Take it step by step
- Read the model size and compatibility information before choosing a pack.
- Download the model explicitly and let initialization finish.
- Choose a document, then ask one narrow question at a time.
- Follow page references and verify the answer against the source before reusing it.
A situation to try.
Ask “Which page states the delivery date? Quote only the relevant sentence.” If the date is absent, the useful answer is that the document does not establish it.
Use the tool’s synthetic example first where available. The example here explains a decision; it is not a reported benchmark result.
Before you call it done.
- Check every cited page.
- Distinguish extracted text from a generated inference.
- Keep context short when memory is limited.
Know the limits.
Local models can hallucinate. Scanned documents may need OCR first, and unsupported browsers cannot run every model.
This tool’s current boundary: Explicit local-model download required. Context is bounded; review facts and citations. Searchable text only; OCR and embeddings are separate packs.
Evidence and scope
Technical source guidance and task-specific output checks. These do not establish that every file, browser or device passes.
Use the checks above on your own exported result. A source explains the format or mechanism; it is not proof that this export preserved every feature.
Technical reference: WebLLM: local browser inference ↗
Choose the right tool for this step.
Chat with PDF
Ask questions about your documents using local AI.
Explicit model downloadWhy use it: Ask a narrow question with page references.
Before you start: Local model download and compatible hardware required.
Searchable OCR
Recognize scan text and add a searchable text layer.
On your deviceWhy use it: Recognize text when the PDF is image-only.
Before you start: OCR errors can affect generated answers.
Preparation tools process locally. Official service links take you to the authority’s website. Inspect any exported copy before sharing it.