AI or API integration

[michal_novomesky]

Hello,

is there any plan to implement AI integration to Help&Manual?

Use cases:
1. I am a techwriter and want to sent my H&M topic to chatbot for validation, using a H&M button, returning suggestions to H&M topic as <draft-comment>
2. Customize chatbot API URL and request scope due to privacy reasons (i.e. do not send sensitive information to openai.com and able to use different/private chatbots)

[Tim Green]

Hi Michal,

We don't currently have plans to have HM projects "proofread" by AI chatbots. Most of our customers are professional technical authors and if at all, they would work in the opposite direction, correcting the many errors and incorrect and misleading fabrications in content generated by AI before release.

As to using AI chatbots as an interface for querying documentation: We have done extensive tests and our clear conclusion at the moment is that they cannot yet be trusted for this. Even if you feed the AI system only your own documentation, the number of inaccuracies and complete fabrications of completely incorrect answers they deliver is very high. When using such a system, you can only know whether the answers you receive can be trusted if you already know the answer, which defeats the purpose of the exercise.

On "AI" in general, it is important to know that the current systems are not really artificial intelligence at all. They are large language models, or stochastic parrots, which have no comprehension of the material they are processing, which would be necessary for any application of the word "intelligence". All they do is predict what words and sentences might plausibly follow upon a query, or upon the last sentence they themselves generate, based on complex statistics. This looks very impressive to untrained observers and for some purposes it can actually be useful. But it cannot yet be used to reliably query documentation and provide dependable results to users who do not yet already know the answers they are looking for.

[michal_novomesky]

I'm just thinking it's worth a try, maybe not now but to be considered in the future.

I agree there are many false positive reports by AI chatbots. But they can detect stylistic issues if you feed them with proper rules.

AI chatbots can, for example, autofill more accurate Topic description (variable <%TOPIC_DESCRIPTION%>), to generate a proper <meta> tag in HTML output. Currently, H&M uses first sentence/segments in topic text if it is empty in an .xml topic file.

[Gerold Krommer]

Have you tried deepls English correction program? I pride myself in a decent English (albeit it is NOT my mother tongue) but the alternative suggestions and corrections from deepl very often surprise me positively.

br,

/Gerold

[Tim Green]

Have you tried deepls English correction program? I pride myself in a decent English (albeit it is NOT my mother tongue) but the alternative suggestions and corrections from deepl very often surprise me positively.

We've been looking at it and do plan to integrate it in a future release. That being said I recently tried feeding it some literary text to test it and the result was beyond awful. Over 90% of the suggestions were the sort of thing that a really incompetent, narrow-minded English teacher with no understanding of creative prose would come up with. So use with great caution...

[Gerold Krommer]

> recently tried feeding it some literary text

That is kind of unfair Literary text makes human creativity stand out from computers and that is why they can't handle that.

I fed it technical documentation that I wrote and the results were really good. Very often the sentence was shorter and cleaner and more concise, some alternative suggestion sounded really cool to my (non-native) English ear.

/gerold

[Tim Green]

That is kind of unfair Literary text makes human creativity stand out from computers and that is why they can't handle that.

True. But it's still important to keep the limitations of AI in mind,

[Ash_Happen]

Hi,

Any update on this? AI has moved so quickly over the last year, I am curious to see what might have changed for you, when thinking about AI and Help + Manual?

[Tim Green]

Any update on this? AI has moved so quickly over the last year, I am curious to see what might have changed for you, when thinking about AI and Help + Manual?

Current LLMs are still not a good solution for searching documentation. They will regularly deliver extremely plausible answers that are completely fabricated and wrong.