The job
You opened a 200-page research paper, a vendor contract, or a regulatory
filing in your browser, and you need to find every mention of a phrase —
"data retention", "indemnification", \d{4}-\d{2}-\d{2} date stamps —
without scrolling, exporting, or piping it through a third-party PDF tool.
Why this is hard without Sephir
Chrome's built-in PDF viewer has a Cmd+F that only marks one hit at a time and refuses to take regex. ChatGPT can't read the PDF directly — you'd have to download it, upload it to a chat, wait for ingestion, and hope it didn't truncate the long ones. The friction kills the question.
How Sephir does it
- Open the PDF in any browser tab (Chrome PDF viewer or a direct URL).
- Open the Sephir sidepanel.
- Ask: "Find every mention of 'data retention' in this PDF with the surrounding context."
- Or for regex: "Grep for
\b\d{4}-\d{2}-\d{2}\bin this PDF — list every date and which page it's on."
Sephir's grep_pdf tool fetches the PDF, parses page text with pdf.js
in an offscreen document (so the MV3 service worker stays responsive),
and hands the agent page-numbered snippets:
12 matches for "data retention":
p. 14: …customer's "data retention" preferences shall be honored for…
p. 27: …data retention periods are described in Schedule B…
p. 58: …data retention beyond 90 days requires written consent…
Why the regex matters
Substring search is enough for "find this exact phrase". A real grep
gives you patterns: every email address (\S+@\S+\.\S+), every dollar
amount (\$[\d,]+\.\d{2}), every section heading (^\d+\.\s+[A-Z]).
You phrase the question once and the agent walks the document for you.
What you can chain it with
grep_pdf→extractAsSchema— find the section, parse the table inside it into JSON.grep_pdf→web_search— find the cited reference, verify it online.grep_pdfacross two PDFs in two tabs — diff the wording of the same clause between contract versions.
What it doesn't do
Scanned PDFs without an OCR layer return no text — pdf.js can only read what's actually in the document. For image-only scans you'll need to run OCR first; Sephir doesn't bundle Tesseract.