Chat with PDF

Ask privately. Check every answer against the page.

Search by meaning and wording to find passages even when you paraphrase. A local multilingual model indexes the document once; optional local AI writes a short answer from the retrieved evidence. Models download on first use. Search misses do not prove information is absent. No document or question is uploaded.

How to use Chat with PDF

1. Open your document from your device.
2. Adjust the settings and preview the result.
3. Read text results on screen or download your file. Your original remains untouched.

Local by design

Your documents are processed on your device. No account, payment or document upload is required.

Enable JavaScript to use the document workspace.

Find answers in your PDF with cited source passages

Document chat is most useful when it helps you reach the evidence quickly. Ask about a term, a date or a named page, then inspect the returned passage in its original context. SandboxPDF’s chat is deliberately evidence-led: it displays excerpts from your document with page references rather than presenting invented prose as an established fact.

The default method searches both meaning and wording using a local multilingual embedding model. It can retrieve a relevant passage even when the question uses different vocabulary. An optional local language model writes a concise answer from relevant passages and includes page citations with exact supporting excerpts. Citation and number checks reject some unsupported output, but do not establish that an interpretation is correct. Both methods process the document on your device. Loading the optional model requires an initial download, and starting it can take time and significant memory. Once initialised, model inference runs locally; the document text is not sent to a cloud chat service.

What the search and answer options actually do

Meaning search compares the question with overlapping passages throughout the extracted document. It does not require every question word to appear in a passage. For example, a question about getting money back may lead to a refund clause, while a question about ending an agreement may lead to a termination clause. Page references stay attached to the original excerpts. Wording matches also help preserve exact names and identifiers.

The first meaning search downloads a quantized multilingual model and indexes the document on your device. Progress is shown and processing can be stopped. Later questions reuse the in-memory index instead of reprocessing every passage. A long document can take considerably longer to index than a short letter. This search model is separate from the larger optional answer-writing model. Turn off Search by meaning to use wording-only lookup without downloading the search model. Explicit page requests also bypass meaning indexing.

Meaning similarity is not proof that a passage answers the question. Related clauses can disagree, apply to different parties, or describe exceptions. The interface labels the results as supporting candidates and lets you inspect their pages. If search cannot locate strong evidence, it says that the search was unsuccessful; it does not claim the information does not exist in the file. Scanned text can still be missing or misread, so check extraction and the selected OCR language before concluding anything about the source.

The optional local language model writes a short answer from retrieved passages. It is not an unrestricted assistant and it does not read unselected source material while writing. Citation and number checks catch some errors but cannot guarantee correct reasoning. If generation fails, you still receive the retrieved source excerpts. Do not rely on it as a calculator or as the final interpretation of a complex agreement.

Ask a question and verify the evidence

1. Open the PDF and let the tool extract its text. For scanned pages, choose the recognition language; OCR runs automatically inside chat.
2. Ask a focused question in your own words. Keep Search by meaning enabled for paraphrases, or request a particular page to inspect it directly.
3. Read the quoted passages and their page references. Compare the relevant sentence with the actual page, especially when tables, footnotes or multiple columns are involved.
4. If desired, initialise the local model and try a focused question. Keep the tab open while the model loads and works.
5. Treat a missing answer as a reason to inspect the source, narrow the request or improve text extraction. It is not evidence that the document contains no relevant information.

Privacy, accuracy and practical limits

Local processing keeps document contents out of a remote inference service, but it does not make a misleading source or poor OCR accurate. Extraction can separate a table value from its label, and a page citation may point to a passage whose meaning depends on the next paragraph. Check amounts, dates, names and units directly before relying on them.

The model and application must be available before offline work can succeed. Do not close the tab and assume every browser will retain all downloaded resources indefinitely. Browser storage can be cleared or evicted, and a fresh session may require downloads again. On a modest device, wording-only lookup remains available while you decide whether the first-use meaning index is worth the wait. For the broader privacy distinction between local processing and network-free browsing, read local PDF processing and privacy.

Frequently asked questions

Are my document questions sent to a cloud AI service?

The document-chat implementation processes its text and questions locally. The optional model itself must be downloaded before local inference can run.

Can it answer questions about scanned PDFs?

Scanned pages are recognised automatically on the device. Choose their recognition language and check recognised numbers and names against the source before trusting the answer.

Why does it return quotations instead of a polished explanation?

Wording-only lookup shows passages without a model download. Meaning search also shows source passages, ranked using both meaning and wording. Optional local AI writes an explanation with citations. If generation fails, exceeds its time limit or returns invalid references, the tool falls back to relevant excerpts. Read the cited page to verify the explanation.

Does it calculate totals across an invoice pack?

Do not rely on it for arithmetic. It locates relevant source passages; verify figures and perform any required calculation using a suitable, checked workflow.

Will it always work after I disconnect?

An initialised session can run local inference without a network connection. Fresh sessions depend on whether the application and required model resources remain available in the browser.

Does a search miss prove the answer is absent?

No. It means the tool could not locate a reliable supporting passage with its current extraction and search. The answer may span multiple pages, use unusual terminology, or exist only in an image that OCR failed to read. Try the relevant page range, inspect the original page, or ask each part of a complex question separately. Similarity thresholds and model limitations can also cause misses; no search system guarantees every answer will be found.

Is the entire PDF sent to the model provider?

No. Model files are downloaded from the provider, but document text, questions and their embeddings stay in your tab. After the search model is initialized, network access in that worker is disabled. Changing the document or leaving the tool destroys its index. Browser caching may retain model files, not the document index. A new session may need to download assets again if the cache was cleared. ## Read the evidence and manage local downloads

Supporting passages appear as short highlighted excerpts. Expand Read full supporting context to see the complete passage and its page links. Repeated excerpts are grouped together while their page references remain available. An excerpt is source evidence, not independent verification of the answer. Read the surrounding context when a condition, exception or date could change its meaning.

The Model downloads & storage panel can inspect cached model files, check published weight sizes, and remove recognized model downloads from this browser. Size checks send model identifiers to their public hosting service, not document content. Tokenizers and runtimes add to the listed weight size. Cached files can be incomplete or evicted by the browser; successful initialization is the readiness check. Removing cached files does not unload an already running model from memory. Close the tab to release it, and expect a download the next time that model is needed.