeExtractPlaceholder(attachment.storagePath) : undefined;
 	const extractContent: ExtractDirective[] | undefined =
 		serverExtract && placeholder
@@ -466,7 +471,9 @@ export async function notifyAgentSaveAttachment(info: AttachmentSaveInfo) {
 			? { inlineContent: parsedContent }
 			: placeholder
 				? { inlineContentPlaceholder: placeholder }
-				: undefined,
+				: pagedRead
+					? { pagedRead: true }
+					: undefined,
 	);
 
 	const systemPrompt = buildIndexingSystemPrompt({