Loading page
About this tool
Review text-based RAG files locally for prompt-injection, hidden-content, suspicious-link, and duplicate signals before indexing.
The RAG Corpus Quarantine Scanner reads .txt, .md, .html, .csv and .json files as inert text in your browser and flags deterministic review signals before you index them: prompt-injection phrasing, hidden Unicode and hidden or active HTML, suspicious links, and duplicate documents. Each file is classified as quarantine suggested when a high-severity signal fires, review suggested when only lower-severity or ambiguous signals appear, or no configured signals. It is a triage step for anyone loading third-party or scraped documents into a retrieval index - the output is a review queue, not a verdict on whether a file is malicious.
Open RAG Corpus Quarantine Scanner on AltFTool — it loads instantly in your browser.
Press Choose files for up to 30 .txt, .md, .html, .csv or .json documents, 2 MB each and 10 MB per batch.
Press Scan corpus locally to classify each file as quarantine suggested, review suggested or no configured signals.
Read the duplicate groups and severity totals, then press Download report for rag-corpus-quarantine-counts-only.txt.
Exact byte-identical matches are grouped apart from normalized matches, where files differ only by markup, entities, whitespace or invisible characters after NFKC folding.
URLs are parsed to flag embedded credentials, javascript:, data:text/html and file:// schemes, punycode hosts, private-network and raw-IP targets, shorteners, executable file endings and token-like query parameters - without any request being made.
The counts-only export lists files, classifications, severity totals, categories and duplicate groups, and deliberately excludes filenames, corpus text, matched snippets and URLs.
Plain text, Markdown, HTML, CSV and JSON, up to 30 files at 2 MB each and 10 MB in total per batch. Each file is scanned to a limit of 500,000 characters, and anything beyond that is flagged as truncated coverage rather than silently skipped.
At least one high-severity signal - for example an active URL scheme such as javascript: or file://, a URL carrying embedded credentials, or a script, iframe, object or embed tag in an HTML file. Any lower-severity signal on its own downgrades the file to review suggested instead.
No. It means no configured rule matched. These are deterministic pattern checks, so novel phrasing, obfuscation or context-specific risks can pass through - treat a clear result as the absence of known signals, not as an assurance.
No. Files are read as inert text in your own browser: no HTML is rendered, no link is followed, no resource is fetched, nothing is uploaded, and nothing is kept after you leave the page.