1/ At #Neeva, design is the act of giving form to an idea: we gather data and inspiration, think, make, and iterate through feedback. 💡
Here's how our team, working alongside the ✨Neeva Community✨, shaped our latest news tool, #BiasBuster...
(read on 📖)
2/ To improve the news experience on Neeva, we solicited insights from users of various news outlets.
One early finding 👉 the journey to get daily news typically started from news providers' sites and apps, but NOT from a search engine.
🤔
3/ So we asked ourselves, when does a search engine becomes necessary and helpful in the journey? 💭
Several users shared that they searched for specific events and stories about which they wanted to learn more.
An avid news user put, "Search is for focused topics.".
4/ The users visited their typical news providers' services to get daily news. But where did they go to get more points of view, especially for political news?
Some said their news sources were sufficient. ✔️
Some went out to check opposing views from the sources they knew. 👀
5/ Checking POVs by visiting multiple outlets takes a LOT of effort.
But do you know who’s capable of making this process effortless? Neeva! 🦸
We pursued the idea. 🧠
6/ Limited sources and freshness were common complaints about news on search engines. Quick retrieval of POVs from broad & fresh sources was also crucial.
The critical aspects of our backend and frontend to rally around were:
✨Broad✨
✨Fresh✨
✨Fast✨
✨Easy✨
7/ Going into detail design development, we generated ideas to express that this tool was about finding and viewing news articles in the spectrum.
Here are ideas we came up with ⬇️
8/ We struggled to find ways to signal and invite the user to drag the control around to set the desired position. 😑
Our caring Neeva Makers community unblocked us with detailed feedback.
Thank you Makers for helping us all along throughout the process! 💓
9/ We investigated UI motions to visually explain news source changes. Also, the motion exercise using the @Framer Motion, an animation library, answered our pathway for UI implementation.
10/ We chose the seesaw figure with color gradients.
We rotated the half-circle within a rectangular container as a mask because it was:
✅ Sufficient enough to express the POV selection
✅ Practical to implement
We prototyped in @principleapp + gathered usability issues.
11/ When an end-to-end build was ready, 80+ opted-in users shared candid critiques.
This sped up the backend improvements & UI polishes:
📌 Wider handle
📌 Clarity on what range is selected
📌 Hover tilt in a randomized direction
📌 Handle dot color changes for a bit of fun
12/ After many meetings, @clickup tickets, and a lot of math, we refined other details, like snapping the thumb to the nearest point on the spectrum and rubber-banding on the edges.
And we shipped in the US just in time for the midterm elections in November. 🙌
13/ What's next? Users shared:
🗣️ Bias Buster can help them relate better to their friends' perspectives
🗣️ Neeva should build an AI summary of opinions
So, the fun of making a better news experience will continue... 😄
We're excited to share we added #NeevaAI:
✅ Answer support for verified health sites, official programming sites, blogs, etc.
✅ Availability in the News tab
🧵
First, at @Neeva we're passionate about generative search engines combining the best of search & AI.
But it's clear generative AI systems have no notion of sources or authority.
Their content is based on their reading of source material, which is often a copy of the entire Web.
On the other hand, search engines care about authority very intimately.
#PageRank (the algorithm that got @Google going) was committed to a better authority signal to score pages, based on the citations they got from other high scoring pages.
Have you seen ChatGPT combine info on multiple entities into an answer that’s completely WRONG? 😬
Generative AI and LLM models can mix up names or concepts & confidently regurgitate frankenanswers.
Neeva is solving this problem on our AI-powered search engine.
Here’s how 🧵
FYI This is a two-part thread series.
Today, with the help of @rahilbathwal, we’ll explain why the problems happen technically.
Tomorrow, we’ll talk through how we’re implementing our solution with our AI/ML team.
Make sure you're following... 👀
In frankenanswers, a generative AI model combines information about multiple possible entities into an answer that’s wrong.
Ex) On this query for `imran ahmed’ from our early test builds, you see a mix up of many intents corresponding to different entities with the same name.👇
2/ First off, we found that there are far fewer resources available for optimizing encoder-decoder models (when compared to encoder models like BERT and decoder models like GPT).
We hope this thread will fill in the void and serve as a good resource. 📂
3/ We started with a flan-T5-large model and tuned it on our dataset. We picked the large variant because we found it to generate better summaries with fewer hallucinations and fluency issues.
The problem? The latency is too high for a search product.