Trace

Inspiration

I kept noticing how easy it is to ask an AI something like "is this worth buying?" or "is this actually true?" and just trust whatever it says. A lot of the time you don't really know where the answer came from, and it can be hard to tell if something was made up. I wanted to make something that showed the evidence instead of just giving you an answer.

That's where Trace came from. The idea was to make research feel more transparent by showing the sources behind the answer, how confident the result is, and where different sources disagree.

What it does

You can ask Trace a question about something you're researching. For example, you could ask if a laptop is worth its price or if a claim is actually true.

Trace searches trusted sources and looks at what each source is saying. It then organizes that information into claims and shows which sources support each one. You can open a claim to see the sources behind it, compare sources that disagree, and see a decision matrix when you're comparing different options.

The main thing I wanted was for every claim to have actual evidence behind it instead of just being something the AI came up with.

Challenges I ran into

One of the biggest problems I had was making the research reliable. At first, I tried having one AI model handle both the web searching and the generation. It would sometimes fail with request size errors, even when the question wasn't that big.

I ended up separating the search and analysis into different steps. This also made it easier to search for opposing opinions instead of only looking for sources that agreed with the first results.

Another problem I thought about was what would happen if the main search service stopped working or ran out of credits while someone was using the app. I didn't want the entire research process to fail because of one service, so I added multiple fallback options. If the main search doesn't work, Trace tries another search method, and then has a basic Wikipedia search as a last resort.

Accomplishments I'm proud of

I'm most proud of getting the research pipeline to handle failures without completely breaking. Having multiple search methods and fallbacks means Trace can still return useful research even when one of the services fails. That was something I didn't really think about at first, but it ended up making the project feel much more complete and reliable.

I also like how the evidence map turned out. Instead of just giving someone a final answer, Trace makes it easy to actually see where the answer came from and compare the evidence for yourself.

What I learned

I learned that APIs that work perfectly in a simple demo can behave very differently when you're actually building a full application around them. Things can fail for reasons that aren't obvious, so having backups and handling errors properly is really important.

I also learned that small details can make an app feel a lot better. Showing real progress instead of pretending something is happening made Trace feel much more honest and useful.

What's next for Trace

I'd like to add more types of sources, improve the rate limiting so it is based more on actual usage, and let users save their research so they can come back to it later instead of losing it when they leave the page.

Built With

• framermotion
• groq
• groq-compound-mini
• nextjs
• tailwindcss
• tavily
• typescript
• wikipedia-search