posible chart error and a quastion

[Ordoumpozanis]

Hi,
Thanks for this great content

while i was reading the /01_atoms_prompting.md
in the chart of quality - tokens is presented a negative slop, that means that as the tokes increase the quality drops. this is the opposite of the example on summarisation where as tokens increase in the max roi area the quality increase that i think is the correct one. I think the chart shows the Loss vs Tokens instead of the Quaity vs Tokens by accident

also you mention token.
what token ?
-input of the prompt command only ?
-input of the prompt command + context ?

for example in summation example with the same exact input of the article context the difference on tokens is only the difference on the prompt part only [some tokens] .

the logic answer would be if i understood correct as more context i give to the summary prompt the best result you may get, bu this does not apply on a summary.. is a bit contusing this part. if you could explain it could be helpful.

thank you

[davidkimai]

Hey @Ordoumpozanis

Thanks for your close reading and thoughtful question!

The chart in /01_atoms_prompting.md is correct: as you give the model more useful information (both the prompt and the text to work with), the quality of its answers goes up quickly at first.

After a certain point (“maximum ROI zone”), adding even more tokens only helps a little (“diminishing returns”), but quality does not go down unless the extra text is irrelevant or confusing.

When the guide says “tokens,” it means the total number from both the prompt and the input/context. For tasks like summarization, it’s usually the context (the article or text to be summarized) that matters most, but also how you phrase the prompt. So, adding more of the right context generally improves results up to a point, and the chart is showing how those improvements slow down.

The main point is that quality rises then flattens out as you add more relevant tokens.