A folder of papers is not yet a usable evidence table.

OGAD (Off Grid AI Desktop) can help you extract and organise details from papers saved on your computer. Use a local model to draft one row at a time, check the original, and keep each finding beside its method and limits. After setup, the text review can run without sending papers to a cloud AI provider.

Download OGAD

Desktop releases

What would you like to do with Off Grid AI?

Have a feature or use case you would like us to support? Tell us what you want to do and which device you use.

Write to support@offgridmobileai.co, join our Slack community, or talk to us on Reddit.

For a student, independent researcher, or small research team, the benefit is a structured starting point for comparison. The model can help with extraction. You decide what the evidence supports and whether the papers belong in your review.

What should an evidence table contain?

Choose columns that answer your research question. Keep the study’s question, method, relevant result, and limitations separate. A short finding without its context can look stronger or more general than the paper allows.

Suppose you are reviewing papers about how people learn a new workplace procedure. You may want to compare the learning task, participants, intervention, and measure used.

A starting table could include:

Column	What to record
Paper ID and citation	The exact source you reviewed
Research question	What the paper investigates
Participants or material	The population or data actually studied
Method	How the work was conducted
Relevant finding	A result that addresses your review question
Limits	Conditions or limitations needed to interpret it
Source location	Section, table, or page you checked

Adapt the columns to the discipline. Do not force different study types into a field that changes their meaning.

What do you need before starting?

Install OGAD and download a local text model that fits your computer. Save the papers and relevant supplementary material locally. Complete app and model downloads before working offline.

The saved-document workflow uses free core features on supported Mac and Windows computers. Release 0.0.54-beta.108 also provides Linux beta packages.

Use readable text PDFs or checked text versions. The document-processing code extracts text; it does not preserve every equation, figure, table layout, or footnote relationship. A paper that opens correctly in a PDF viewer may still extract poorly.

Keep the original available throughout the review. Record which version you used, especially when a preprint and later publication differ.

How do you extract the first row?

Start with one paper and a limited set of columns. This makes errors easier to spot than importing a large folder and asking for a finished review.

1. Select a downloaded local model in Models > Text.
2. Start a new chat and choose + > Attach files.
3. Attach the paper and inspect its text preview.
4. Compare headings and a few important passages with the original.
5. Ask for a draft row with supporting passages.

Use:

Draft an evidence-table row for this paper using the columns below. Use only the paper. Give a short supporting passage and section label for each factual entry. Mark information you cannot establish as “Not established from supplied text.” Keep the authors’ interpretation separate from the reported result.

Review each cell. The aim is to make checking easier, not to remove the need to read the relevant sections.

How should you handle numerical findings?

Verify the value, unit, denominator, time point, and comparison directly in the paper. Text extraction can separate a number from its table heading or omit a note that changes its interpretation.

Do not let the model silently convert a percentage, combine different samples, or substitute a relative change for an absolute one. Preserve the paper’s reported form until you have checked the calculation yourself.

A useful follow-up is:

For this numerical finding, identify the original value, the group it applies to, the measurement time, and the table or passage behind it. Flag any part that is unclear. Do not calculate a new effect or fill missing values.

If the source relies on a figure the model has not read correctly, inspect that figure yourself. Leave the table cell pending until the needed information is confirmed.

How do you keep different papers comparable?

Use the same definitions for your columns, but keep meaningful differences visible. Two papers can use the same word for different measures. A result from a short exercise may not address performance in an ongoing workplace task.

For the learning example, record whether the outcome is recall, task completion, or a self-reported view. Do not merge these under a generic “improvement” column.

After checking several rows, ask:

Compare these reviewed rows. Identify differences in population, task, measurement, and study design that limit direct comparison. Refer to the paper IDs. Do not rank the papers or infer a combined result.

Use that response to improve the table’s clarity. Any formal appraisal or synthesis needs the method appropriate to your research question.

What if a paper is long?

Work through the relevant sections separately. A model’s context must hold source text, instructions, conversation, and its answer. A visible full-text preview does not guarantee that every passage informs the response.

Keep a checklist for methods, results, discussion, and supplementary material where relevant. Ask about the section you need, verify it, and then add the checked entry to your table.

You can keep the papers in an OGAD project for later lookup. Create it under Projects > New project, add files in Knowledge & settings > Knowledge base > Add files, wait for indexing, and use Chats > New chat. Project retrieval selects passages; it is not an exhaustive review of every paper.

How do you preserve a useful review record?

Save the checked table in your normal spreadsheet or document editor. Include a review-status column so a drafted row cannot be mistaken for a verified one.

Status	Meaning
Drafted	AI-assisted extraction awaiting source review
Checked	Listed fields compared with the original
Needs clarification	A missing or ambiguous detail remains
Excluded	Kept in your record with your stated reason

Keep citations accurate. Do not accept an invented DOI, author name, publication year, or page number simply because it looks plausible. Copy identifiers from the source or verify them separately.

Build one row you can use

Download OGAD and start with one paper relevant to your question. Create a draft row, verify its evidence, and save it with the source locations.

The local model helps you organise the reading. The result becomes useful when each cell says something precise, preserves the paper’s limits, and leads you back to a passage you have checked.