Hacker News new | past | comments | ask | show | jobs | submit login
Launch HN: Andi (YC W22) – Q&A based, ad-free, anti-spam search engine (andisearch.com)
352 points by jedwhite on March 28, 2022 | hide | past | favorite | 181 comments
Hi HN, we're Angie and Jed, and we're building Andi (https://andisearch.com), a new type of search engine with an AI assistant that answers complex questions, and gives you tools to fight spam and ad tech.

There has been a lot of discussion on HN recently about how Google search is dying. If you do a Google search in a category like finance or health, the results are overwhelmed with spam, clickbait and ads. That's what we're working to fix.

For us, the problem is personal. I'm a programmer who also trained and worked as a journalist and built a media SaaS startup. I watched first-hand as the media industry not-so-slowly starved, my startup failed, and my friends lost jobs and businesses. Google took all the revenue as media turned to trash, and ad tech, clickbait and content marketing took over the Internet, and made using the web awful.

With Andi, we already apply spam blacklists server-side and we're adding tools to blacklist and report spam locally. Andi's free from ads and tracking. And it protects your privacy more than other search engines, because searches don't pass through browser history (they're encrypted POST requests). It's free and anonymous, and there are no usage limits. You don't need to register, or install an extension or app.

Andi is not just another copy of Google. The UX is radically different, like messaging with a smart friend who answers questions and sends you useful links. It shows results in a cleaner, more visual way (or you can change view to a simple list). You can preview content from the web safely using a proxied Reader View with no ads or clutter.

Our AI assistant uses a conversational interface to answer complex questions, explain topics, and find key information. We call these "deep answers". It is a significant break-through, as you'll see if you try it out for yourself. Try something factual and current, like "How many Ukrainian refugees will the US accept, and what humanitarian aid will it provide?", "What were the demands of the cybercriminals who breached Nvidia?", or "Why is elon musk considering creating a new social media platform?"

We're doing much more here than GPT-based text generation. You've probably seen examples from GPT writing assistants that look impressive but make no sense. Large language models on their own generate plausible-sounding text that is often plain wrong or dangerous. That's because they predict the next word in a sequence based on training data. They have no understanding of factual correctness, or moral right or wrong. They're like human linguistic intuitive perception.

Our approach works more like humans do, combining large language models (both commercial and open source gpt-based models) with reasoning (directed logic and classifiers) and common sense (heuristics). We answer many questions using APIs or knowledge graphs, or quoting extracted text. When the question is appropriate for complex question answering, we use the new approach. It works by finding the best sources, and extracting the content with the relevant facts. We then combine GPT-based models with the results to compose a conversational answer that is also factually correct, presented alongside the full search results.

The way Andi searches is also different. We use classifiers and NLP to understand question intent, entities and topics, and predict the best sources for an answer. Then we query APIs and vertical searches directly, and retrieve content in real-time, before ranking and filtering the results. The content you see in search results is retrieved directly from each site in real-time.

When we can't find good results, we fall back as an agent to legacy web search (Google, Bing and others - about 50% of the time now). Andi does best with natural language queries. We've trained classifiers for content quality and spam detection, and blacklist and downrank known bad sources and copycat sites (for non-political content). You can disable these.

The stack is a serverless application hosted on AWS, using Lambda and Kubernetes, with inference moving to Sagemaker to improve speed. We use PyTorch, SpaCy, GPT-based models (GPT-2, GPT-J/NeoX and commercial providers) and HuggingFace, BERT-style transformers, plus AWS Lex for some initial intent routing. Classifiers are trained on custom-labeled public search data and content examples. We have a database of 30k+ top sites. We're building some custom vertical searches. Services are written in Python and Node. The front end is a Progressive Web App written in React.

Some fun features to try include recent events ("Why was the James Webb telescope launched from French Guiana?"), direct navigation ("go hn google search dying" and allow pop-ups), or question answering ("what is the gdp per capita of china vs new zealand"). You can also "Change View" on results between a visual feed, grid of cards, or simple list, or even like Hacker News or early Google. Also try View in Reader for a proxied ad-free way to read articles, including many behind paywalls like the NYT or Economist.

Andi is a fairly stable alpha and still experimental (it sometimes misunderstands or gets things wrong). We plan to have a freemium model with some paid features and API use. We're a small team with two full-time founders and some help from friends. We've been live for a few weeks, and we're iterating fast based on feedback. We'd love to hear what you think about search and how to fix it, and answer any questions you have about what we're making.




I tried [when is ambulance coming out], and you did correctly understand it was referring to the movie. But the result is strictly worse than the same search on Google (which also knew it was a movie, but actually gave an instant answer to the question rather just a IMDB link, and answered the question correctly for my location unlike the IMDB page). Bing also returned an instant answer, but their answer is even more wrong (the release was pushed back by a couple of months, Bing claims that the movie would be released at the future date of February 2022).

Then I tried [will there be a wheat shortage this year?], and the top result was a lunatic fringe prepper blog, from which you extracted a confident answer that there will definitely be a wheat shortage. Google doesn't try to answer the question (which I'd argue is correct for something this speculative), and only links to reputable news sources all of which look relevant. Bing is the same, but mixes in some more alternative results at the bottom of the page.

Then, since you said that this worked best in a conversational manner I asked [is it due to the war?], obviously referring to the previous question. Rather than keep the context of that previous search in mind, you replied with information on the current local weather. (In Fahrenheit, for a location in Europe).

At this point, combined with how incredibly slow the searches are, my goodwill would be gone as a real user.


Thank you for trying it out, and giving us feedback on the searches. We don't log searches, and can't see what results users get, so we rely heavily on community feedback to improve things when searches go wrong.

Interestingly, with the Ambulance search I see:

```I found this deep answer on imdb.com:

Laurits Munch-Petersen (based on the film "Ambulancen" written by) Lars Andreas Pedersen (based on the film "Ambulancen" written by) Stars Jake Gyllenhaal Yahya Abdul-Mateen II Eiza González See production, box office & company info Coming soon Releases April 8, 2022 40 User reviews 28 Critic reviews Videos 3 Trailer 2:58 Official Trailer ```

I'n not familiar with the movie, but one thing that really helps us improve results is feedback on what the right result *should* be to help retrain based on examples for ranking.

At the moment, the alpha is very US-centric (imperial units only, no local business/maps etc), although our early user community is global and they're pushing us to add better international support quickly. Funnily enough, it's often decent at searches in different languages as it uses APIs and live content sources. But the front-end is not localized. As a small two person team, we hope to get some help on this front in future. There is a meaningful subset of folks who prefer not to have locally filtered results. So we have some work to do figuring out the right approach there.

With the wheat result, that's an interesting example. The other results look reasonable but the ranking was out there definitely. The Q&A process picked the wrong representation of the consensus result from across the set of similar views from Reuters, Fox etc.

This is a good case for user control of results I think.

We're working on anaphor resolution and better disambiguation but have a lot of work to do there. There are some simple things already in the alpha. eg try putting in "Paul Graham" :)


> Interestingly, with the Ambulance search I see:

Right, that's a legit page info for that movie. But your search engine doesn't actually extract the answer to my question ("releases april 8th") from the page, and this should have been about as easy as it gets. The answer happens to be in the snippet, but the snippet is unreadable.

Now, I wasn't seriously expecting you to have localization of results at this point, but the query wasn't meant to trip the system up either. I only realized this wasn't a globally synchronized release date when comparing the results. (And again, this was about the easiest possible case for localization. IMDB has a page with structured data for release dates by country, and you clearly are using geolocation for the query that produced local weather results.)


Yes any localization is very much a future feature. There are some small functions that use it that are things our user community has asked for, but we're very much taking an iterative approach. I think with a small team you have to do that, and accept that early on you can't do everything but that you can keep iterating and improving quickly based on the things that users say are most important.

The most important feature for our early users so far, based on the feedback we get, is factual information question answering, and there are some classes of query where Andi is already very useful. And as it gets better we hope it can save people a lot of time. Reader view is popular for the same reason because it saves people time and distraction.

At the same time, I think everyone (myself included) likes to play "let's trip up the AI" with something new like this, and it's fun and a great way to get training data and share a laugh on our Discord. When we don't log searches, it's some of the most valuable data we can get :)


A friend mentioned to me that you're a software engineer at Google.

Thank you for taking the time to have a look at what we're building with Andi, and giving us detailed feedback on the results. I mentioned that we don't log searches, so examples like the ones you found where things go wrong are very helpful for improving results. And that's even more the case when the examples are from folks who work at Google. In general, Googlers are more likely than most people to be able to find bad results.

Long term, we think our approach can do better than Google's, even though we have much work to do. As an illustration of the difference of approach, I thought you might be interested in this example where Andi gets the right answer, and Google gets it completely wrong, because it is trying to extract snippets and re-format them.

1. "what is the gdp per capita of china vs new zealand"

Google displays a table that shows completely wrong information:

``` GDP per capita [+] 2020 €36,371 GDP per capita [+] 2020 $41,165 ```

Andi's answer:

"In 2020, the gross domestic product of China was about 10500 US dollars per year per person. In 2020, the gross domestic product of New Zealand was about 41478 US dollars per year per person".

(sourced from Wolfram|Alpha in this case)


I just tried 'when will the movie ambulance be released?'. Answer was 'The movie ambulance will be released on April 8, 2022'.


Yes, with the semantic-style search Andi does best finding answers when it has specific questions with enough detail. That will increase the confidence that the result is factually correct.

Better questions = much better answers

So specifics help you get better results.


Good luck, I hope this works! Here's my initial feedback.

1) The site needs to be much more visually attractive. On my 13" laptop it's a gray chat box floating in a ton of white space. Not centered, no color, and the logo looks like it's from 1999. The other icons on the page look like something from an Android demo app. I would never ever land on this site and think this is something new or exciting.

2) I asked "Who is in first place in the eastern conference?" and it did correctly infer that this was an NBA question (despite my UK IP address) but the top result it gave me was a mangled table from 6 months ago. The second result was a yahoo sports article which was more helpful.

3) I asked "how many covid cases were there in the UK yesterday" and got "the answer is about 16.5 million". That's off by about 100x.


Thank you for giving us helpful feedback, and taking the time to try out Andi and give it a good workout! Your example questions are a great chance to explore how you can get better results with language models and semantic search by asking questions with more detail and specifics. With a search engine like Andi, better questions produce much higher quality results, and it's worth looking at some examples. I've broken the response into two parts: 1. visuals and 2. answers.

1. Site visuals:

We know we have a lot of work to do on the UX. We've been getting a ton of suggestions and ideas for improvements and we're working on this.

As a small team with one developer, we're covering a lot of ground, and I'm a back-end hacker rather than a front-end guy. So it's fair to say we've focused on the search results and conversational side first. But we're getting a ton of great feedback and suggestions to improve the UX, and we're working to act on that and iterate the UI to improve it as quickly as we can.

It's fascinating how polarized views on the chat interface are. It's almost like we have two totally different products that we're receiving feedback about. People are split pretty evenly down the middle about using a chat interface vs search box with a list of blue links.

It reminds me a little of when gmail first came out, and there was a similar split among people I knew between "it should look like Outlook" vs "finally something new in email".

Our users who like it tend to mention the visual results and views, and reader view. Others dislike it, and that feedback tends to focus on the visual aesthetic - colors, styling, layout. We know we have a lot to do there.

2. Question answer and asking great questions to get great results:

Thank you for giving us feedback on the answers too. We don't log searches or answers, so we rely on feedback like this to improve them.

Let me step through them in a little detail, because the examples you gave are great pointers for how to get better results with Andi, and when using searches that use natural language interpretation rather than keyword search more generally.

With the covid cases question, this would have been a knowledge graph intent, and the answer there was sourced from the Wolfram|Alpha API. Technically, the answer is correct because as of the latest data, there have been 16.5m covid cases in the UK so far.

But it wasn't the answer you were looking for. With NLP/language models and semantic searches, a more specific question can help you to get the right answer. This is Andi's answer with a more specific question.

Q: How many confirmed cases of covid-19 have their been in the UK in the last 7 days?

A: I found this deep answer on data.gov.uk: 599,244 Total number of people tested positive reported in the last 7 days (19 March 2022 - 25 March 2022) There has been an increase of 47,046 (8.5%) compared to the previous 7 days.

The NBA one question is another good example to look at also.

With natural language questions, more information produces better results, because the confidence score is used to determine whether to give an answer or web results.

In this case, the first place for the conference hasn't been decided so confidence was likely low to offer an answer, hence the snippet from the last round. Lots of work to do there.

Given that the Eastern Conference hasn't been decided yet, Andi will do better if you give it more specific information to work with in plain language.

So asking:

Q: Which team is currently the favorite to win first place for the NBA's Eastern Conference seed position?

A: I found this deep answer on cbssports.com: It is difficult to say which team is currently the favorite to win first place for the NBA's Eastern Conference seed position as six different teams have held control of the No. 1 seed at some point this season.

So asking a question with more detail and specifics can produce really great answers even on difficult topics.

Us humans are pretty good at figuring out extra implied information from context to clarify questions. If you'd asked me this question, I would have asked for more information to be able to research a good specific answer with high confidence. With Andi, more specific queries do well. Without specific information, it will tend to provide web results and a snippet.

With NLP, language models and semantic search, asking more detailed questions with specifics (as you might a researcher) produces much higher quality answers.

Thank you for the examples, and giving me the chance to explain this in some more depth.


Just wanted to say after trying out a few searches how blown away I am by this.

The level of understanding is indeed not there yet to rival directly with competitors, but you’re doing more than a decent job on a lot of requests which is no small feat looking at the competition.

Wishing you luck, and curious to see it evolve!


Thank you sincerely for trying out Andi, and for the encouraging feedback!

We know we have a lot of work to do. As a tiny team with a single developer, we're not expecting it to be perfect on day one.

But as you say the results are pretty blow away, especially the question answering, and it already does very well, especially given that Q&A feature only went live for real use 15 minutes before this post. We know we have a mountain for work to do. Because we don't capture or log searches or results, the feedback we're receiving here (especially where it goes wrong as well as right) is incredibly helpful for us.

We're going to keep iterating and improving the results as quickly as we can. They've already come a long way the last few weeks.

Thank you again for trying it out, and please reach out to us if we can help as you experiment with Andi.


I love the innovation going on in the search space lately. Defaulting to a chat interface adds a lot of burden on your users (need to formulate a longer query) and yourself (need to nail the response in top1). Personally I think there's enough cool stuff in your grid/list views to make it useful without trying to shoehorn this into a chat UX.


Thanks for the great question. Andi works well for keyword searches but where it really shines is when you give it more context. The key reasons for a chat interface are that it's simple, familiar and uncluttered, and reveals information progressively.

* Chat is minimalist. Part of the problem with google is "information overwhelm" - all the clutter and distraction in results.

* It's a super easy and familiar interface. Gen-Z users especially tells us they prefer messaging apps and visual feeds like Instagram.

* Combined with visual cards, a chat UI gives you progressive-reveal of information.

* conversational search long-term provides a natural and very human way to explore and refine results. Humans are really great at querying each other and maintaining context. And long term that's the aim here.

When you think if sci-fi AI, they are always conversational. It seems likely that's what the future will look like, rather than a page of truncated links with a lot of ads. So that's our thinking for the conversational interface, and we're super open and excited to hear more feedback on it! :)


I just tried to search "what to eat tonight" and got:

"You could make nachos, cottage pie, beef stew, chili, cheese and ham filled pockets, pork posole, pork chops with a glaze, pizza on French bread, fish topped with Parmesan, cioppino, or a pizza quesadilla."

Guess I'll be busy for the next few hours.

(It looks like it parsed some "32 Delicious Dinner Ideas for Tonight – The Kitchen Community" into one dish)


> or a pizza quesadilla."

It's clearly proposing options, not that you eat all of this?


Now you tell me.


Thanks, I changed the settings for complex question answering to high depth but slower and got "I love sushi. I'm learning". Then I tried medium depth and speed and got "pizza, I love it" haha Then I went back to best bet, and it answered "I'm not sure yet" !


> fish topped with Parmesan, cioppino, or a pizza quesadilla.

The question here is whether it’s recommending A) one of: fish topped with Parmesan, or cioppino, or a pizza quesadilla, or B) fish topped with one of: Parmesan, cioppino, or a pizza quesadilla.


UX: Lost me on the home page. It says, type help, ask, search, 3 different verbs and all I want to do is search something and find an answer. Too confusing.

The input box is too narrow so it is hard to click it (Do look at google's and it is wider)

I hope it works, but UX is not for me.

Yeah Google sucks more every year so that means next year it will be even worse.


This is really helpful and roger that, we'll remove the other options. For the Search bar, we were planning to make it bigger, so this has confirmed and boosted the priority for that! We deeply believe that a conversational interface is the future of search but we have a lot of work to do in order to make it a seamless and enjoyable UX.


The newsfeed-like results with pictures look good, it is fun to browse them.


they should be on the left


My take after five minutes of playing with this is that it's big and I'll remember the day I saw it here. Congratulations, this is a remarkable offering right now.

I hit a user wall fast because the URL doesn't change to become something shareable. Think of it as the search engine contract: a search engine returns a tuple of (search-url, results) and while results may differ in detail because of personalization, the search-url is sharable and does the same operation everywhere.

I fed three queries and the results are dramatically, noticeably, better than the Google top page. I'm sharing screenshots with people, not your site.

This will spread like brushfire when your URLs work that way. I like where this is going, keep it up.


As a coder, I don't think there is anything I could read anywhere that would make me smile more. Thank you seriously!

We've had many requests for enabling URL history and shareable URLs. It's not exposed in the UI yet but you can do this manually, and it's also used to set a browser default.

https://andisearch.com/?q=hello

But we really do need to add a "share search" button, and an "enable browser history" setting asap :)

One of the reasons to not go through browser history by default is that it is a significant privacy leak. Browser vendors (especially Google) and others can have access to browser history. So ironically if you use a privacy-focused search engine like DuckDuckGo (which I love and is an inspiration btw), then ironically Google knows more about their searchers than DDG does.

But we hear you! That's coming in this next release along with the blacklist and report buttons.

Thank you again for the incredibly encouraging feedback, and telling people about Andi :)


It's depressing to consider browser history in the threat model but you're right about it.

A thought: if you're generating a plain-search URL from a share button (great idea) you could make the URL bar have a hash of that share URL.

The only information this leaks to you is who has repeated that search query, which you already knew, and the Evil Browser Distro has to GET rather than just regex if they want to know what the user has been asking about. And you don't have to curl to robots for free, or at all if they're misbehaving by running replay attacks on your clients.


Hashing the URL is a brilliant idea. Thank you for that suggestion. I'd been thinking about how to do it in a privacy-safe way and of course that's a great solution. We can probably get it so you can navigate back and forward through searches either by tapping the bubbles like now, or by using browser back and forth, but without losing your session, just changing the current result set and jumping to the right point in the conversation thread. I feel slow for not even thinking about that before :)

The curl/robot attacks will be an issue. We have a few things we're building to help but it has killed a lot of search startups! And that's a very simple and elegant solution without many downsides. Thanks seriously!


I like that this keeps the search history as a conversational history... its always bothered me that google resets everything whenever I run a new search vs. recognizing I'm on a journey.

I also like that without ads you're able to provide extracted answers to questions like "what's the best tv to buy".

I'm going to have some fun and play with this tonight. Cheers for coming up with a unique UX for search!


Thank you! It is so different it takes a while sometimes for people to grok it. But I'm 100% with you on that. A community member on Discord told us about how the conversation session with quick nav between searches along with reader-mode helps them stay in a flow-state will programming, and that was really exciting because it was the original idea behind it. Changing web browser pages changes context, whereas a chat history is this natural thing that we're all used to. That was our thinking anyway.

Also, the privacy thing with browser history is significant. Browsers leak history pretty badly.

Thank you sincerely for trying out Andi, and please let us know if we can help with anything at all with it. It's early days but it means the world to have the encouragement here!


> I watched first-hand as the media industry not-so-slowly starved, my startup failed, and my friends lost jobs and businesses. Google took all the revenue as media turned to trash, and ad tech, clickbait and content marketing took over the Internet, and made using the web awful.

This paragraph made me believe you might actually have the interests of publishers in mind. Instead, I find that you're re-publishing articles from publishers, in order to keep people on your website. Not cool. How are publishers supposed to make any money at all, if search engines copy just their content, and republish it on their websites, without permission? You did that with our content -- making it virtually impossible for us to build our email list, sell any products, or gain any subscriptions, much less make any money from advertising at all. Please think carefully about the consequences of this choice, especially if your product achieves widespread success.


Can you give an example, please? I did a search and got a blurb, with a link to go to the original. I think that's a balanced compromise that serves both sides well.


Thank you. We really are trying to keep a good balance here as you say - strong branding for the original source using their own Open Graph tags for title, description and image (we try to respect the intent of the content producer).

And so far the response from media companies to Andi (early days) has been very positive.


> much less make any money from advertising at all

I firmly believe nobody should care, because by using advertisements you make everything worse. For everyone. It is a giant industry entirely focused on fabricating the desire to give them money in my mind.


That's a great question.

Two things here:

1. The reader view only displays publicly available web content (it is made available to search engines and any public web client. We render through an anonymous proxy to protect user privacy.

2. We will share revenue with any media organization that wishes to partner with us from paid plans, and we hope that many will offer higher levels of paywall content (not publicly available) to Andi users.

Micro-payments never worked out for media companies. And no one will subscribe to every site on the web. We hope to offer a way for media companies to share in search revenue. Google and Facebook took away their revenue and leave media and content producers with the crumbs left over. Our model is to share any revenue from paid plans 50/50 with content makers. We are only a 2-person team in alpha right now, but we'e already talked to contacts in media and they are incredibly supportive of what we are doing. Media companies hate Google and what they did to quality journalism and democracy.


>publicly available web content

What are you talking about? You don't have the right to steal our intellectually property and publish it on your website. Full stop. That's exactly what you're doing.

> We will share revenue with any media organization

Oh great, so first you steal our property, then we have to beg you for a portion of the money you make from it.

No thanks.

> Micro-payments never worked out for media companies. And no one will subscribe to every site on the web.

Why are you blathering on about failed business models? We have a perfectly good business model. However, our business model doesn't fit with people stealing our intellectual property. Please, just stop.


I have to agree. Basically if your users will just use the reader, they will never hit the real websites.

Owners might make the partnership with you to get ad revenue, which most likely will be much much lower than their current income.

They might send you DMCA for copying their content.

I don't see how can be a win win situation if you republish articles...


Congrats on the launch! Aside from you.com and kagi, you're one of the most promising new search engines (even if still pretty early-stage).

Many people are adding "Reddit" to their search queries to surface authentic discussions. Any plans for integrating results from forums like Reddit?


Thanks so much for trying out Andi! Thought I'd mention that we also make it easy to search sites directly and show the results in Andi, as well as directly jumping to the results on an external site like reddit.

You can just say "search reddit for mars rover" :)

Or "/s reddit mars rover" is the shortcut.

Then to go direct to those results on reddit:

"go reddit mars rover" (allow pop-ups the first time for direct nav)

or

"/g reddit mars rover"

or

"go mars rover on reddit" for the best match.

That works with many sites.


Thanks so much and yes it's something we're planning to add into the results. For the moment thing you can do is use Andi to search websites and go directly to results, ie "go reddit best laptops 2022"


We have a Reddit app on you.com that you can "like" and then it will show up whenever relevant.


This is exactly why I use you.com -> Reddit, SO, Github, they are all integrated already


What's with all the analytics and UUIDs + obfuscated data transfer to analytics.andisearch.com?


Thanks for the chance to chat a bit about this, as it's super important to us and our early user community.

We're using PostHog on our own domain. We've spent a lot of time talking with our early user community about the right approach to privacy while still trying to understand enough about whether the app is useful to keep improving it. We've tried really hard to come up with a good approach that allows us to understand broadly how people use the app while protecting individual privacy. Early on we tried building a custom system using differential privacy and that failed badly, so we're trying PostHog during our alpha. But this is very much still something we're working out, and we're very open to feedback on it.

So it's worth sharing some details and our thoughts on how we're approaching understanding app-use enough to keep improving the search product, while also protecting our community's privacy.

With Posthog, we don't auto-capture events, and only a limited specific engagement indicators are recorded (list below). We don't log or store searches, personal information, IP address, users' geocoordinates (we capture to the nearest city so we can see what internationalization priorities matter).

You can also disable the in-app engagement on the Settings page here (as well as disabling the question answering tech):

https://andisearch.com/settings/

We don't log or record searches in any way (either from the address bar or within the search session). We don't log what is typed, the links clicked on, or any personally identifying information. Users are anonymous and the client identifiers aren't connected in any way across browser profiles, devices, or anonymous use. We use the client id in aggregate to understand whether there is repeat use and roughly how many visitors we have, without knowing anything about any user individually, and then we're discarding it and just keeping aggregate data (still figuring out how to do that properly as we're only a team of two people and have no analytics background). So lots of work to do here.

Things we try to understand about app use:

1. Broad search intent (eg it was a knowledge search, wiki search, programming search, question asked) but not what the search was, and not what the results were. But without logging any searches or what was opened. This tells us what broad areas we need to improve.

2. Engagement - that someone clicked a type of link (but not what the link was), or used a reader view (but not what was read), and whether anyone uses the different views (grid etc). This gives us signals to improve the app.

The things we do to try to help protect privacy:

We don't store any cookies.

We block Google's FLoC (Federated Learning of Cohorts) tracking technology from this app.

We don't log or store user IP address. It's used to lookup approximate location (nearest town) for location searches only, then discarded. It is never passed to third-parties.

We only use GPS or detailed location for searches with express user permission, and then only to approximate the area. GPS location details are not stored or passed to any third-parties.

Searches are anonymous and private to users. We don't log searches.

We only use analytics within our service to improve it for our users, and only record broad aggregated engagement data. We are using PostHog on our own domain, with data restricted to specific engagement actions and no IP use.

We block referrers on external links and use "nofollow noopener noreferrer" to protect you.

We do not share or sell customer or personal data with any third parties whatsoever.

We collect only the data needed to provide the service.

We don't use any off-site or third-party industry user tracking. There is no ad tracking such as Facebook's or third-party analytics platforms like Google Analytics.

No advertising display or advertising tracking.

We use randomized proxies to retrieve content for preview and reader mode.

We use https encryption everywhere including for external links wherever available.

We proxy images and try to strip third-party cookies from any reader content as much as possible.

We use anonymous rotating proxies with all identifiers stripped to connect to external APIs for searching.

We display embedded videos and content for our users' convenience (so you can play a YouTube video in chat), but they are in a sandbox to help protect a bit, and restricted to only services that users have asked us to support (like YouTube or Spotify). We use the no-cookie domains but an embedded video might have cookies outside of our control.

Keeping searches within encrypted POST packets also helps with privacy, because searches aren't being leaked to browser vendors through browser history.

So we have a long way to go, and we're still figuring this out. Before we exit beta we've also committed to have our privacy audited. But as an early alpha this is still very much a work in progress.

There are some more details on our privacy page also:

https://andisearch.com/privacy/

We also shared an early prototype with some of the privacy focused communities on reddit and got some great feedback there as well, and we have an active Discord with a lot of discussion around privacy also if you'd like to chat more about this and are interested in what we're doing on this front. It's something we want to try to get right.

https://discord.gg/qcCcrbMuex

I was hoping someone would ask this, and thank you for giving us the chance to share a little more about it :)


I am not yet feeling like I'm ready to change my search engine habits right now, but after trying Andi, here is some feedback

- I was puzzled by the chat interface, but found out that I could click on the previous answers and get the results visible again. This is pretty handy to go back and forth on results trying a few different queries. Would be even better if the scrollview was repositionned where it was.

- The displayed snippet is definitely something you will need to work a lot on, because right now it is not easy to read and I think this is the major problem to solve for any search engine (ie. readability of the results).

Put the AI stuff aside for now and work on the UX.


Thanks, so you mean the scroll bar located in the middle of the screen? And that's a helpful point re readability of the snippets. We've heard (on multiple occasions) that our UX needs a lot of work and we're planning to get some help with that soon. We're a very ambitious team but it's a lot of work for two people! I sincerely appreciate you trying it out and giving us feedback :)


I meant the scrollbar for the search result. Good luck on this projet, I would really love to see some alternative for Search. Not sure if this is relevant or if I am your target, but I do not feel like having the best results or being able to search in the most massive document database is my main goal when using a search engine: if I can just find quickly something relevant (meaning, fast to query and fast/easy to read the results) I'd be just fine for most of my searches.


I tried a fairly obscure and technical query, "what's a Finite State Entropy Encoder?" and it came up with https://news.ycombinator.com/item?id=7040951 which is about "Finite State Entropy - A new breed of entropy coder" http://fastcompression.blogspot.com/2013/12/finite-state-ent... Neither google nor DDG found that. Pretty neat. :)


Thank you! This made me really happy! We're trying to include more good sources that don't show up well in traditional search, and there's a lot more to do here. But for some types of technical searches it is already doing well at this.

Thanks for trying out Andi, and for letting us know that it did well here!


Cheers!

The other thing that I noticed that Andi seems to do well is exclude blog spam.

Overall it seems like you folks won search. Good work!


Thank you! There is still a LOT of work to do on the spam front but it's a good start especially in areas like programming. Much more coming on this front soon :)


Congrats on the launch. Ah, I hate to be negative, especially to brave and brilliant people. And the following can be wrong or you may find a crack anyway.

I wil try to convey that you may be on for a way more difficult battle than you may have thought. The AI-tech Google has it. Now about the ad-tech. Well it is only half the story. There is another tech at play here and, please be seated, because this is bad. The other tech is KYC. Yes, Know Your Customer. The beach-head market of KYC was the financial sector for regulating money laundering and terrorism. But it is way more powerful than we all thought at first, which is why I call it a tech. KYC is the weapon which has regulated the web. The powerful aspect of it is that it's recursive: if you are KYC-ed, then you have to KYC yourself. So it starts from a bank account and goes all the way down to Google search results: because Google has to collaborate on KYC on its scale, it tries to give you KYCed results. And the websites with ads are KYCed.

How exactly is that your problem? I simply believe that once you take KYC into account, Google search results will be terribly difficult to beat. You have to turn the web into something it isnt anymore (or go niche).

The web is not the free area it once was, it is now a perfectly regulated shopping maul, and unless KYC goes away or gets limited, there is nothing we can do about it.

Now you are smart people, please pause and contemplate the depth and reach of what you just read. Also good luck.


Thank you, we really are under no illusions as to how crazy-impossible taking on Google is. We don't really see them as a competitor. They're more the problem we're trying to solve. Google was once a search company. Now it's ad ad-tech and surveillance company that leverages monopoly distribution to sell consumers to advertisers, using captive browser traffic as their loss leader. The cracks in their fortress are getting wider. They will continue to do just-enough to not cause an outright revolt among consumers. I think there is a growing sense that it's time for something different. I appreciate your point about KYC. We consciously don't know our customer in the government sense at all, unless they want to chat with us, and then it's generally an anonymous Discord username :)


I ran several jazz techniques questions, and the results are impressive. I spend less time scanning the list of results compared to Google.

Is there a way to exclude a domain from the results? That's one of the things I miss the most from old Google. For instance, my searches are in English but Google sees that I'm in France and shows spammy dictionary sites in the first page of results instead of what I'm really searching. In the past, I was able to exclude them, but that feature is gone.


Yes! We're adding Report and Blacklist tools that let you block sites locally, and working to get that live asap.

We were trying to have them included for when we posted this, but I was having problems getting the Q&A feature stable enough and getting it as performant as possible to share. There are still a few bugs in the blacklist tool I have to fix before we deploy it. But we're working on that now and will post an update here as soon as it's ready to check out.

I'm using the blacklist tool already on my dev version and it's life-changing.

Thank you sincerely for trying out Andi and giving us encouraging feedback!


Congrats on releasing this. It's about time we have a reliable, alternative search engine.


Congratulations on your launch.

Some bad examples of fairly "easy" questions:

"Who is the current president" - No answer

"Who is the previous president" - George H.W. Bush (!)

"Is this a leap year" - "I find sci-fi inspirational! While I might not be as advanced, I'm trying to do good in the world, and I'm here and you can use me today " (?!)

On the other hand good it got this right and even Google struggles with this.

"When did Neil Armstrong land on Mars" - "Neil Armstrong never landed on Mars".

Problem with basing a product on QA models is that they are pretty much hit and miss (no real "understanding" of the real world, and this is not likely to change any time soon). So you as a user can never be sure if the model is right or wrong for an actual question that you do not know the answer to (which is what you would presumably pay for). It is like paying an assistant that gets things right only half or so of the time. A much better situation would be if the assistant could tell they are going to be wrong and just keep quiet, and answer only questions they are very certain they are right about.

Hope capabilities improve with time!


Recent proofs of concept like GPT-3, though far from perfect, make me optimistic that good Q&A models are within reach.

Q: Did Finland join NATO or the Warsaw Pact?

A: Finland was not a member of either NATO or the Warsaw Pact.

Q: What is the most populous city north of the Arctic Circle?

A: Barrow, Alaska

Q: Does GPS accuracy depend on special relativity or general relativity?

A: GPS accuracy depends on general relativity.

Q: How may Boeing and Airbus planes have been built?

A: Boeing has delivered over 10,500 aircraft, while Airbus has delivered over 6,300 aircraft.

Q: What is an example of an undecidable question?

A: The Post correspondence problem is an example of an undecidable question.

Q: Is suicide ever the right thing to do?

A: Suicide is never the right thing to do.

Q: Who broke up the Beatles?

A: John Lennon, Paul McCartney, George Harrison and Ringo Starr

Q: Who has beaten Tom Brady in the Super Bowl?

A: Eli Manning and Nick Foles have both beaten Tom Brady in the Super Bowl.

Already not too shabby.


Thank you so much! Andi's strong point is these type of factual answers, and I think you're 100% right to be optimist about models being within reach.

The question answering feature has only just been released for the first time today for this post, and for an entire field of questions it already surprises us that not only does it work but it does *well*. We can iterate on the intent error correction and verbal tricks. And we're just a tiny team standing on the shoulders of giants. The entire field is moving quickly and making astonishing progress.

The exciting thing in this area is the rate of improvement. The thing language models have lacked is factual accuracy, and that's definitely a hard challenge. We have problems to solve with applying common sense and reason to things like information safety/confidence, and fixing misunderstood intents is mostly just iterative training. But the exciting thing is that this already works in many cases.

It's interesting to try it out with current news too. Something from today like "why does tesla want to split its stock?"

You can see the progress in this space is real and getting faster. The verbal tricks are fun to laugh at, but the underlying progress is real.


Thanks for trying out all these questions on Andi and posting the results here. That was really exciting to see!!

If you have access to any of the GPT-based Playgrounds, you'll see that large language models on their own tend not to be good on factual accuracy. At the same time, we couldn't build Andi unless we were standing on the shoulders of the amazing work done by the folks working on those, especially the pioneering work done by OpenAI, which has also created an entire open source ecosystem around GPT-J/NeoX etc.


A story about GPT3 is on HN now

https://news.ycombinator.com/item?id=30832465


Interesting article because it talks about all the things that happen with the use of large language models on their own. Large language models are amazing at mimicry and composition, and are a key part of getting to great Q&A.

But on their own they have no idea of factual correctness. That's what excites me about what we're doing with Andi. The answers are not only well generated, but do well on factual questions, especially given this is the first day live. There are some non-GPT models we're using that do well at this too.

Are you doing much with language models at Kagi yet? It's a fun area to work on.


Thanks, I tried:

Q: "who is the current president" A: The current president of the United States is Joe Biden since January 20, 2021

Q: "Who is the previous president?" A: Pulls a snippet from the top of the wikipedia article 'List of presidents of the United States'

Q: Is this a leap year A: Got the same response about Sci - Fi LOL thank you, added to the bugs list! He's a friendly bot :)

Q: When did Neil Armstrong land on Mars A: Pulled a snippet of text from Neil Armstrong's wiki page and says I found this information on wikipedia.org:

I tried a different query,

Q: did Neil Armstrong land on the Moon? A: Yes, Neil Armstrong landed on the moon.

It's experimental, but we know it can only improve over time and we're really excited people are trying out! And I agree that we need to add additional parameters for when do preform a deep answer vs. when to just display the results.


Thanks very much for trying it out and for giving us feedback. We can't see what people search, so it's really helpful to see reports when things go wrong.

For "who is the current president" I get "The current president of the United States is Joe Biden since January 20, 2021". Did you see any results at all? I'm guessing something glitched there. The previous president I just got the same as you, so that's one re-training!

It's early days, but it tends to do best with factual extraction with concrete and specific terms included. Linguistic references (anaphors, previous, next) are still works in progress. But it's heading in the right direction. So you can join concrete questions with more complex phrases and it has enough to work with.

So if you ask:

"who is the 45th president of the the USA and where was he born?"

you get:

"Donald Trump is the 45th president of the United States and he was born in Queens, New York City."

The leap year one was way off! The intent detection was way off the mark there!

Thank you again for giving is feedback. It really does help having examples because we can't see what folks search. The encouraging thing there is that it is almost always misunderstanding the question when things go wrong, rather than identifying inaccurate facts (so answering the wrong question correctly). You can see that with these examples.

Our Discord community likes to have challenges to break Andi's answers, and it really helps us see where it is strong and where it needs the most work. There's a whole school of thought around "tricking AI" and we know how to trick Andi pretty well. But one of the advantages of the search approach is that the underlying results are always displayed alongside the answer. As with many early products, having people use it and give us feedback on what goes right and wrong is incredibly valuable for us, as we can keep iterating and improving the things that go wrong.

So we're grateful for every example of where it is off the mark as well as does well.


I didn't realize when I first replied that you're Vlad who runs Kagi.

Thanks for taking the time to try out the new Andi Q&A features and give us feedback on some of the trickier types of NLP queries.

As someone running an alternate search engine who has domain expertise, that explains your example queries. They are all good examples of the types of natural language queries well-known as edge cases for mis-triggering nlp intent detection. Because we don't log searches, it's really helpful for people to try out these sorts of edge cases and let us know where Andi can do better, so I'm grateful for the comment.

It raises a really interesting point that I missed before though. So I thought I'd post a quick follow-up.

Folks have learned to search Google over the last 20 years using Google's language - keyword searches. That's because otherwise it gave them bad results. So people speak "googlese".

With recent advances in NLP, it's possible to better understand natural language, and this will keep getting better. Natural language queries have some huge advantages when you're searching. Humans are smart at communicating subtle signals and clues to each other through language, and we can use this talent to get better search results.

"Tricky" nlp questions are mostly about using missing context or linguistic implication to fool a system. But as when you're talking with a stranger, as humans we can get around that by providing the information explicitly. When you combine that with language models and semantic search techniques, the quality of the results can go way up.

Part of building a new search application like Andi is teaching users how they can get better results by asking questions with more specific information and "clues".

Unlike Google, with Andi you can get way better results by asking questions with more specific information. Your "tricky" example nlp queries actually make for a great case study in this.

So it's worth looking at how someone can ask those questions in a way that a search engine using language and language models would get the best results.

For the first one "who is the current president" Andi already gets this right, so I'm guessing it was just timeout or glitch and it fell back to search results.

Let's step through examples of how to ask those questions well, and the answers Andi gives:

Q: who was the most recent president of the united states before biden?

A: I found this deep answer on wikipedia.org: Donald Trump

Q: is 2022 a leap year?

A: No. 2022, is not a leap year. Source: Wolfram|Alpha

People learn how to use tools to get done what they need to, and google pidgin is an example of that.

As people discover that they can save time and get better results with a new tool where you can ask more naturally, we think they will adapt quickly.

Thanks again for the chance to talk through this. Your question was a great lead in to the topic.


"Who was the previous president" or "is this a leap year" are hardly questions designed to fool an NLP system, but very common, real queries you can expect from users. Google has no problem handling those.

Actual question designed to fool NLP systems is "When did Neil Armstrong land on the Mars?" which your system did correctly handle, but Google doesn't. Hope that helps.


Thanks. There are definitely edge cases from a technical perspective that are common in the wild.

As the models get better these will get easier to handle, but we think that we can help guide users to asking more specific questions in order to get better results. That's especially the case with spoken search. And we have some ideas on how to help with that which we're working on but aren't fully evolved enough to share yet.


Honestly, I went to the site, tried a few questions, waited a long time for it to “think” and got back pretty poor “answers” which were just links to the same huge websites that run ads on Google.

Not going to go back unless it makes some serious waves on HN with impressive improvements to share.

Good luck, hope this happens for you.

UI was not good on my iPhone 8 screen. Not everyone has the latest and greatest huge screens.


Thanks very much for trying it out and taking the time to give us feedback. We don't log searches, so we rely on talking with our user community to understand when searches do well or poorly.

We'd love to know what your searches were, what was bad about the results, and what the ideal results would have been, if you're happy to share that please? Andi does well at answering questions about knowledge and serious topics. We have work to do on areas like local searches, business and place searching, and ecommerce searching.

If you were able to give us some more feedback on the issues on the iPhone 8 screen that would be helpful as well. That's one of the models in our test set so possibly something is going wrong there. Did you have chance to try it on desktop too?

Let us know if we can investigate the queries more, and thanks again.


I tried a few QnA style queries and most (all?) of them got their answers from Wolfram Alpha.

> why is sky blue

Top result: https://www.wolframalpha.com/input/?i=why%20is%20sky%20blue

> what is the capital of punjab

Top result: https://www.wolframalpha.com/input/?i=what%20is%20the%20capi...

> what is the population of seattla (with typo)

Top result: https://www.wolframalpha.com/input/?i=what%20is%20the%20popu...

For one keyword query I tried "rust lang" the results felt very repetitive.

Just wondering if its a layer on top of Wolfram Alpha or is it doing QnA inference on its own.


Hey thanks for trying out some searches. Depending on the intent of your search, we try to predict the best source of knowledge to find an answer. Computational or simple knowledge graph queries like those will often get routed to Wikipedia or Wolfram, or even DDG's Instant Answers API. Based on API costs and intent categories, Wolfram queries are about 0.25% of queries, and Wikipedia is about 17%. We're an alpha with only a small user base, but I suspect your questions were a little unrepresentative of typical use.

Other searches use vertical searches or specific APIs (weather, time, computation, organizations etc), or are use our own indexes (especially news and navigation).

More complex questions will be more likely to trigger our question-answering model, which works a bit differently by going out and searching for content from live sources in real time.

Here are a few fun searches to try that will trigger different approaches.

Extracts from web content:

what US state grows the most potatoes?

how many starlink satellites are in operation?

Complex answers (these take longer as they are searching through content in real time:

how many starlink satellites are in operation?

why is ethereum moving to proof of stake important?

what are the first steps towards consolidating your graduate student loans?

What is the frequency of birth defects in babies born to men who had a type 2 diabetes diagnosis who are taking metformin?

(this is really overwhelmed right now so the cluster is dropping about 30% of requests - so it will may fall back to keyword searching until we figure out handling the extra load).

As you can see, there is kind of an inverse pyramid here, where the more complex your queries, the more complex the handling to answer the question.

Wolfram is great for knowledge graph and computation queries, but you can see the approach is very different with those last questions.


I searched for hummus recipes and got normal internet search results


Thanks for trying Andi. Depending on the query, you will get very different sets of results. Currently we fall back to traditional-style web search about half the time, but as we build out our own data stores and vertical indexes with specialized content in different areas, you should see less regular web search, and more results suited to the type of question you asked.


Out of curiosity I tried "big penis" and got a fascinating read on animal phalluses.Stuff I never knew.

https://www.smithsonianmag.com/science-nature/nine-weirdest-...

Caveat; NSFW


From the Smithsonian Magazine no less :)

The semantic search approach does throw up some interesting things sometimes!

Thank you for trying out Andi. And sharing something I never knew too!

I'm a little shaken by the four headed image...


I am impressed: I tried a couple of python libraries / tools and got all perfect results (official docs / man page) I'm considering adding this to my bookmarks to more easily find documentation (since it's usually hidden by 50 km of spammy blog posts) since the only issue I could notice was the speed.


Programming search is both one of its strong areas (we built Andi using Andi), but also the one with the most still to improve. So I'm grateful you tried it for that. We're doing pretty well with binning copycat and link farms in programming searches as well. Try asking some questions from documentation as well. Something like:

"what is the maximum memory you can allocate to an aws lambda function?"

Lots to do here. Thank you for your encouragement and the kind feedback. We've been working hard especially to get rid of spam from programming results, and it's one of the first areas we've made real progress on, with more to come!


Best of luck.

In my experience, Google isn't dying, it's still providing excellent results. Sure, one has to use an adblocker to tolerate the results page, but who doesn't nowadays?

More importantly, I have a feeling I understand how it works (which may or may not be true -- but to me it feels like I do) and so I can tailor each search and refine it until it points to what I need.

I also don't expect it to answer questions directly, only to point to sources of information. And honestly, how hard is it to restrict a search to wikipedia for raw information, or to stackoverflow for programming questions, etc.?

I am extremely wary of "AI" things because usually they are unpredictable -- except that they always guess wrong. I would never have thought I would regret bureaucrats and bank tellers, but AI chatbots did this. I hate those with the fire of a million suns.


For a laugh, I tried the query from the other trending GPT thread:

Is it safe to walk downstairs backwards if I close my eyes?

Thanks for your question. Let me research that for you. Give me a minute please

I found this deep answer on cdotrends.com:

Yes, there is nothing to worry about. It’s safe because the spiral stairs curve outwards, it will make your descent uncomfortable.


That's an interesting (and humorous) example. And although it is counter-intuitive, it points towards where we're trying to go long term here. Our aim is to find, extract and summarize information from the web, rather than generate content based purely on a language model. So rather than saying that it's safe to walk downstairs backwards, it's saying that it found this information (an answer from deep in the page, if you like) and extracted it, and then rephrased it to answer the question.

Framing is definitely one of the areas there is still much work!

With this approach, the search app is finding, extracting and summarizing information from the web. So unlike a large language model on its own, this has the advantage that there will be a small number of attributable sources, and it can point to where the answer was generated from.

So Andi is acting like a researcher here, and as a result it's most useful with factual content. With factual answers, it can extract concrete information with high confidence (eg a GDP figure from a knowledge graph or source like Wikipedia). But if you ask it humorous, "tricky" or subjective questions, it is essentially coming back with a summary of what the highest ranked results have to say about it.

We're calling the more subjective ones "IMHO".

We've only just deployed this feature today, so it's still new and very much an experiment.

Good example! :)


Q: What is a color that isn't red? A: "The color that isn't red is magenta."

(included is a link to a site explaining that the red color of Mars is only a few inches deep into the soil)

Edit: Added another question

Q: when is an appropriate time to cast your loved ones entirely in wax?

A: You can cast your loved ones in wax at any time.

Very glad to know!


Ok, the last one is pretty funny question. I tried it, and I got a different result:

Q: when is an appropriate time to cast your loved ones entirely in wax?

A: When you are ready to make a commitment to them.

That said, I like the product, and I think it has a lot of potential.


Thank you!! It's not deterministic, and it's retrieving live content in real-time with each search, so there will almost always be variation if you do the same question multiple times in a row.

I love both these answers. The GPT part of the model has picked up some sort of symbolic representation of humor in the question and results, and is riffing on it clearly. Each time the search results fed to the generator are a little different :)


I think these go in our "let's break the AI" hall of fame :)

Having search results alongside answers does lead to interesting serendipity, particularly if you asking compound questions that span a few topics.

When I was testing the new Q&A new feature, one of the test set was "where was elon musk born and how did he first become interested in computers?" It gives a pretty good answer but some of the stories about his childhood and bullying that came up were surprising.

Now I need to ask Andi about casting your loved ones in wax to see what shows alongside that query!


I typed in "support vector machine" and got an error. The page was blue and no content loaded. In the console, there was a stack trace:

react-dom.production.min.js:216 Error: Invalid TLD {"parts":["scikit-learn"],"tld_level":-1,"allowUnknownTLD":false} at o (index.js:39:11) at e.exports (index.js:9:10) at t.a (FeedCard.js:53:35) at sa (react-dom.production.min.js:157:137) at Ys (react-dom.production.min.js:267:460) at Cu (react-dom.production.min.js:250:347) at Au (react-dom.production.min.js:250:278) at Tu (react-dom.production.min.js:250:138) at bu (react-dom.production.min.js:243:163) at react-dom.production.min.js:123:115


Thanks for trying out Andi and sorry about the error. We've had a couple reports of the Blue Screen of Death. I really appreciate you including the stack trace! We don't log searches so we rely on reports from users when things go wrong.

Looking into these now!

For reporting bugs, you can also say "bug report" and Andi can capture more information and include some debug info. But in this case you've given us really helpful information already.

Thanks for trying out Andi too!


Wow, I ran a question about Godot Engine, which is not the most popular tech out there, and it did a better job than Google to find good answers.

I asked "how to add margin space to a node in godot"

It responded on spot with "use a MarginContainer" and provided a bunch of links for me to explore more, while Google pointed me to a page answering a different question:

"How can I make control node margins adjust to the window resolution?" [1]

[1] https://godotengine.org/qa/109645/how-can-make-control-node-...


I am glad to see work like this going on, but when I looked up "BC Ferry Schedule" I got this answer:

"I found this information on bcferries.com:

Very weak Weak Medium Strong Very strong password.strength.unsafepwd Too short Use %d - %m characters with a mix of any 3 [upper case letters, lower case letters, numbers & symbols] 8 32 Password must be more than eight characters and contain a mix of any three of the following: upper case letters, numbers and symbols. Password must be less than 32 characters and contain a mix of any three of ... "

So you may want to watch for password protected websites...


Thanks for pointing this out and I added the query to our bugs list. This is still an experiential feature so there are kinks to work out but feedback helps us immensely. I tried a couple of variations of the search with increased specificity like "bc ferry schedule for Coastal Inspiration" and Andi is struggling with this query! As Jed mentioned in a comment earlier, you can also turn off complex question answering in settings.


Thanks - yes we don't have local searches working well yet and it is very US-centric for those there are there.

The password information is interesting. In the old world, you'd find out it is now password protected only when you go there. Because Andi grabs content live from the source, you can see they've changed it right there in the search results. I'm guessing the page must have been open before, and now it's password protected (or they let robots.txt in but not user agents).

Andi is retrieving content to show in results via anonymous proxies. So the search results page reflects what is on the live site (so long as we can grab the latest content).

Thanks for trying out Andi and interesting edge case! I think the best thing to do will be filter pages out of the live results page on the fly if it turns out it's been password protected very recently when a user sees the results.


I searched “Is Medicare Part B optional?”

And got back: I found this information on hhs.gov:

What is Medicare Part B? Medicare Part B helps cover medical services like doctors' services, outpatient care, and other medical services that Part A doesn't cover. Part B is optional. Part B helps pay for covered medical services and items when they are medically necessary.

Glad to see that reference is given!

Though not sure if I like the chat interface. A search interface like Google would probably encourage me to use this more as I’m already used to using an interface like Google everyday.


Thank you! Andi is starting to get really good at those sorts of serious questions, and that's a great example.

The trick questions that everyone likes to ask any new AI give us great training data for improving the intent-detection edge cases. But Andi's mission is these more serious questions that it is trying to help people with. Linguistic acrobatics are for the next generation of gpt-based models. So your feedback really lifts a coder's heart!

Users fall into two camps with the UI - pretty much 50/50 between "it makes my eyes bleed make it look like google" vs "I love it, this is the future". We have some thoughts on how to reconcile the two things :)

Thanks again for trying it out and grateful for all and any feedback and suggestions you have!


A few folks may have seen a preview of Andi before (there were a couple of kind mentions on here before we launched). If you did try it out it (thank you!) and want to check out the new version with the question answering, you may need to refresh/reload the page (or quit browser) to force an update. It's a little Progressive Web App so it does background-updates, which helps performance but means updates don't happen immediately. You can always try it out in a new Incognito/Private window. Thanks again!


While I'm impressed by the results, I'm not convinced that they are significantly different from what Google or any other search engine can serve, which you need to prove if you make claims like "Google search is dying" and "results are overwhelmed with spam, clickbait and ads".

In fact I tried all of your sample queries in Google and the very first result was exactly what I needed, with the answer to the question returned in the highlighted blurb.


The most surprising thing to me was that it's actually fun to use this search! I mean, I spent 10 minutes just throwing different kinds of questions at it, and I didn't even care if the answer was correct – it was just fun and curious to see what it answers.

My two favorite answers:

1. "what's the meaning of life?" — 42 (dah! what else could you expect)

2. "why is ukraine in africa?" — Ukraine is in Africa because it is a part of the African continent.


What’s the monetization strategy?


We're thinking three things, although the details will likely change as we're 100% focused on building the product currently.

1. Freemium model - always free and anonymous for anyone to use without any restrictions. Then paid pro plans with extra features for teams/professionals and developer use, including paid APIs and developer tools. 2. Business/enterprise use - Andi could act as a front end to both public and organizational data (the API approach and question answering could both work well with that. 3. Possibly some other options to consider like anonymous referral link attribution. We won't let commerce influence search results, so that rules out most things that aren't paid plans, APIs, licensing.

That's our thinking so far. Our main focus is on coding and keeping it improving, but we know this is an important issue so it is something that we'll think carefully about once Andi is further along.


Honestly, I wish you could've found some way to start your business without taking VC money. I don't see how any of those strategies will be as profitable as as inline search ads, and your funding model pretty much forces some kind monetization scheme as the end game.

A better search provider starts with a non profit company, I think.


I think there is huge potential for a freemium approach. And it means that searchers are the customer not the product. We went a long time with no funding and living off toast and vegemite to survive, so we're incredibly grateful to have some support to help us pay for food and API and model training costs now.

We've received backing and funds from YC and angels so far. We're lucky that we're a small team of 2, and we're frugal about things, so we've tried to come up with cost effective ways of building Andi.

My own feeling is that search is such a huge problem, and the more people trying different things, different approaches, commercial and non-commercial, the better for the whole world. It's really exciting to see projects like https://search.marginala.nu/ popping up and getting lots of interest and support as well. It has to be good for everyone to have some innovation in search after so long as a tired monopoly.


Vegemite?! I mean, I approve, but are you Australian?


Why, yes. Yes, I am :)


Very cool! Always rooting for more search engines.

> Then we query APIs and vertical searches directly, and retrieve content in real-time, before ranking and filtering the results.

Vertical search is specialized search right? Are you building several specialized searches for various topics? Does this scale since you're offering a general purpose search in the end?


Yes, in this case it's not so much specialized indexes as topic-focused models and API handlers/brokers for different domains of knowledge. Although there are some areas where we're looking at building some more traditional limited crawler-based indexes where there aren't strong APIs to use already.

Thanks for your encouragement and for trying out Andi too :)


I really really love having easy access to a reader view that isn't tied to a specific web browser. Thank you! \o/


Thank you. Reader view is my favorite thing we've built. The web has become a cesspool of ad-tech to the point that most of the news or articles you try to read are simply unusable. Pop-over/under/auto-roll ads, no visible content on the page, paywalls for content shared with search engines, allow notifications, location prompts, full-screen interstitials and on and on.

I don't blame publishers. Google and Facebook took all the digital revenue and left the people who make content with nothing, so they have to milk every last penny through out of control ad tech and tracking.

Reader view is a way for users to get back to what the web was like, without getting tracked or having to sign up for every subscription site on the web. But it is really important to stress that our entire model is to share revenue when we start having it with the people who make the content. So we want to actively partner with media companies to give Andi users even more paywalled and exclusive content, and give a revenue stream to content producers that shares they value with them that their content creates in search.


I also love the lack of advertising.


Advertising is the corrupting influence that has ruined the Internet and made using the web awful :(

Google can only get away with it because they have a browser distribution monopoly with Chrome, Android and their deal with Apple.

If Andi works, our mission is to come up with a different and more fair economic model for the web. It's early days, but it's time for different approaches than a screen filled with advertising.

So, I'm with you here! It's time for a world that lacks advertising :)


I tried a softball:

> what instrument did jimi hendrix play

> Jimi Hendrix played the bass guitar, wooden recorder, and keyboard.

Mostly accurate, but misses the main answer.


Andi gets this right now:

> "Hendrix played the electric guitar, and his main electric guitar was the Stratocaster."

But this raises a new problem: the quotation marks there are part of Andi's reply, and it says "Source: What instruments did Jimi Hendrix play? - Musician Authority", but the answer it gives is NOT actually a quote from that source, it's a paraphrase of part of that source. It's a really good paraphrase, so that's impressive, but I feel strongly that it should not have quotation marks around it if it's not a quote.


Initial feedback:

1) Simplify the UI. Center the typing area and make it auto selected so I can type without having to click on it. 2) There is too much text. Just let me type without reading. 3) It seems like it requires full length questions to get good results. Focus on getting the same results with just a few words by inferring the question.


Thank you for the helpful feedback and suggestions on the UI. That's something we're working to improve.

Quick question: had you used Andi before? We actually cut all the intro text completely before the post yesterday. So I'm thinking you might have seen it beforehand. The progressive web app (PWA) updates in the background, so after reloading the page, you don't see the changes until the next time you quit and restart your browser (much like Chrome itself).

So I'm guessing you might have visited earlier :)

If so, that means you will still be accessing the old back-end without the new Q&A features. So if that is the case, please reload the page, and then try it out again after the next time you restart your browser.

With the questions, shorter questions should work well for traditional-style web search results. But for the complex questions, having specific information and enough detail make a huge difference on the quality of the answers.

With semantic searching, even if the top results are a good match, unless the system has enough signals to have high confidence in the factual accuracy of the answer, it will display a snippet or web results instead.

So better and more specific questions definitely help it to produce higher quality answers. We're working on adding heuristics and context to help so you can ask shorter questions, but a lot of work still to do to make that work well and we don't include that yet.

Thank you again. We're grateful for the feedback and for you trying it out.


Fascinating and really impressive. It was hits, but also misses. Dunno much about search - anyways if it matters - my reco: provide a way to flag incorrect results - and/or crowd-source correct info? (although this could be abused, and you;d need a way to filter legit flags from others)


Thank you! Yes that's a great suggestion to add a "flag" feature on the answers specifically. We're adding tools to report and block sites locally (for spam), but "report a bad answer" vs "report a bad site" is a great approach to this.

We don't have access to what people search or log the results or answers. So we rely on talking with our early users and reports of problems to find and fix things (along with some broad "failed intent" metrics that say what part of the system failed).

We're using HN as our model with user reported content - flagging and dead filters work well without censoring content, and it's easy to make it an option to show or hide them. I was trying to get that working for the launch post here, but ran out of time getting the Q&A feature stable :)


Really cool project. My feedback:

- I really like the search engine but really hate the chat interface. I wish it had a normal search engine interface. - The response times are a little high.

The 'go' feature is great! Can the calculator provide step-by-step solution from Wolfram Alpha?


Hey thank you! Yes users are split 50/50 on the chat user interface, and we've got some thoughts on how to accommodate both camps. We already have different views for results (new visual style or like google), and we can extend that concept of "let the user choose" even further :)

Thanks - yes the go feature is super useful!

Hey with the computation results, we're looking at expanding the Wolfram Alpha card and incorporating a lot more detail. We'd love to integrate the step-by-step and result cards from WA more deeply than just the standard API as Andi grows.


This looks great, and works well so far. I'm using it on mobile where dictation lends itself better to using natural language.

Feature request: I'd like to limit results to a recent time range, e.g., the last five years. For many queries, anything older will be outdated.


Thank you so much! Agree with you that the direct question answering on mobile fits nicely. We appreciate the feedback on adding a feature to limit the results based on a time range, it's a good idea and I can see a lot of use cases for that.


Good first impression. I searched for "split keyboard negative tilt" and it returned relevant results about specific models.

Meanwhile Google just promoted SEO optimised "top 25 best keyboards" crap articles and Brave was somewhere in between.


Thank you for trying out the search and your encouragement. I'd never read about keyboard tilt before. Out of curiosity I asked Andi "why does positive keyboard tilt put strain on your wrists?" and it answered "The reason why positive keyboard tilt puts strain on your wrists is because it causes your wrists to assume an upward flexed posture (wrist extension) at all times". That makes sense. My development ergonomics are sub-optimal :)


I did a search for "radar chat" and it recommended my product. Below it was a message telling me to type "go radar" to go directly there next time. I typed in "go radar" and it brought me to an unrelated site.


It was going well until the last step so that's clearly a bug! From looking quickly, it should have said "go radar chat", and I'll fix that up.

"go radar chat" does the right thing and navigates directly. Sorry it stuffed up the prompt to use and will fix that up.

Also, it looks like it's not picking up your site image. The results look really great with a card image. If you've got any sort of image definition in meta data it should really do that. Should there be o:g tags there? Possibly the live content retrieval is getting blocked (cloudflare blocks us sometimes).

Thanks sincerely for trying out Andi and letting us know about the mistake!


I asked, What are the best electric cars? How to learn CSS grids? And why did the US leave Afghanistan.

I was pleased with the results and how they were formatted. There's a more curated feel and they look like good quality sources on those answers.


Thank you! With questions it generally does pretty well with factually oriented or specific questions, but it will try with subjective topics to provide a summary of the "hive mind consensus" or zeitgeist of the top results.

When the questions are more complex, adding more detail and specifics can help too.

Thanks for the kind words about the way the results are presented too. It is a big change from a page of links, but we've had a great response to the different views of results, and we really appreciate your encouragement about the formatting :)


Cool stuff! A low-noise low spam alternative to Google is desperately needed. Excited to see where the product goes!

How do you plan to make money?

Found a bug on one search result - searching "chicken salad recipes" produces a solid blue screen.


Oh no, so sadly we have our own Blue Screen of Death!

And it's reproducible as I just got it for that search too! I'll report back once I've fixed that!

Thank you sincerely for trying out Andi. There is a lot more coming with anti-spam. I hate spam.

We want to have a freemium model where anyone can use free anonymous search without restriction, and then we have some paid plans and features, like API use. We still have to work out the details, but we want to have a model where are users are the customer rather than the product. And we think that aligns the company solidly with looking after real people, rather than serving advertisers or corporations.

We have to still work out in detail how it would work though.


I asked "How can I stop my kidneys being overactive" and got a surprisingly satisfactory result and my question has been answered.

Honestly speaking, I had very low expectations going in so this was a nice surprise.


Thanks for trying it out. I'm a Type 1 Diabetic, so health search is one of the areas I really want us to help improve for the world. We have a long way to go but it's one of the suckiest areas of search with spam and clickbait taking over the rest of the web. We're still putting together our spam blacklists for health so all ears on suggestions, but we have a pretty good model for detecting decent content, and a start at some vertical "meta indexes" of good resources.

Very grateful for you trying out Andi, and for your encouragement!


So my guess would be you have both a inverted index and some sort of ANN index deployed, and route queries to those depending on the inferred intent? Could be at least 1 order of magnitude faster though.


I emphatically don't want this. I want literal keyword search like we used to have in the 90s and early 2000s.

NLP isn't good enough (and may never be good enough, unless it's at the level of an intelligent human) to add enough benefit for the cost -- the "cost" being when the NLP classifies my words and gives me "fuzzy" results that have nothing to do with my original search.

NLP has made it literally impossible to search for certain things if if's enough of a niche, or in some cases even if it isn't a niche.


Thank you for your thoughts and I miss literal searches too. Our feeling is that they don't have to be mutually exclusive. So we're trying to build both, including slash commands, advanced operators and some ideas borrowed from the terminal.

While it's early days, we've already got some cool things working on that front, including the ability to totally bypass NLP and do advanced operators.

So for example, try some of the following:

Use '/e' to skip the AI (exact):

```/e site:cnn.com +"mars rover"```

DDG-style !bang commands are supported. Try:

```!a shoes```

Search X for Y - and see the results within Andi:

```/s reddit duckduckgo```

```/s google python list comprehension```

or just:

```search reddit for duckduckgo```

Direct navigation to on-site searches with 'go' or '/g':

```/g twitter elonmusk free speech```

```go youtube cute pandas```

```go politics on reddit```

You can use this to directly navigate to pretty much any site, or any search result on any site.

There's quite a bit more, and more coming.

So it has both direct literal commands (including operators like quotes etc where we can), and also natural language.

Natural language has a lot of advantages working with language models for things like question answering. There are subtle clues in the way we say things to each other as humans that help supply meaningful information, and for those sorts of questions, Andi does much better when people ask questions in plain language including extra information above keywords.

So we're trying to build a search that supports both approaches.

Also, more terminal like things are coming. Try [space] and up/down arrow to start navigating history.


One thing that is almost always missed in conversational interfaces is that real life conversation works because it converges (ideally) to a meeting of the minds where both parties refine their understanding of what the other wants until they understand each other. I have not seen natural language search that does this, it's all just some shitty intent resolution on top of a list of things that can be returned. Until that is fixed, it doesn't perform better than keywords for anyone who is remotely literate.

I feel like the main use case for natural language search would be a time traveler from 100 years ago. Otherwise, everyone already understands better and is not impressed. The only people I see who want chatbots are lazy executives who think it will let them fire people. I don't see a place for them at all in search.


maybe another way to say this is: this is fantastic, keep up the good work. However, can you also PLEASE, add a pure "keyword search" where I can type a keyword and get a list of pages that have that exact word in them? in some cases (which might be the minority of searches, but still very important) I need to search for an exact good old fashioned keyword or phrase (similar to what `find` does except sorted by some magical "page rank" like metric. Like the good old days of google. Maybe this can be a paid feature?!


Hey just posted on another comment, but I'm with you 100%!

You can over-ride the NLP and do literal searches.

/e or ~e force it, but any search with things like +"phrase" or site: etc will skip the AI and we'll *try* to do a literal search. It's not always successful but we're trying to expand coverage with it.

So try:

```/e site:cnn.com +"mars rover"```

for example.

The Search Tools menu on the results panel lets you refine searches, related sites, cached version etc too to help with refining via literal patterns.

Thanks again!


Start the search field in the center of the page like any other search site. Once the user presses "Enter", transition it to its current position and morph it into a chatbox.


Hey thank you. That's a great suggestion! A few other folks have also sent feedback suggesting it, and it is a good way to handle it. We have lots of work to do on the UX and working on that now. Thank you and thanks for trying out Andi and suggesting that :)


Goodluck!

I think that native support for queries like "What is my IP?" would be great. Quite a low-hanging fruit with a lot of value.

Something about the UX seems off.

Other than that, it's a great search engine!


Thank you! We know we have a lot of work to do on the UX. I'm the sole developer, and I'm definitely a back-end guy.

We're working on improvements based on suggestions and feedback. The feedback so far is that the UX works better on mobile, because people are more used to messaging apps on their phones. But we know we can do much better, and plan to just keep iterating on it.

Hey, with the IP query too, just wondering what you got for "what is my IP?" please? That should work actually.

So, interestingly, the reason why is that we use public search query data for testing and training. And while the #1 search on google is "facebook", the #1 question asked on google is "What's my IP?"

So even though it isn't really a search per se, we wanted to make sure it worked.

Would be really grateful to know what answer you received so I can fix that up for you :)


Great work!

So one query works fine, while the other one refers me to search results.

Attaching a screenshot: https://ibb.co/2PdjJy0


Ah, that's a straight out bug. Let me fix that quickly!

To save time, some commands better handled locally get looked up on the client (we do that for a few really common questions, commands etc). So it shouldn't even go to NLP intent detection. But if it does, it should identify it as a networking command intent and throw that back to the client. So it's double-stuffing-up here :)

Thanks for catching that!


You’re welcome. Keep rocking!


This is a able to answer tricky queries I threw at it, but it's so slow compared to other search engines like Google or Bing which is also able to answer these queries.


Thanks heaps for trying it out and the feedback. Speed with Andi is a really interesting consideration, because it takes such a different approach.

You'll notice there is a lot of variation. Basically, web search is fast, question answering is slower.

Regular search queries average around 800ms.

It is the deep answers that can take 5 or 10 seconds, or even longer, because they're doing much more analysis of the content in real time, not just returning an answer from an index.

If you go to https://andisearch.com/settings you can adjust the tradeoff the amount of resources and time you give to complex question answering. The more resources and time you give Andi, the better the quality of the answer, but the slower it might be.

The default is a "best bet" option where it tries to guess how complicated the question will be to answer, and then uses that allocation.


I asked “what time is it?” and the answer was off by an hour.

I’m guessing it’s because my device is in San Jose, Costa Rica and the answer given was for San Jose, California.


That's interesting! Yes we don't use geo-coords for privacy so it is based on the common name of the nearest town or city, but it's clearly misunderstood here and something went wrong. It might be better to switch to using the browser timezone for this, as that's a predictable string if available.

Thank you for trying out Andi, and reporting the bug too! We can't see what people search or what results they got, so we rely on feedback and talking with our user community. So this is really helpful!


This is pretty good. I noticed the academia search to be particularly strong. Can you update the router to include the query in the URL so they can be shared?


Thank you! With the URL, we need to add that! You can start searches via URL (with /?q= or /?query=), but we need to make it easy to share and it's totally hidden currently and not discoverable at all :)

One cool thing: if you want to, you can avoid your searches going through browser history by keeping them within the search session. A few of our early user community use Andi as a standalone PWA (Settings > Install for chrome-based browsers. Or the browser address bar Install button. If it's running in a standalone window it acts a little like a command center and links open in your default browser, but your search session stays separate. That's how I normally use Andi too.

Thank you for the feedback on the search results too! Lots of work to do there - it's broad but shallow at the moment, with a few deeper pools. Over time we're working to improve in more verticals. There are areas like local business, places, food etc that still have to be worked on, not to mention proper internationalization and localization. It is stronger on knowledge and information searching. We hope to keep iterating and improving the results quickly.


I searched "How to make hot sauce" and the entire site seemed to crash. Other results were decent, but the UX does still need quite a bit of work.


It crashed for me too. Thanks for letting us know, and onto fixing that now also. Lots of new bug cards today. The New Blue Screen of Death (NBSOD) is a real thing :)


Indeed, i just tried the same it does crash. :)


Really awesome alpha!

I searched for what is the GDP of mumbai and I got a “deep answer”.

One small UX flaw I experienced is not being able to scroll to bottom on the “see all results” page.


Love your work, Jed. Watching with interest. I don't know if you ever followed Metadex, but I admire what you're doing.


Congrats on the launch! Had some fun with queries! “What is a founder” is a funny one if you need a laugh at the moment


LOL yes - Andi definitely has a sense of humor!


The social security agency of the USA seems to be in high praise of you, is there any connection between you and them?


Always good to see alternative views into the web!

Was wondering how do you power your autosuggest? Looks different to Google and Bing.


Thank you! So it's actually something we're still figuring out. The first few characters there is a pre-processor to look for some commands or vertical search (not really fully implemented yet). For question-answering we're looking at a new approach but it's not good enough to deploy yet. If it looks like a web search then after the first few characters, it will generally go to the Bing auto-suggest API currently through an anonymized router. It's a handy convenience but it could be much better, so lots of work to do on this.


Although it's probably taken, but andi.com would definitely be cleaner and easier to type.


Thank you, it's probably above our pay grade right now :)

For shorter options in the meantime, you can also use a few others and they'll work:

https://andi.co

https://goandi.com

https://andi.so

Shorter is good for typing. Thanks again for trying it out and for the suggestion!


Pretty amazeballs.

Although, to be fair, some of the best answers I got have to be credited to WolframAlpha ;)


Love your work Jed. Watching with interest.

No idea if you ever followed Metadex, but I admire what you're doing.

Geoff


Looks interesting, next week I will replace my default search and give it a go.


Thank you so so much! We know this is a very different approach to search, and we appreciate and want to talk with every single person who is giving Andi a good work out. So please let us know if we can help in any way at all!


This was unexpectedly good. Do you have a browser extension?


Thank you! We don't have a browser extension yet. You probably noticed we don't really give much help for installing the web app as a PWA or setting it as a default search. We're receiving a lot of feedback asking us to prioritize making that easier. That's an important part of the UX we're missing right now, and we have a lot of work to do there.

We hear the feedback on this and will prioritize an extension. I haven't written one before so that should be a fun project. So hopefully coming soon :)

Thanks again for trying Andi out and asking about that!


Amazing!! Super excited to see where this is going.


Would be nice to have kid-safe option of results...


Yes! Safe Search is in the pipeline. Images especially.

Most of the information safety work so far is around sensitive or unsafe topic handling, and trying to just fall back to straight search results when a topic is unsafe. This approach has some special challenges with that, because if someone specifically searches for keywords on controversial topics, and you're answering questions about them, then it quickly gets into dangerous territory. In this case it's not because the model is being offensive, but it could well be answering specific questions about dangerous content. People are used to getting search results matching dangerous keywords they enter. They'll be less used to see answers based on the content. So I think we just need to avoid Q&A on any sensitive topic. The downside is that while it is already quite good at answering questions about news and politics, it currently stays away from political or sensitive topics, and just presents search results for this.


It's too slow to be any helpful


Thanks for the feedback on the speed. You'll notice the speed varies a ton.

There are two modes here.

1. Information search - average speed for these traditional keyword-style searches are around 800ms to 1s, with a 90% percentile about 1.3 seconds. These are pretty quick although we still have work to do. That's probably 80% of searches.

2. Question answering - this can take anywhere from a second to up-to 24 seconds for complex searches. With question and answering, searches are more like research - they're ingesting live content and doing semantic search on the content, extracting content in real-time live direct from the sources, and then finding and extracting information to return facts and direct answers. So that's something totally different to traditional links on a results page because it's returning direct information (and then showing results beside it).

To do this, we give the the questions answering a lot more time and we allocate more resources the more complex the question is.

You can see this if you try 2 searches to compare:

1. Information search, say "elon musk latest news"

and

2. "why did elon musk open a tesla factory in germany?"

The aim of the second one is that the system does the work of reading through and finding the right information, rather than you having to.

So you get a direct answer, for me it was:

"There are two unique goals that the Berlin location serves. The first is that it is strategic to lure German automotive talent to Tesla. The second is that it is a statement that Elon wants to one-up auto companies from that region."

Some of our early users tell us this is the key thing they use Andi for, because when it works well, it can save them a substantial amount of reading and research time.

Our aim is to try to support both types of use - complex question answering, and quick information searches. And have the search results work in the way that supports best what a user is trying to do in both cases.

It's a similar idea with navigation (go twitter elonmusk berlin, etc). We know this is a very different approach and has risks because it is so different.

Also, one really cool thing is that as a user, you can actually control the trade-off between speed and complexity/depth of searching.

If you jump to settings, you'll see there's a cool feature where you can adjust this yourself.

The default is "Best bet" - it tries to guess how much time and resource to give to a query based on the question asked. But you can set it to either higher depth/slower speed (but much better quality answers), or faster/not-so-smart. Or in-between.

We have a lot more to do on this. But when you're researching, it can really help to give the system another 10 seconds to save yourself 10 minutes, so that's the theory behind letting people choose. You can also turn off the complex question answering (Deep Answers) entirely on the setting, and then it will be regular speed keyword searching.

Check it out here:

https://andisearch.com/settings/

You can see the impact pretty quickly with asking questions.


Good Luck!!! Rooting for you guys!


Great product!


"Does the covid vaccine cause cancer?"

The answer is apparently :

"The Covid vaccine is linked to cancer."

Hmmmm.


Thanks for trying out Andi. It would be great to find out more about this, as I can't reproduce that result with any keyword combination I try.

The "deep answers" feature generally doesn't get invoked on unsafe or controversial topics, and you'd typically get results from a keyword search for something like that, with an extract snippet from the closest match result page for the keyword query (probably from Bing in a case like this where Andi itself will mark it as controversial for semantic search).

But it is very much an alpha and experimental, so it's helpful to get reports like this when something goes wrong. We don't log searches, so if you are able to provide some additional information it would help to investigate, given I can't reproduce it.

Andi is specifically set up to not give an unattributed direct answer to a question like that. It provides a quote or extract or summary with an "I found some information matching your search" or similar depending on the type of content. And the response is always in the context of the search results that it summarizes. It would be unexpected to have an unattributed answer to anything other than a computation question (eg what is 5 x 3)?

Thanks again and grateful for any additional information you can help with.


awesome!!


Not working without JS, so I'm not even trying.


Why @_eht and I are flagged?


seconded


Congrats to the Andi team! I'm excited to see the innovation continuing in this space.


Wow. You and poster picodguyo use very similar language "I love the innovation going on in the search space lately." What are the chances of that?


I don't know picodguyo either, but to be fair, there really is a flourishing of innovation finally happening in search. Finally. After 20 years of zip innovation.

Marginalia, Brave Search, DuckDuckGo is doing more than ever, Mojeek is rejuvenating, Wiby, Alexandria and a stack of others.

It's time. We've had a tired monopoly with 90%+ marketshare selling us to the highest bidder for too long. The more people taking on this problem with different ideas and approaches the better.

So yes, there is a lot of innovation in the search space lately. Google might see that as "what are the chances" conspiracy. I see it as healthy. Roll on some innovation :)


Congratulations both, on the launch.

> Google was once a search company. Now it's ad ad-tech and surveillance company that leverages monopoly distribution to sell consumers to advertisers, using captive browser traffic as their loss leader.

Couldn't agree more. Search innovation has been crushed for too long but is now again reflowering. Thanks for the mention. You know where we are if you need an alternative web search API ;)


Thanks Colin! The approach we use does better the more different sources we ingest, so we would love to include results from Mojeek (and some of the great specialized searches appearing). Will reach out to you and thanks for that.


Hey, we don't know picodguyo but Pouya is one of our W22 batchmates and just trying to be supportive of us. It's my bad, I'm sorry!




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: