The semantic search sounds useful, but I very much doubt that the average newcomer will use it or even be aware of its existence, hidden as it is within the Advanced Search and even then only available after you have done an Advanced Search and changed the previously hidden Sort by parameter.
Keep in mind that we got along very well for decades with finding web content through keyword searches alone. It isn't the end of the world if a newcomer doesn't discover the feature.
Advanced search isn't really that hidden. When you do a search via the field on the header that returns more than five results, the dropdown has a "More..." link that leads the user to the advanced search.
Note what I wrote in my previous reply:
So the AI-based search results will be accessible by default for newcomers. It is only in the case that the user has already changed the selection of the "Sort by" menu that the AI-based results are not accessible. The type of user who is advanced enough to want to adjust the sorting of their results is also likely to be sufficiently savvy to figure out how to access the AI-based results.
Note that the feature is still surfaced to the user even in the case where something other than "Relevance" is selected in the "Sort by" menu:

There is actually an "AI embeddings semantic quick search enabled" site setting that makes the feature available via the "quick search" field on the forum header:

I enabled this option at the time I enabled the "AI embeddings semantic search" feature per @6v6gt's suggestion. However, it didn't produce the result shown in the video demonstrating the feature that was published by Discourse:
I concluded that the feature must be broken, and so didn't feel it was worth mentioning in this discussion. However, I now see that it does work, but just with a different behavior from what is shown in that three year-old video. The AI-based search results are now automatically included in the results shown in the "quick search" field dropdown in the case where the search query doesn't return sufficient results from a keyword search alone to fully populate the dropdown.
I actually prefer this new design of the feature because it is seamlessly integrated into the forum UX instead of forcing the user to first perform another operation just to access those results. One of my gripes with the AI era is the cluttering of user interfaces just to show off that the AI hype train has been boarded (I now have an instinctive negative reaction to seeing the
emoji that has been co-opted as the symbol of AI). The user of the quick search only wants to quickly access some good search results. They don't care what specific method will be used to produce those results.
"Site AI" needs another name. Right now there is a trough full of AI slop in the topics that is getting deep.
Maybe Arduinoplasty, or AVR1138, or HAL-R4, but not BOTox... juust anything but AI, plz.
I'm not sure I understand what you mean by "site AI". Are you referring to the wording of the title of this topic "This site needs more AI !"? Or are you referring to something else?
The initials. They are not wanted.
It is now implemented or have I misunderstood something ? I've just tried the test phrase below in the "Quick Search" bar on the top right and it does not bring back useful results because it does not appear to do synonym susbtitution, in this case "twice" (which most people have used when reporting this phenomenon ) for "2 times".
Of course, when I click on "more" then it goes into a full "Search" GUI and there I can select the "new" related results feature which bring in the "*AI" result set.
Where it now claims "*AI" this type of search method has been in use for decades. A thesaurus is used to find synonyms and these are used to enhance the search phrase. Maybe there is scope to argue about what has been used in the background to build the thesaurus.
In order to communicate effectively, we must use established terminology. If I write something like "AI-enhanced searches", the reader will have a good idea of what I am trying to say. If I instead write "Arduinoplasty searches", then I will not have achieved communication with others.
I have been a bit frustrated to see the change from using established terms like "machine learning" and "computer vision" to "AI" as everyone tried to capitalize on the AI hype train. However, I have also noticed that the accurate use of more specific terms like "LLM" is becoming more challenging now as the diversity of low level aspects of AI increases, while the high level user interfaces become more abstracted.
For example, in 2022 it would have been accurate to use the term "LLM" when referring to ChatGPT. However, the ChatGPT service is now actually a collection of various types of models and tools. The user prompt may be broken up into multiple separate tasks, which are each handled by the appropriate system, and not all those systems are necessarily large language models. Obviously I don't want to go into all that when I only want to refer to chatbot services in general, so "AI" is actually the most appropriate term.
The ability to get AI-based results from the quick search has been present from the moment I enabled the "AI embeddings semantic search" feature five days ago. I just didn't realize it was working until I happened to perform a search that only returned a couple of keyword-based results a few hours ago.
I thought I should add a post mentioning the existence of this capability, since my previous communications indicating the AI-based results were only available from the advanced search page had resulted in discussion here.
The reason is that your query produces enough keyword-based results to fully populate the quick search dropdown.
Try running this query from the quick search field:
"The procedure entry point GetPackageFamilyName could not be located KERNEL32.dll"
The query only returns three keyword-based results, so the AI-based results are also shown in the quick search dropdown.
Where it now claims "*AI" this type of search method has been in use for decades. A thesaurus is used to find synonyms and these are used to enhance the search phrase.
I understand that, and your complaint that the forum's keyword search does not take the "fuzzy" search approach used by other search engines like Google for many years prior to the AI era. However, the "AI embeddings semantic search" feature of the Discourse platform does truly use AI.
It is actually pretty interesting how it works:
An embedding model creates a numerical representation of the semantic meaning of every forum post. This information is stored in the forum database (meaning each post only has to be processed once, rather than redundantly for each search query).
An LLM is used to generate a hypothetical post from the search query. In the case of the query you ran:
arduino restarts 2 times when a new sketch is loaded
This is the hypothetical post that was generated:
Hey everyone, I'm running into a strange issue with my Arduino Uno. Every time I upload a sketch, the board seems to restart twice before actually running the code. The first reset happens as expected when the bootloader jumps to the application, but then it resets again a second later.
I've tried:
- Disconnecting all peripherals
- Using different USB ports/cables
- Re-installing the Arduino IDE
Has anyone else experienced this double-reset behavior? Is it related to the auto-reset circuit or maybe a memory leak in my
setup()function causing a watchdog reset? Any advice would be appreciated!
The same system used to generate embeddings for the real forum posts is then used to make an embedding for that hypothetical post. This process is called "Hypothetical Document Embeddings (HyDE)".
The embedding of the hypothetical post is then compared against the embeddings in the forum database, and the posts with embeddings having the highest "nearest-neighbor similarity" to the HyDE are selected as the AI-based search results.
The post embeddings (but not HyDE) is also used to generate the "Related topics" list you see at the bottom of each forum topic.
we must use established terminology
True... but the wrong definitions need to be given another name. Look at the now broken word "drone." In the "bee" world, a drone is given food and has one role. In the military targeting world, it was given AZ, EL and fuel... then it "sounded cool" to call the highly sophisticated, EOIR, PTT/PTL GPS, Satellite, unmanned aircraft "drone" that soon stuck with the hyper-smart, mass RC quads that fly in formations by the thousands.
The word "drone" is not recoverable.
So too is the "I" word.
I don't get exactly the same results as you in that I don't see the message:
Neither have I found a "preference" which could explain the difference. I have also tried in another browser without being logged in so "dark mode" was not used. This is, anyway, what I see.
Interesting what you have found regarding the search process, that is it is building a fictitious forum post out of a search phrase then scanning a database of pre-analyzed forum posts to get a match. I guess that is similar to a sort of "distillation" process where one AI model "picks the brains" of another AI model during a training process by inventing a series of user queries.
I don't get exactly the same results as you in that I don't see the message:
I see. The problem is that my post #30 in this topic created one additional keyword match result for the search, bringing the number of results up to four. Apparently the inclusion of the AI-based results only occurs if there are three or less keyword match results.
Try this query instead:
"the procedure entry point getpackagefamilyname could not be located in the dynamic link library kernel32.dll"
There are currently only two results for that query, so even with the one additional result this post will create, it will still be under the threshold for inclusion of the AI-based results.
I see. The problem is that my post #30 in this topic created one additional keyword match result for the search, bringing the number of results up to four. Apparently the inclusion of the AI-based results only occurs if there are three or less keyword match results.
OK. Thanks. That seems to have been the explanation. I got a whole stack of results with that search phrase. The algorithm for deciding whether to include "AI" results or not isn't very intuitive. The problem would be solved by the not trying at all to differentiate between "AI" results and non-AI results and simply provide the best matches however found. Google and probably every other popular search engine would be a good example of that desired behaviour.



