How Sparkbot works
A Sparkbot is an assistant trained on one kind of page and on what you care about. It knows the questions to answer, how much each one matters, and the right answers it has to match. Training takes a short conversation and happens once. After that, a Sparkbot runs in about a third of a second.
Training a Sparkbot
You train it on a real page, so every decision is checked against something you can see. The careful work happens once, during training, which is why running a trained Sparkbot stays quick.
Tell your Sparkbot what you want it to answer
Open the panel on any page and ask in your own words, like "Is this a good deal?"
Training splits your request into specific questions and suggests the factors that usually decide broad ones, with the ones this page covers already selected. None of the page is kept once training is done.
Check it and adjust it
About 15 seconds later you see real answers on that page. Keep what matters, drop the rest, and adjust your Sparkbot by giving it more context.
Each factor becomes a detail to pull from the page or a check to answer, with an importance you can change. Edits re-run instantly, and marking an answer as wrong records the right one so the Sparkbot is adjusted until it gets it right.
Save it and reuse it
From then on it runs by itself on every page like this one, in under half a second. Share it with your team or publish it to the Gallery.
It learns where it applies from the pages you run it on. You can see those sites in the panel and remove any of them. Sharing passes on its instructions, never your pages.
What training means for your data
You're teaching a Sparkbot what matters to you, not handing over your content. Here is exactly what is kept and what is shared.
Training teaches it what you care about
Training a Sparkbot means telling it which questions matter to you, how much each one counts, and what a right answer looks like. The result is a set of instructions. It isn't a model built from your content.
Your pages and data stay private
Sparkbot doesn't keep the pages, documents or data you use while creating a Sparkbot, and nothing you browse is used to train anything, for you or for anyone else.
Sharing shares the instructions only
When you share or publish a Sparkbot, you share its questions, factors and weights, which is what it should care about in that situation. None of your pages or data go with it.
The parts you can change
Every Sparkbot is made of the same five parts. You can open any of them and change it, and your edits apply the next time it runs.
- Details
Facts read off the page, like a price, a date or a phone number. Numbers can count toward the score, such as lower mileage being better.
Price, mileage, seller phone
- Checks
Questions it answers about the page. Each one knows whether an answer is good news, bad news or just information.
"Does the listing report an accident?" where a yes is bad news and critical
- Score
An optional 0 to 100 score with a verdict, built from the details and checks you care about most.
A good deal score of 78, worth a closer look
- Known answers
The right answers you confirmed while training. They act as the Sparkbot's tests, and it can only be published to the Gallery when it matches every one. The pages themselves aren't kept.
Title status, marked as not stated
- Sites
Where the Sparkbot applies, learned from the pages you run it on, plus PDFs.
Listing pages on a car marketplace
What happens when it runs
A trained Sparkbot answers in under half a second on a typical page, and its score shows on the toolbar badge moments after the page loads.
It reads the page you're on
Web pages and PDFs both work. Menus, ads and other clutter around the content are ignored.
It answers every question
Every answer points to the line on the page it came from, and anything the page doesn't say is marked as not stated.
It scores the page your way
Each item counts by how much you said it matters, and you get a score with a verdict you can open to see why.
How the score works
A score should be easy to trust. You can see what went into it, and it won't give you a number it can't back up.
- You decide how much each item matters, and you can change it whenever you like.
- Serious problems count for more than minor ones, so a repaired scratch won't sink a listing the way major damage does.
- One deal-breaker is enough to keep a page from scoring well.
- If the page doesn't mention something, it's left out of the score. If the page is missing too much, you won't get a score at all.
- You can open any score to see exactly what went into it.
Where it runs, and what it sends
Suggested when you open the panel. On a page none of your Sparkbots know, opening the panel sends only the page's title and address. You get the Sparkbot that fits, the others that might, or an offer to train a new one. Nothing is sent as you browse, and pages are read to answer your questions, never to train anything.
Learned when you run it. Running a Sparkbot on a site teaches it that site, for you. The next page there runs by itself. Learned sites are listed in the panel, and you can remove them.
Private until you share it. Share a Sparkbot with specific people who can use it or edit it. They get its instructions, never your pages or data. Publishing to the Gallery requires it to match every answer it was trained on. Anyone can copy a Sparkbot they can see and train the copy on their own priorities. Making a Sparkbot private again removes every share.
Your next page, answered in under half a second.
Add Sparkbot to Chrome, open a page you check every day, and train your first Sparkbot on it. Free for 15 pages a month.