# Chat with document

> Ask questions using retrieved passages from an uploaded document.

[Open tool](https://www.toolcabana.com/is/chat-document) · [Gervigreind og fyrirmæli](https://www.toolcabana.com/is/category/ai-and-prompt-tools)

Tool ID: chat-document. Requested language: is. Description language: en. Complete guide translation: no; untranslated sections use English.

## Overview (en)

Chat with document answers a question about a text-based PDF, TXT, Markdown, CSV or JSON file, or pasted text, using a small model that runs in your browser. The document is split into overlapping passages, the six sharing the most words with your question are sent to the model, and it is told to cite passage numbers and say when an answer is unsupported.

## Supported tasks (en)

Downloads an inference model on first run. Chat and generation require WebGPU; smaller recognition models use browser WebAssembly. AI output can omit or invent details. Review it before use.

## Steps (en)

1. Choose a source file. If the tool accepts multiple files, arrange them in the order you want.
2. Set processing, question for pasted source.
3. Run chat with document, review the output, then use the available copy or download controls.

## Settings

- Processing (en; key: inference; type: select): Local browser [en]
- Question for pasted source (en; key: question; type: text)

## Limitations (en)

Chat sources: 8,000 characters; generation: up to 1,400 output tokens. First model downloads can be hundreds of MB. WebGPU chat and browser memory requirements vary by device. Model output is a draft.

## Example input

```text
A useful tool makes everyday work easier. Review the results before sharing.
```

## Privacy and connections (en)

Local mode downloads public model/runtime assets and keeps input in this browser. Models may remain cached on this device.

## Questions (en)

### Can I chat with a scanned PDF?

Only PDFs with a text layer work, because the tool reads the embedded text from up to the first 30 pages. An image-only scan yields no usable text, so run it through OCR first and upload the result.

### Where do I type my question when I upload a document?

With a file selected, type the question in the main text box. When you paste the source text instead, the main box holds the document and the question goes in the Question for pasted source field, which defaults to What are the main points?

### How does chat with document pick which parts of the file to use?

It cuts the text into 1,200-character passages with 100 characters of overlap, scores each by how many words of three or more characters from your question it contains, and passes the top six. Questions worded differently from the document may miss relevant passages.

### Where is my input processed?

Local mode downloads public model/runtime assets and keeps input in this browser. Models may remain cached on this device.

### What are the input limits?

Chat sources: 8,000 characters; generation: up to 1,400 output tokens. First model downloads can be hundreds of MB. WebGPU chat and browser memory requirements vary by device. Model output is a draft.

## Related tools

- [Private local AI chat](https://www.toolcabana.com/is/local-ai-chat)
- [AI flashcard maker](https://www.toolcabana.com/is/flashcard-maker)
- [AI quiz maker](https://www.toolcabana.com/is/quiz-maker)
- [AI study guide](https://www.toolcabana.com/is/study-guide)