[BUG] File attachments cause massive token bloat — file data resent with every API call

[MinaMan01]

System details / 系统信息

Environment - DeepChat version: v1.0.7 (AppImage) - Model: deepseek-v4-flash - OS: Linux Mint 22.3 (Zena) — kernel 6.17.0-35-generic - Desktop: X-Cinnamon - CPU: Intel Core i3-1115G4 - RAM: 15 GB - GPU: Intel UHD Graphics G4 (Tiger Lake)

What's wrong? / 出了什么问题？

Describe the bug

When a file (image, PDF, document) is uploaded via DeepChat's message input box, the file's content is embedded inline in the message history. Every subsequent API call resends the entire message history — including the full file data from every previous message. This causes:

• 500KB+ per request for a single image (base64-encoded)
• Cache misses balloon because file content changes the context hash
• Token consumption spikes even when the file is no longer being discussed
• Chats become "heavy" and slow after just a few file attachments

To Reproduce

1. Upload an image (e.g., 500KB PNG) to a new chat
2. Ask a question about it → ~500KB cache miss tokens
3. Ask an unrelated follow-up → 500KB file data is resent AGAIN as cache miss
4. Repeat → every request burns the file data as uncacheable context

Expected behavior

File content should either:

• Be stripped from old messages after the model has processed it, OR
• Be stored externally (referenced by path/ID) rather than embedded inline, OR
• Old messages with large file attachments should be eligible for compaction

Workaround

Instead of uploading files to the chat UI, provide the file path directly:

"Read /home/user/Downloads/image.jpg"
"Process /home/user/Documents/report.pdf"

The assistant reads the file directly from disk via its tools. The file data never enters the message history, keeping token usage lean.

Environment

• DeepChat version: v1.0.7 (AppImage)
• Model: deepseek-v4-flash
• OS: Linux Mint 22.3 (Zena) — kernel 6.17.0-35-generic
• Desktop: X-Cinnamon
• CPU: Intel Core i3-1115G4
• RAM: 15 GB
• GPU: Intel UHD Graphics G4 (Tiger Lake)

[zerob13]

Thanks for the detailed report and the suggested workaround.

This is not considered a bug in the current implementation. When a file is manually attached to a message, DeepChat intentionally embeds it as part of that message so the model can process it together with the conversation context.

That said, this is a valid use case. Large attachments being carried into later requests can make conversations heavier, especially when the file is no longer relevant. We will consider improving this flow in the future, allowing users to choose how an attachment should be handled, such as embedding it into the message or referencing it in a lighter way.

I’ll treat this as a feature request rather than a bug. Thanks again for raising it.