AI & LLM Testing
Test fixtures for people building AI agents, LLM-feeding scrapers, and RAG pipelines. Includes prompt-injection isolation testing, a RAG testing dataset with a documented answer key, and RAG testing PDFs and documents built to stress-test extraction.
Prompt injection
A page with safe, labeled example injection patterns and a benign canary phrase, so you can check that page content stays isolated from an agent's instructions.
- medium Prompt injection test page — Seven hidden-text vectors (HTML comments, display:none, off-screen, color-matched, zero-font, alt text, meta description), each with a unique canary.
RAG & retrieval corpus
A small RAG testing dataset: fifteen short, topically distinct articles paired with a documented answer key, enough to test chunking, embedding, and retrieval end to end and check whether the right result comes back.
- easy Corpus & answer key — Every article in the dataset, plus sample queries mapped to the correct expected result.
- easy Example article — Available as HTML, plain Markdown, or PDF, and in bulk as JSON at /ai/corpus/articles.json.
- easy Full dataset bundle — The whole RAG testing dataset in one download: every article, every format URL, and the answer key in a single JSON file.
PDF documents
RAG testing PDFs for exercising PDF-to-text extraction, from simple single-column articles up to layouts built specifically to trip up naive parsers.
- easy PDF documents — Every RAG testing PDF on the site in one place, including a plain PDF version of each corpus article.
- medium Table extraction test — A drawn table (an order log) that tests whether extraction keeps each row's cells together instead of reading columns out of order.
- hard Two-column layout test — Two unrelated essays placed side by side to test whether extraction respects column boundaries instead of reading straight across the page.
- hard Running header/footer test — A multi-page document with repeated header and footer boilerplate on every page, testing whether extraction strips it out.