Reading the answer

Under and next to an answer, MindDuck shows where the model got its information and what changed in your data.

Fragments from memory

Before answering, MindDuck itself searches memory for what relates to the question and passes what it found to the model. These fragments are shown in the report under the answer: N fragments from memory. Each one leads to its source: a message in another chat or a place in a document.

Sources under the answer

If the answer relies on your past conversations, source cards appear under it. Each shows the chat title, the date of the message and the sentence the answer relies on. Imported conversations carry a mark such as Imported from ChatGPT · 14 Mar 2023. Click a card to open that message in its chat.

The quote is picked again on the device whenever you open the answer: MindDuck doesn't keep the text of found fragments in the answer's history.

App actions

The model can not only answer but also act in the app: search memory, read a document, create and edit documents, suggest facts, create projects, search the internet. Everything it did is listed in the App actions block. Each action leads to its result: opening the document, the fact, the page.

If the model called nothing, it says so: The model did not call tools, so no app data changed.

Links in the text

Links to chats, documents and facts in the answer are clickable: they open the source. External links open in the browser.

Usage

Under the answer you see how many tokens went to the request (in) and the answer (out), the approximate cost and the response time. The cost is approximate; the exact amount is in the provider's account. See What it costs.

Work details

Work details shows how the model worked: the steps, the time and, if the provider passes them, the model's reasoning. What is shown is set in Settings → Assistant → Model work display:

• Detailed: the current step, the full timeline and reasoning;
• Compact: only the current step and time while the answer runs;
• Developer: everything, plus technical event data.

Waiting for the model API to respond… means the request is with the provider: a reasoning model may think for a long time before it starts writing. The app hasn't frozen.