You remember a participant describing a confusing handover. Which interview was it, and what did they actually say?

OGAD (Off Grid AI Desktop) lets you add transcripts to a project and ask questions against the indexed text. A local model can help you find relevant passages without uploading the interview collection to a cloud AI provider. Check the source before using a quote or turning an observation into a finding.

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.

This suits a small research consultancy, design studio, or product team that needs to return to interview evidence between sessions. It gives you a way to find the material behind an idea while the final interpretation stays with your researcher.

What can local AI help you find?

It can retrieve passages relevant to a question and use them in an answer. Ask about a concrete task, event, or difficulty. You can then inspect the transcript and decide whether the passage supports the point you are making.

Suppose you have interviewed people about handing customer issues from support to operations. Some describe missing context. Others describe waiting for an approval. Those may require different product changes even if both use the word “delay.”

Useful questions include:

• “Which passages describe information being lost during a handover?”
• “What did participants say they do while waiting for approval?”
• “Find examples where a handover worked well, including the reason given.”

The third question matters. Searching only for failures can leave you with a one-sided account. Ask for the conditions that made the process work as well as the points where it broke down.

How should you prepare the transcripts?

Use one clearly named text file per interview. Preserve participant codes, speaker labels, and section markers when you already have them. Remove unnecessary identifying details before import if your research process calls for that.

A practical filename is study-handover-p04-2026-09-18.txt. Keep the participant-code key separately. The filename makes returned sources easier to trace without placing a person’s name in every answer.

Material	How to prepare it
Checked transcript	Use TXT or Markdown with consistent speaker labels
Interview notes	Label them as researcher notes, with the session date
Research plan	Add it if the model needs the study question and definitions
Earlier findings	Label them as interpretations rather than participant statements
Scanned transcript	Create and check readable text before import

OGAD supports text documents, PDFs with extractable text, and DOCX. A scanned page needs a usable text layer; do not assume document import will read every image inside a PDF.

Avoid mixing unrelated studies in your first project. One study and a few checked transcripts make it easier to judge whether an answer used the right evidence.

How do you create a searchable research project?

Install OGAD, choose a downloaded local text model, and add the transcripts to a project. Finish the local indexing-model setup while connected before testing offline use. The app needs its model resources on disk to process the material without a connection.

Projects and uploaded-document search are free core features. Supported Mac and Windows builds are linked on the download page. 0.0.54-beta.108 also provides Linux packages as a beta; this workflow does not depend on Pro screen or meeting recording.

1. Open Models > Text and select a downloaded local model.
2. Open Projects > New project, enter the study name, and press Enter.
3. Open Knowledge & settings > Knowledge base > Add files.
4. Choose the prepared transcripts. Wait for indexing and leave their retrieval switches enabled.
5. Open Chats > New chat inside that project.

The project implementation provides the file list and retrieval controls. Project chat also has conversation context, so use a fresh project for a new study and keep speculative analysis clearly labelled.

What should your first question look like?

Ask for a small evidence table rather than a finished research conclusion. A table makes it easier to check each claim against a file.

Search the uploaded interview transcripts for descriptions of losing information during a handover. For each relevant passage, give the filename, participant code if present, a short exact quote, and the surrounding situation. Keep interviewer suggestions separate from participant statements. Say when the evidence is unclear. Do not claim this list includes every interview.

Then open the cited source and check the words. The expected result is a set of useful passages to investigate. The source check establishes whether each item belongs in your analysis.

For the handover example, record three things in your own research notes: the situation, the observed difficulty, and the participant’s explanation. Keep your proposed solution in a separate field. This prevents an AI-generated suggestion from being presented as something a participant requested.

Can you count how many participants mentioned a problem?

Do not use a retrieved answer as a complete count. Project search returns a limited set of relevant passages. It can miss a differently worded comment, and several passages may belong to one participant.

The project chat path retrieves a bounded context for the answer. It is useful for discovery, but it is not an exhaustive coding pass over every transcript.

If you need a count, review each interview against a written coding rule and keep a participant-level table. You can use the local model to suggest passages one transcript at a time, then accept or reject each suggestion yourself.

For example, define whether “handover delay” includes waiting for approval, waiting for missing information, or both. Without that rule, a neat total can combine different problems and give you a misleading result.

What if you only have interview recordings?

Create a checked transcript first. On Mac, select a local speech model in Models, set Current model and Spoken language in Transcription settings, then use + > Attach files in chat to process a saved audio file.

Open the resulting text preview and compare difficult sections with the audio. Copy the text into an editor for corrections, then add the checked version to the research project. Speech setup can differ on other desktop platforms.

This audio attachment workflow returns transcription text. Do not assume it supplies reliable speaker identities or exact timestamps. Preserve those from your existing research records where available, and check quote attribution against the recording.

How do you handle weak or conflicting answers?

What you notice	What to check next
A quote is too tidy	Compare it character by character with the transcript
An interviewer prompt becomes a finding	Inspect the speaker labels and the participant’s response
Several sources disagree	Ask for each account separately with its source
An expected interview is absent	Review that transcript directly with a narrower question
An old interpretation keeps appearing	Check project documents and related conversations

For contradictory accounts, ask what differed in the situation. One person may handle urgent requests while another works on routine cases. Preserve that context rather than forcing both experiences into a single conclusion.

Try one question you already know how to check

Download OGAD and add a small set of checked transcripts. Ask for evidence about one task, inspect the sources, and compare the result with your existing notes.

Local inference keeps this processing on your computer when you use local models. It does not determine your recording permissions, retention policy, or what happens when you share the files elsewhere. Keep those parts of your research process explicit.