DEV Community

Cover image for Photos and voice memos from your phone turn into text when you import them (This week in Graphium, Nov 9 - 15)
Masaya Kumagai
Masaya Kumagai

Posted on

Photos and voice memos from your phone turn into text when you import them (This week in Graphium, Nov 9 - 15)

This week I worked on three moments: collecting records, checking what the AI answered, and showing something to other people later.

1. Import a photo or a voice memo from your phone, and it is already text

When you import the photos and audio that arrived from your phone, the text in the photos is read and the audio is transcribed, without pressing anything more. Photos can then be found by the text read from them. If that extra press were left, I think the records would end up collected but never read again. The reading runs in the desktop app, and one audio file can be up to two hours. (v0.137.0)

The inbox on the desktop: a photo, a memo, a URL and a video that arrived from the phone, with buttons to import the selected ones or all of them

2. See where the AI looked, on the map of your own notes

When you chat with the AI without opening a note, the graph beside the chat lights up the pages the answer read and the pages it used. Pages outside the folder you chose are dimmed, so you can also see whether it stayed in scope. Whether I can trust an answer depends on where it looked, so I show that on the map as well as in the written sources. The graph sits beside the chat in a wide window. (v0.128.0)

The chat view: the answer with its sources in the middle, and on the right a graph where the pages the answer read and used are lit and pages outside the chosen folder are dimmed

3. Jump from a sentence the AI wrote to the passage it came from

On a topic page rewritten from its sources, click the source at the end of a sentence and the PDF opens at the passage behind that sentence, with the passage marked. I ask the AI to copy the passage it paraphrased word for word, and Graphium checks it against the full text of the source before saving. That check does not use AI. A passage that cannot be found in the source is not dropped silently; the sentence gets a note saying so. (v0.133.0, v0.135.0)

A topic page where each sentence ends with a link to the note it came from

4. Read and write the same notes from the AI app you already use

The ChatGPT desktop app can now be connected with one press of a button in Settings. It works in Work mode conversations. I want the memory to stay on my own machine rather than on the AI company's side, and to be the same memory whichever AI I use. In apps that can draw a view inside the conversation, such as Claude Desktop, the assistant can show the whole graph and mark the pages it read. (v0.134.0, v0.136.0)

The whole graph drawn inside a conversation: pages the assistant read are pale green and the links it followed stay as thick green lines

5. Show later that this content existed before a certain date

Seal a saved point in the history as evidence, and it gets outside timestamps and can be exported as a zip to show to a third party. I do not promise that it cannot be deleted; I only promise that a change can be detected. The timestamps are attached in the desktop app. (v0.127.0)

Two rows of the history: the upper one has a Sealed mark, the time it is proven to exist before, and buttons to verify and to export


The full list of changes, small fixes included, is in the release history of the manual.

https://kumagallium.github.io/Graphium/manual/release-history

Graphium is open source on GitHub.

https://github.com/kumagallium/Graphium

Top comments (0)