Shopdog AI

Hey everyone. Kiril, from Toronto. I'll lead with the awkward part, since my handle gives it away anyway: I built ShopDog, a small tool that digs repair answers out of factory service manuals. There's AI in it, and I'm guessing that word doesn't win many friends around here (which is completely fair). Confidently wrong advice on a fifty-year-old bike is worse than no advice, so it was built backwards from that fear: answers come from the manual, the specs and torque numbers are over 99% accurate, and when the manual doesn't have an answer it says "the manual doesn't cover this" instead of improvising. The KZ1000 is one of the bikes it covers, which is what first brought me to this board. Here's my deal as a guest in your house: that's the last time I mention it. My goal is to use the tool to answer forum questions as they come up as a way of showing value. So if there's a post made by me in response to anyone's question, it was written by ShopDog. Mods, if you'd rather handle me through Vendor classification, I'm glad to do it properly. And if the honest answer is that the community would rather I didn't post here at all, I'll respect that and quietly show myself out.Mostly, though, I'm here as a guy with a '79 Shovelhead I'm turning into a chopper, which has humbled me on carbs, charging, and one memorable transmission seal. So you'll find me in the forums, trying to give back a little of what forums like this one taught me. Anything I quote will have been checked against the factory manual before I hit post.Glad to be here, however long you'll have me.

Welcome from the UK Kiril, from my point of view as long as you are using the tool to provide information to the members which they can chose whether or not to follow (as with all other advice given here) I don't see an issue so long as you aren't trying to sell your product here.

   This little air-cooled corner of interwebs Motorcycle Madness is populated with bikes the majority of which have a 'Z' somewhere in their moniker.  Therefore, welcome to The aZylum.

   Rabbit hole warning - I'm intensely curious from an IT / document processing perspective:
   ShopDog implies a pre-existing searchable document file.  The last 10 yrs. or so of slmjim's previous IT life was deeply involved in a tangent of printing; creating searchable .pdf files by scanning (sometimes tens of thousands of) hard copy pages to a Fiery server that was running OCR software on the scanned images to RIP them into searchable .pdf documents.

   Are searchable vintage shop manuals somehow created from hard copy? By what process?  Or, is AI finding complete vintage manuals (files) online that predate OCR by decades, making the RIP process unnecessary?

EDIT: Related thoughts & questions:
1) Where would one find, for instance, a searchable file of any edition of the original Kawasaki hard copy Z1 'White Manual' circa early 1970's ?
2) Can the AI include the very useful illustrative figures / photos / drawings / tables (Fig. 1, Table 2, etc.) from the original manuals?
3) Does AI take into account and include info in subsequent hard copy service bulletins that correct / update info in the manuals?

Good Ridin'
slmjim

Hahahaha the aZylum is great!

slmjim, your experience means you'll get the product pretty quickly and hope it's cool! Technology has come a long way recently... Same starting point as your old pipeline: scanned hard-copy factory manuals, since almost nothing from that era has a digital original. But instead of OCR toward a searchable PDF, each page goes through AI vision extraction into structured pieces that keep their metadata: which bikes and years a passage applies to, which section, page, and figure. Retrieval matches meaning rather than text, and the constraint runs the other way too. An answer traces back to specific passages, shows you that trail, and says "the manual doesn't cover this" when it doesn't, instead of guessing.

I also started with some scanned manuals but I also bought some reproductions of popular manuals and used a service to get them scanned. Once I had them scanned I ran the same AI process to ingest them into my database. It was actually quite a complex process to get right and took months of experimentation to get done accurately. 

To answer your three questions:

1. I don't know of a searchable early-70s Z1 White Manual I'd vouch for. Scans exist around forum archives, but they're images of pages, the exact problem you spent a decade solving from the other end. Full disclosure on mine: KZ coverage today is the 81-83 KZ1000/1100 factory manual. The early Z1 books aren't in yet. If you wanted a searchable copy you'd have to buy a physical book and then send them into a service that can scan them for you. Feel free to message me and I can send you who I used... they were pretty good. I think I paid around $15-20 per manual to get a great scanned copy. 
2. Yes. Figures, diagrams, and tables come back with the answers when it makes sense. It definitely adds context and helps out for a lot of questions. If you want to see an example of what it looks like check this out:  beta.shopdogai.com/a/ptxfwq7q8p
3. Not yet... Factory manuals only right now, so a bulletin that superseded a spec after printing isn't in there. It's on the roadmap but I'm still in the early stages so really trying to figure out if it makes sense to built more. 

Just trying to answer questions here and If the document-nerd curiosity goes further, message me and I'll set you up with free access so you can play around with it! I would love for someone with your experience to test it out and try to break it.
 

wrote:

... I don't know of a searchable early-70s Z1 White Manual I'd vouch for.  ...  The early Z1 books aren't in yet. If you wanted a searchable copy you'd have to buy a physical book and then send them into a service that can scan them for you. ...
 

   slmjim here.   The Lovely Z1BEBE's eyes glazed over early on while reading your response, but tech isn't her thing anyway.

   Thanks for the responses Kiril !  Lots of good info & food for thought.  Puts the geek within at ease in many respects and, intrigues at the same time.

   I appreciate the free access offer to tinker with your product.  It's clearly orders of magnitude steps forward from the relatively crude standalone OCR platforms I dealt with years ago.  It'll be in the dead of the coming winter before I'll have the time to start up that steep learning curve.

   Re: the Z1 White Manual; I've seen somewhat disturbing chatter that when bound hard copy publications are scanned for AI consumption, that the binding (spine) is cut off, pages scanned, then discarded.  That may be acceptable for a book that was published in the tens or hundreds of thousands or more, but I would not submit a vintage White Manual (or any other limited-supply volume)  to destruction to get it scanned.  Do you know if a non-destructive imaging option is available?  Example - can photographic images be processed by AI vision extraction?

  On a related note; microfiche of technical documentation sometimes exists alongside the matching hard copy manuals.  So too, do natively-searchable .pdf files of the same manuals that were offered to end users by the equipment manufacturers. 
   Do you know if fiche and / or natively-searchable .pdf files can be utilized for direct AI input?

  To clarify the "natively-searchable" above, I'm referring to a file that had originally been created within, say, a word processor application.  One that had not undergone the sometimes - erroneous OCR process.

  Does a manual comes back as a file from the scan service you've used? If so, in what format and, is it searchable on receipt?

  I'll not bother you with any further questions.  I'll happily retreat to the geek cave & crunch the replies.  If deeper conversations are to be had, perhaps by email if that suits you?

   Appreciate your time!

Good Ridin'
slmjim & Z1BEBE  

While I would not expect Shopdog to give away any proprietary information, I am interested in seeing this topic stay public here at KZR. 
The seemingly sudden arrival of AI in the public sphere has caused quite a bit of chatter and sometimes angst. Yes, I know it is not new technology, but it has only recently come into public view and usage. ( Henry Kissinger published a book or two about it many years ago.) As a mostly retired educator, I have been grappling with some of the deep and important questions the arrival of this technology has raised. I think it would be interesting and useful to learn about developments like this from the inside. 
I hope y'all will consider dragging some of the rest of us down your rabbit hole.

slmjim, no rush at all. Whenever winter sets in and you have the time, it's at beta.shopdogai.com. 

Your scanning concern is legitimate. The standard high-volume process is exactly what you heard: the spine gets guillotined off, the loose pages run through a sheet feeder, and the carcass goes in the bin. Fine for a paperback printed a million times. Feeding a White Manual into it would be vandalism so I absolutely didn't do that. Everything in ShopDog came from PDFs that already existed, or from reproduction manuals I bought specifically to be cut up.

Photographs: yes. An image is an image, and the software doesn't care whether the pixels came off a scanner bed or a camera. A copy stand, a V-cradle book scanner, even careful phone shots if you're patient. What hurts quality is glare and page curl near the gutter, not the fact that it's a photo.Natively-searchable PDFs are the best possible input. If the text layer is real there's no OCR step at all, so no guessed characters and no torque spec quietly mangled at the source. Microfiche: not directly. It has to be imaged first, and then you're back to OCR on whatever that image looks like. Usable, but it's the long way around if a native PDF exists. And yes, the scan services hand back a searchable file, normally a PDF with an OCR text layer under the page image. 

Email any time: Kiril@shopdogai.com.

TexasKZ, happy to keep this in the open. There's no secret sauce. The hard part was never the idea, it was months of grinding through manuals and getting it accurate.

The whole thing in plain terms: Every manual gets broken into pieces, page by page, each tagged with the bike, the year range, and the page number. Those pieces get indexed by meaning rather than keyword, which is why "bike won't charge" finds a section titled "Charging System Inspection." When you ask something, the front end works out which bike and system you mean, then pulls the handful of pages most likely to hold the answer. Only at that last step does a language model get involved, and its job is narrow: read those specific pages and tell you what they say, with the page cited. It cannot answer from its own memory. If the pages don't have it, the right answer is "the manual doesn't cover this."

So the AI part is smaller than the word suggests. Most of the work is cataloguing and routing. The model is just the last mile that reads the page back to you in English.It's built that way because of the failure I was afraid of. Ask a general chatbot for a KZ torque spec and you get a fluent, confident, invented number. It doesn't know it's guessing, and neither do you. Tying the answer to a citation, and letting it say "not covered," is what makes it safe to point at a fifty-year-old bike. 

Kiril,
   What you're pursuing is a tall order, not least of which is the need for much, very detailed riding herd on data in the background by a knowledgeable human.  I'm satisfied that you're on top of things to keep accuracy in and AI data slop, as in 'guessing', out.  That's to be commended.

TexasKZ,
   Truly appreciate your interest & suggestion top keep this discussion public.  I suspect there's at least mild underlying interest from other members that has yet to be expressed.  Was not my intent to keep anything secret.  Everyone is welcome!. 

KZR site Mods,
   May I suggest this AI discussion be moved to Daily Chatter & renamed appropriately to make it easier to find?
   Shopdog's OP in New Members laid the foundation for this ongoing technical discussion that's now evolved far beyond a simple new member intro.  If kept in 'New Members' it will get buried quickly.  A perhaps-abbreviated-version of Kiril's intro OP could remain in the New Members subforum, while the meat of the AI discussion would reside in Daily Chatter.  I'd request that Kiril suggest an appropriate Daily Chatter title for this discussion that would pique interest.

Good Ridin'
slmjim

   

I appreciate the support Slmjim, I'm building Shopdog for myself as well so I definitely want something that's actually going to be accurate! Accuracy is probably my number 1 concern and it's something that's taken the most amount of time. The technology isn't really that complex but getting it accurate is a pretty big challenge. 

KZR site mods - I would love to have more conversations around this! Happy to answer questions and if anyone wants more information I'm more than happy to share. The tread could be titled "Using AI effectively for accurate diagnostics and repair" Or "AI - The helpful companion or the the world ender" Either work! 

Looking forward to continuing these discussions. 

BTW - If anyone's actually interested in using Shopdog in the future. I would highly recommend making an account even if you don't have plans to use it yet. The first people who sign up will have perks for life as a way of saying thank you! I'm still working out what those are, but I'd rather the folks who have been engaged from the start get the perks! 

No that's not me. I'm Shopdogai.com 

Wookie58 wrote:

Welcome from the UK Kiril, from my point of view as long as you are using the tool to provide information to the members which they can chose whether or not to follow (as with all other advice given here) I don't see an issue so long as you aren't trying to sell your product here.

I think following on from your latest post I needed to make some changes. I have moved this thread to the "Vendor" area as when "you get down to it" your presence here is simply to "market your product". I have looked at your site and all the signs are that you are a new business that has a product which may or may not be useful to our membership (I'll let them decide) so with that in mind we will let you run for now and monitor feedback from the members. However any attempt to "direct market" the members via the PM system or any other means is strictly forbidden and will result in immediate removal of your account.
Apologies if my approach seems a little draconian but our members and their security is paramount (this forum reaches to all corners of the globe) and as you will be aware the "on-line" world is full of scammers and threats so we have to be diligent.

Lately I've been experimenting with AI on my second hobby, which is geneaology.  I asked Gemini some questions about one of my ancestors and got a totally fabricated story.  When I pointed out some errors in the AI response I got a reply that Gemini was hallucinating a response because it couldn't find anything but it wanted to please me.  Hmmmm.

I have a son who works in AI in SF.  We spoke recently and he remarked about how quickly AI is improving to a point where he no longer needs to employ new engineers to assist him.  He is now beginning to wonder how long his own job will be needed.