Ask Your PDF – Local AI (Experimental)
Ask a local language model about text excerpts from your PDF. Document text stays in this tab and its local worker.
Experimental integration: actual GPU inference has not yet been validated on a real device by us. Do not rely on answers for professional decisions.
1. Check your device and load the model
Loading downloads runtime code and model files from jsDelivr, Hugging Face and the MLC project. Expect hundreds of MB and substantial GPU memory use (roughly 1 GB or more). Downloads may be cached; file text and questions are not saved to that cache. Use Wi-Fi and a capable computer. Two basic test questions run locally before document selection is enabled; passing is not a guarantee of answer quality.
2. Select a PDF and review its text
3. Ask about the selected text
Context sent to the local model
Only the displayed excerpts are analyzed. A source page in a citation does not prove the answer is correct.
AI response · verify it
Read the guide: ask about PDF without uploading it
How it works
Text extraction and excerpt selection happen locally. A small Qwen2.5-0.5B model generates an answer in a dedicated WebGPU worker; the worker blocks new network requests after loading. No cloud fallback is used. Extraction may omit diagrams or reorder text. This small model can hallucinate and may be weak on Arabic or complex questions.
Prepare a file without an AI model · OCR for scans · Privacy policy
Questions
Is this actual AI?
The integration runs a small Qwen2.5 model through WebLLM on a compatible device. Answers are not generated by a server.
Does it work on every device?
No. WebGPU and sufficient memory are required. Model loading can fail despite passing the preliminary check. Actual GPU inference has not been verified in the current development environment.
Does it analyze the whole document?
No. Text is extracted from up to 20 selected pages, and up to three short excerpts are used for each question. Summaries cover only the displayed excerpts.
Can it read images and tables accurately?
The model is text-only and cannot see images. OCR is not performed here, and extraction may lose table structure.
Are citations guaranteed?
No. Model-generated page references may be wrong. The tool displays actual context excerpts for review and flags page numbers outside that context.