fix(file-parsers): keep text-free PDFs empty and count only pages that were read

Gate the truncation notice on the sanitized, trimmed body: a text-free
multi-page PDF collapses to a lone separator, so the previous length check let
a notice turn a document callers treat as empty into one that looks like it
holds content.

Also stop counting a page the budget cut off before it yielded anything, so the
notice no longer reports one more page than was shown.

Author: waleedlatif1

apps/sim/lib/file-parsers/pdf-parser.test.ts

@@ -33,6 +33,24 @@ function buildTextBombPdf(repeats: number): Buffer {
     Buffer.from('<< /Type /Font /Subtype /Type1 /BaseFont /Helvetica >>'),
   ]
 
+  return assemblePdf(objects)
+}
+
+/** Builds a PDF whose pages carry no content stream, so nothing is extractable. */
+function buildTextFreePdf(pageCount: number): Buffer {
+  const pageIds = Array.from({ length: pageCount }, (_, i) => 3 + i)
+
+  return assemblePdf([
+    Buffer.from('<< /Type /Catalog /Pages 2 0 R >>'),
+    Buffer.from(
+      `<< /Type /Pages /Kids [${pageIds.map((id) => `${id} 0 R`).join(' ')}] /Count ${pageCount} >>`
+    ),
+    ...pageIds.map(() => Buffer.from('<< /Type /Page /Parent 2 0 R /MediaBox [0 0 612 792] >>')),
+  ])
+}
+
+/** Serializes numbered objects into a PDF with a matching xref table and trailer. */
+function assemblePdf(objects: Buffer[]): Buffer {
   const chunks: Buffer[] = [Buffer.from('%PDF-1.4\n')]
   const offsets: number[] = []
   let offset = chunks[0].length
@@ -88,4 +106,11 @@ describe('PdfParser', () => {
     expect(result.content).toContain('AAAA')
     expect(result.content).not.toContain('truncated')
   }, 30_000)
+
+  it('reports a multi-page PDF with no extractable text as empty', async () => {
+    const result = await new PdfParser().parseBuffer(buildTextFreePdf(3))
+
+    expect(result.content.trim()).toBe('')
+    expect(result.content).not.toContain('[...')
+  }, 30_000)
 })

apps/sim/lib/file-parsers/pdf-parser.ts

@@ -125,7 +125,12 @@ async function extractTextWithinBudget(pdf: PdfDocumentProxy): Promise<BoundedEx
     const { text, used, completed } = await readPageWithinBudget(page, remainingChars, deadline)
 
     remainingChars -= used
-    pageTexts.push(text)
+
+    // A page the budget cut off before it yielded anything was never really
+    // read, so it must not count toward `pagesRead` or add a blank separator.
+    if (completed || text.length > 0) {
+      pageTexts.push(text)
+    }
     page.cleanup()
 
     if (!completed) {
@@ -181,18 +186,22 @@ export class PdfParser implements FileParser {
           logger.warn(PDF_TRUNCATION_WARNING, { totalPages, pagesRead, textLength: text.length })
         }
 
+        const body = sanitizeTextForUTF8(text)
+
         // Callers only ever read `content`, so without an inline notice a truncated
-        // document is indistinguishable from a complete one. Empty text yields no
-        // notice, so a text-free PDF still reports as empty rather than as a lone notice.
+        // document is indistinguishable from a complete one. Tested after sanitizing
+        // and against `trim`, because a text-free multi-page PDF collapses to a lone
+        // separator — appending a notice to that would turn a document callers treat
+        // as empty into one that looks like it holds content.
         const notice =
-          truncated && text.length > 0
+          truncated && body.trim().length > 0
             ? truncationNotice(
                 `PDF text truncated at parser limits, showing first ${pagesRead} of ${totalPages} pages`
               )
             : ''
 
         return {
-          content: sanitizeTextForUTF8(text + notice),
+          content: body + notice,
           metadata: {
             pageCount: totalPages,
             source: 'unpdf',