I am a frequent user of Kagi Assistant, primarily for text summarization purposes. However, the recent implementation of the following feature has significantly degraded the usability for my use case:
"Larger copy-pasted content is now automatically converted to a .txt file and works like any other attachment. This ensures the full original content is always preserved, even in very large threads that hit context window limits."
I would like to report three critical issues that have emerged from this change:
Issue 1: Pre-summarization of Content Before AI Processing
The most severe problem is that the content of the .txt file appears to be summarized and significantly reduced before being passed to the AI model. As a result, I am receiving a "summary of an already summarized text" rather than a summary of the original content.
This behavior is fatal for text summarization use cases. The AI cannot produce an accurate summary when it only has access to a pre-digested version of the original text.
Issue 2: Thread Titles No Longer Reflect Input Content
Previously, thread titles were automatically generated based on the actual content of the input text. Now, titles are generated based on the prompt text or filename instead (e.g., "Discussing message.txt").
This makes it difficult to distinguish between threads in the history and forces users to manually rename every thread to maintain a useful archive.
Issue 3: Character Encoding Issues with Japanese Input Text
When pasting Japanese text that gets automatically converted to a message.txt file, the file content displays correctly initially. However, after reloading the thread or navigating away and returning to the thread, the displayed content of message.txt becomes corrupted with garbled characters (mojibake). This prevents users from reviewing the original input text.
This appears to be a character encoding issue in how stored .txt attachment content is retrieved and displayed, specifically affecting non-English languages.
Example Thread
Here is a test thread demonstrating these issues:
https://kagi.com/assistant/81581014-2452-451e-ac2d-8b06c4d79046
Thread Title: 入力された文章の要約
(Summary of the entered text)
The Desired Functionality
Full content preservation for AI processing: It would be ideal if either an option were available to disable automatic conversion to .txt files, or if the complete, unmodified original content could be passed to the AI model when processing attachments — not a pre-summarized version.
Content-based thread titles: Thread titles should be automatically generated based on the actual input content rather than generic prompt text or filenames, enabling users to easily identify and navigate their thread history.
Proper character encoding: Character encoding should be handled correctly for non-English languages in stored attachments, ensuring content remains readable after reloading or switching between threads.
