Bug Report: AI X-Ray Progress and Character Hallucinations

[Rogerthis]

I've encountered two main issues with the AI X-Ray feature.

First, the progress display was inaccurate. Despite being at 20% of the book, the progress bar showed 1%. I managed to correct this by going to 'Reading statistics' and selecting 'Synchronise now', which updated the percentage to the correct figure the next time I used the feature.

My second, and more significant, issue is with the quality of the results. The AI-generated content is filled with hallucinations, inventing characters and providing incorrect details about relationships. This was particularly frustrating as I used the feature to understand the minor characters, but the AI failed to recognise them.

This leads me to a couple of questions:

Is the AI only able to identify characters if it finds information about them on the internet?

Is it possible to upload the .epub file directly to the AI to provide it with the text as a reference?

[omer-faruq]

Unfortunately, this is a limitation of the method we're using. I don't think we can provide the entire book as data to the AI ​​via the API. Even if we could, that would require a significant number of tokens, which would be quite expensive, and the free version doesn't have sufficient limits. And it will take a long time to load the content of an ordinary book into AI as data.

Instead, you could try uploading it to platforms like Google's NotebookLM and using an X-Ray prompt.

I tried to do this by first converting the epub to an MD file using MarkItDown (Microsoft's freeware), then uploading it to NotebookLM and using our X-ray prompt. I think it took about ten minutes to get a result from it. NotebookLM does not allow importing the EPUB format, and converting the file to an MD file decreased the size from 5.2 mb to 700 KB, so it is useful, I think. Because the book consisted of multiple independent stories, the result was a bit inefficient, producing a single, mixed-up result for all the stories. However, this could easily be resolved by creating a separate file for each story, or specifying this in the prompt might produce a more useful result.

[Rogerthis]

Thanks for the explanation.
I used Calibre to convert my .epub to a .txt, and I used this to upload to Gemini using your prompt.
It replies back within about 30 seconds.

[omer-faruq]

What is the size of the txt file? If it is a large one and the whole text is considered, I think it is a good result. 👍

[Rogerthis]

The text file that I used to upload to Gemini was 517KB, which is bigger after converting from the .epub, at 394KB
The file was 2232 lines long. The paperback was 272 pages.

[omer-faruq]

Good news is that I modified code to add a functionality to consider text of book on x-ray and recap.
There is a setting for it on the settings page :

also there are two settings on the config about this

[omer-faruq]

Update: the options on the settings menu have moved into the Assistant menu: