Подготовка PDF для ИИ

It follows the LlamaIndex document schema: one document per page with that page's text and metadata fields. It's produced by PyMuPDF's LlamaIndex integration, so it can be loaded directly into LlamaIndex or any pipeline that accepts per-page text with metadata.

PDF to Text gives one plain text stream, and PDF to Markdown keeps headings and structure for reading. This tool splits the content page by page and wraps it in structured JSON with metadata, which is what retrieval pipelines need for chunking and citing sources.

Only if they already have a text layer. Scanned images produce empty or near-empty output, so run them through OCR PDF first and then extract.

Yes. Upload multiple files and you get pdf-for-ai.zip containing one yourfile_llm.json per PDF; a single file downloads as yourfile_llm.json directly. If one file in the batch fails, the rest still complete and the summary tells you how many failed.

No. There are no options; every page is extracted with its full text and metadata. Do chunking in your own pipeline, where you can pick sizes that fit your embedding model, and drop pages you don't need there.

Yes. If a file is encrypted, you're prompted for its password before extraction starts, and the text is pulled from the decrypted copy.

No. Despite the name, nothing is sent to an AI provider or any server; the extraction runs with PyMuPDF compiled to WebAssembly inside your browser. You decide afterward where the JSON goes.

Feed it to a RAG pipeline, build a knowledge base for a chatbot, run LLM analysis over reports, or turn a PDF collection into training data. Because the content is split per page with metadata, answers can point back to the page they came from.