Calibre working AI Model?

[SASSBS]

Hi All,

I am looking for some advice from anyone using the AI features, what AI model are you using that gives accurate results more often than not.

I enabled the AI options and installed LM Studio and used the default model of openai/gpt-oss-20b however no matter what I ask the model for any book it gives me a highly inaccurate answer.

Some of the answers have been very very funny though!

Appreciate any recommendations for models to use.

[Comfy.n]

The model you tried is one of the best available among the free ones. I see a 120b (not the 20b) free model also available when using the OpenRouter provider, can't remember if LM Studio has that one.

But IMO, unless you can find a way for the model to read the entire content, it won't provide accurate answers. For instance, my RSS Reader plugin has an option to send the entire article. Works with some models.

[kovidgoyal]

Gemini 2.5 flash is free too via the google provider for a limited numberof queries per day. Worth a try, since it uses RAG for answer grounding.

[Comfy.n]

I wonder if RAG can work reliably. I tested it just once for summarizing books, using the LM Studio experimental feature, at the time, IIRC. Results were off the track. I think this was the moment when I realised RAG just incorporates summarized *key* information. So, a book summary would be the summary of a summary...

[kovidgoyal]

Yes RAG basically includes search results for the book summary. Assuming the book you are searching for has decent summaries available online I have generally found the results to be acceptable.

[Comfy.n]

Ah, sure. In this case, I wanted to know how much of each chapter it was actually reading. So I asked a bullet-format summary for each chapter.

[SASSBS]

thanks, sounds like we are not quiet there then for AI working with ebooks. I would love a way to have a model read my books and then answer questions.

I tried the Gemini models but the answers were still not great - slightly better but not ideal.

[Capricorn]

No point in using AI when it can be manipulated so easily:
https://www.bbc.com/future/article/2...ook-20-minutes

[Sirtel]

Originally Posted by Capricorn

No point in using AI when it can be manipulated so easily:
https://www.bbc.com/future/article/2...ook-20-minutes

AIs aren't any more reliable than any other information found online, or than an average human, for that matter. Why should they be? The internet is full of lies and fabrications; so is life. As with any information, always check the sources.

[kassandr]

Originally Posted by SASSBS

I would love a way to have a model read my books and then answer questions.

This is actually possible right now. I built an open-source tool called ARCHILLES that does exactly this: it indexes your Calibre library (and/or Zotero, Obsidian, a Folder) — full text, metadata, annotations — and makes everything semantically searchable.

ARCHILLES uses RAG (Retrieval-Augmented Generation), meaning the AI doesn't hallucinate from its general training data but retrieves actual passages from your books before answering. You ask a question, it finds the relevant paragraphs across your collection, and responds with proper citations — author, title, page number. It connects to the AI model of your choice via MCP (Model Context Protocol), so it works with Claude, ChatGPT, or local models.

Everything runs on your machine — no cloud, no data leaves your hard drive. It's MIT-licensed and on GitHub: github.com/kasssandr/archilles. Still early days, but the core is functional. Worth noting that this is a different approach from calibre's built-in Discuss feature, which sends individual books to an LLM for conversation. ARCHILLES works across your entire library at once, so you can ask questions that span multiple books and get cited answers from whichever sources are relevant.