Can you elaborate on what you mean by “it’s almost assumed..”? I’m actively working on datasheet parsing at scale at Zenode.ai, and I haven’t read anything specifically about this…
Thank you! Yeah, missing pricing and availability is a tough one, but I'm really hoping we'll be able to get it from the distributors in return to funneling better traffic towards their sites...
We've actually heard this from a few industries, there's definitely a lot of different vertical AI opportunities here! Right now we're very focused on just doing electronic components, but maybe someday!
Yeah, this was actually a nightmare . We had to create "canonicals", which ended up being a mix of regex, machine learning and good old manual human labor. Fortunately, we could usually rely on the datasheet as the 'source of truth', but even that approach failed when faced with the overwhelming amount of part data we were shifting through.
From all of our customer interviews, almost everyone used Digikey and the occasional manufacturer website to find parts.
Because of that, we're seeing many people put the parameters they're looking for directly into the search. After a few searches, it starts to be clear that you can add more information directly into the search, so queries get a little more complex.
We're getting ready to add features to the AI to help you (interactively) navigate the range of available parts to narrow down your choices and understand what's available when you don't know where to look or what parameters to use. You should be able to ask 'I need to convert a parallel RGB display bus into MIPI' and the AI will show you different families of parts that could do that.
ChatGPT has definitely spoiled us all, but we're working to improve performance every day. Thus far we've usually prioritized accuracy over speed since there isn't much room for error in this industry, but hopefully our current response time is still faster than reading everything yourself. The sources not linking is a bug, should have it fixed today/tomorrow!
Digikey's parametrics were the best in industry prior to ours, but they were still really painful to use because of the strings. But to even get to the parametrics, you first have to know in advance the category and subcategory names of the part you're looking for. There are over 3000 different categories, so even once you're pretty good, you're always asking yourself "is what i'm looking for in the logic category or the amplifier category". With Zenode, all this is handled using natural language! And we did a huge overhaul on the parametrics, we turned the strings into structured values and built modern filters that are just waaay easier to use.
Octopart is a bit of a different beast from us. They're crushing the pricing and availability data, which is actually hard to get (a LOT of business development to get access to the realtime values from the distributors themselves). Plus they have a really massive catalog, definitely broader than ours, but they don't have the depth (i.e. the detailed spec and documentation data) that we do because we went to the manufacturer websites to obtain that.
We're learning a lot here, since this always happened AFTER we handed over the BOM. It certainly seems like a pain, particularly when a supplier ends up being a critical bottleneck (always late, overcharging, etc). The number one requested feature from these folks is a way to validate an alternate component, something they can use for themselves BEFORE passing it over to us design engineers for approval (heard some pretty spicy takes on doing this, apparently working with EE's isn't always sunshine and rainbows...)
Yeah, Moat's are tricky in the age of AI. We think there's always room for a standalone interface in this niche sector, given the sheer amount of data that needs to get presented (with no room for hallucinations). But beyond the data curation (which is actually quite hard) and the custom UI, there's also all the tuning data to optimize for the workflows that come next, like Alternates, schematic awareness, etc.
But yes, OpenAI continues to be the 800lb gorilla 'boogeyman' crushing startups with every new update. Our goal is to remain cockroaches until we find solid distribution!
Thanks, that's why we built it! We're big digikey fans, and have spent hundreds of hours on the site, but that doesn't mean it's without room for improvement!
Solid questions, you've clearly been burned before lol.
1) Document revisions can be a bit hard to handle, but we currently look for any dates or document numbers present, and when possible we just show the latest. However, plenty of times there are no dates OR revision numbers, and we'll get different documents from multiple sources. In this case we show all the documents, and have prompted the AI to flag this risk to the user if the question is answered differently between documents. Not the ideal approach, but working with what we're getting...
2) Hmm, not really. You can share the chat, and the specific response, but the only way to export the source would be to screenshot it... Adding it to our list of potential next features!
Awesome, that was one of the top reasons we built it, tired of clicking on strings lol. What are you working on specifically? Would love to chat and see what we can do to help!
This is actually well said. One of the issues we had to overcome with Zenode was the fact that people reference the same information in completely different ways. For example, "what's the nominal power draw" and "what's the typical power output" are the same query, but to an LLM, these are different.
We had to do a ton of work to get the tool to consistently return the same value (and still sometimes it shits the bed).
So yeah, even light technical jargon can confuse the heck out of an LLM, which is why that wasn't the only tool that's necessary to build this.
Try us out, would be curious to get your feedback! www.zenode.ai
Damn, yeah, we added the ability for the AI to set filters based on the input over the weekend and it has definitely slowed things down. We've got a rust script thats going out today to speed things up and add more functionality (like giving you the count of parts that each value in a filter is associated with), that should speed things up pretty dramatically....
Thanks for the try though, we'll definitely be improving this dramatically over the next few weeks!