Yes. A local AI model can use documents that were never part of its training when the app supplies relevant text with your question. OGAD (Off Grid AI Desktop) does this through Projects and document search. You add the files, the app indexes their text locally, and the selected local model uses retrieved passages to prepare an answer. The model does not need to be retrained for each document.
Download OGAD | Desktop releases
What would you like to do with Off Grid AI?
Want help with this workflow, or a feature you would like us to add? Email support@offgridmobileai.co with your devices and the task you want to improve.
Understand the difference between training and context
Training shapes a model before you use it. Context is information supplied for a particular answer. A document assistant can use context from a new file even when the model has never seen that file before.
Suppose your team wrote a project brief yesterday. A downloaded model cannot have learned yesterday's private brief during its earlier training. You can still add the brief to an OGAD project and ask about the deadline it states.
The important limit is that the model answers from the text made available for the request. It does not automatically absorb every file on your computer or permanently learn the contents in the way training would.
What happens when you add a document?
The app extracts text, divides it into smaller pieces, and builds a searchable representation. When you ask a question, it retrieves relevant passages and includes them in the model's context.
This process is often called retrieval-augmented generation, or RAG. The useful idea is simple: find source material first, then generate an answer with that material available.
The answer can cite a filename and text part so you can inspect the evidence. A citation is a route back to the source, not proof that the model interpreted it correctly.
Prepare a readable source
Use a text-based PDF, DOCX, TXT, or Markdown file. Start with one short document containing an answer you know.
The desktop document extractor reads PDF text and DOCX text. An image-only scan needs a suitable text version for this path. Charts, diagrams, and layout may not survive text extraction with their full meaning.
Check the file's version before import. If the brief changed, use the current approved source rather than relying on an old filename that says “final.”
Set up a local document project
Install OGAD and download a local text model that fits your computer. Complete the initial local search setup while connected. Projects and document chat are core features.
- Select the downloaded local model in Models > Text.
- Open Projects > New project, enter a name, and press Enter.
- Open Knowledge & settings > Knowledge base > Add files.
- Add the source and wait for indexing.
- Keep its retrieval switch enabled.
- If Include captured memory is available, turn it off for this document-only test and save.
- Open Chats > New chat inside the project.
The Projects interface provides these controls. Earlier conversations in the same project can also supply context, so use a fresh project when testing a clean source set.
Ask a question with a checkable answer
For the fictional new brief, try:
What review date does this brief state, and what must be completed before that review? Name the source file and quote the relevant passage. If the retrieved text does not establish a condition, say so.
Open the original brief and compare the answer. Check the date and the dependency separately.
This small test shows the mechanism more clearly than asking the model to “understand the entire project.” The model can use a new fact because the app supplies it as context, not because the model secretly knew the document.
Know why an answer can still be wrong
Several stages can fail:
| Stage | Possible problem |
|---|---|
| Extraction | The file has little usable text |
| Indexing | The source is not ready or enabled |
| Retrieval | The relevant passage is not selected |
| Generation | The model misreads or invents a detail |
| Review | The reader accepts a citation without checking it |
A better model may help some generation problems, but it will not fix a missing text layer or the wrong source version.
For a weak answer, first check the project and file status. Then ask a narrower question using the document title, section, or exact term. Inspect the original passage before changing several settings at once.
Do not expect every page in every answer
Retrieval selects passages within a context budget. A broad summary may miss an appendix, an exception, or a detail near the end of a long document.
For full-document review, use the table of contents as a checklist. Ask about each section, check the answers, and combine the reviewed notes afterward.
If a section is repeatedly missed, add a checked text extract or paste the relevant passage directly. Very large files may also need splitting because import and context limits are separate constraints.
Keep source knowledge and model knowledge distinct
A model may add general information from its training. For a source-specific task, make the boundary explicit:
Use the supplied project documents for factual claims about this project. Keep any general suggestion separate and label it as a suggestion. Do not fill missing project facts from general knowledge.
Review whether the output follows that request. A prompt can guide behavior but does not guarantee obedience.
If the document concerns a specialist decision, use the model to locate and explain text, then obtain the appropriate review. New-document access does not make the model an authority on the decision.
Update the source when the work changes
Adding a document is not the same as connecting a live source that updates automatically. When a new approved version arrives, review the project files and disable or replace obsolete retrieval sources as appropriate.
Keep historical material clearly labelled if you still need it. A date in the filename helps, but the model should not be expected to resolve every version conflict by itself.
Try it with a document the model could not know
Download OGAD and add a short note you wrote today. Ask a precise question about it and verify the answer. That demonstrates the useful capability: local AI can work with new private material through supplied context, while you keep checking the source.
Top comments (0)