DeepSeek-OCR integration for the compression of tool_responses.

[magikRUKKOLA]

Suppose we are using the Ling-1T or any other LLM in the loop of tool-calls. Suppose we are having the prompt:

Output the mandelbrot fractal using python in ascii format /python

For example the DeepSeek-R1 would chose the 80x40 resolution. So when we will get the tool-response and it would go to the model's prefill, it would eat up 3.2k tokens (roughly).

What if, for example, for the long tool-responses we would just convert it into the graphical representation for the DeepSeek-OCR to produce a text/markdown as a short text representation of the tool call execution. And instead of the original long tool call response we would send what the DeepSeek-OCR would produce.

[1] https://github.com/deepseek-ai/DeepSeek-OCR