Hacker Newsnew | past | comments | ask | show | jobs | submitlogin
Show HN: I made a tool to receive alerts when answers change (alertfor.com)
157 points by saran945 on July 28, 2024 | hide | past | favorite | 81 comments
Hi HN,

I've created a tool called Alertfor that scours the open web to find the most relevant and up-to-date answers for complex questions. You can set up alerts to receive continuous updates whenever there are changes or new information becomes available for a given question.

I used an agent framework (Autogen + Sibyl) to collect and answer questions, and I schedule a Celery job to run the same query continuously every six hours.

I would love to hear your feedback, suggestions, or anything else you’d like to say.

Note: I'm submitting this for a second time; I'm not sure if this is against HN policy.



A public demo of some kind is necessary. I saw nothing that was available to view without registration. This made me lose interest very quickly.

Even a quick screenshot would be better than nothing.


Made a quick video/demo and added in the home page. Hope this helps to understand the product.


Show HN rules requires you to have allow people to use it, ideally without creating an account. Otherwise it can just be a regular non-“Show HN” post.

You can find the rules for “Show HN” posts here: https://news.ycombinator.com/showhn.html


The rules suggesting that I would get more feedback if there is no signups. Anyways, luckily I have got enough feedback :) next release will have some public access without signups.


I don’t see it. I see that there’s a link to a tweet [0], but it isn’t accessible without a Twitter account.

The only visible message is:

> Made 50% progress on Alertfor. Its getting smarted in collecting web data.

> Alertfor works like this - "You ask question, it find answers in the web then you can tract the answer for changes" (AQTA - ask question track answer)

> Here are some use cases I tested . . .

I’m assuming the demo is a self-reply to create a thread, but those require a Twitter account to view these days.

[0]: https://x.com/saran945/status/1815053542610067487


Yep don’t have a twitter account, don’t want a twitter account and hate how common it has become to link to tweet threads I can’t read.

The product looks cool, but plus one on a demo video on the page.


Yup, pretty much never clicking anything with an x.com or twitter.com domain again. Ever.


You may be using Mobile, Its showing me in my Laptop. I will work on style issues on mobile. The twitter link is my blunder mistake. .


I'm on desktop (Firefox) and I saw the same thing.


I see it now. It's an embedded vimeo video.


Love the smooth zooms in that video.


You would probably love it less if you watched the video in full screen on a big monitor, the constant zooming made me feel slightly seasick...


but you didn’t??


Please bear with me, Its not showing on mobile.

https://vimeo.com/991338343


I went and looked up the library I didn't know and learned a lot.

A demo would be nice too but I might not have done that otherwise.


my bad, Sure, will make one and upload soon. I did not expect this much interest.


Some public Alerts would be cool to showcase the idea. Potentially publicly visible alerts could be free.

A generic version of https://istheshipstillstuck.com/


Typically users prefer to have their Queries private, may be I should add public/private flag, and make all public alerts as searchable/free to follow. Thank you for suggestion.


That's your Premium tier right there


The page would benefit a lot from an example or two.

"Who is the current NBA champion?"

"How many MPs do the Toddies have?"


I linked to my tweet from home page - https://x.com/saran945/status/1815053542610067487

it has some advanced usage examples. you are right, I will add some sample questions at home page. Thank you.


X only shows the single linked tweet, not the thread or comments, to users that are not logged in. It's not a good way to share information on the web anymore, if it ever was.


I did without thinking much, links are dead within social media. I will do the changes.


Yea, you shouldnt use twitter to share info.

I have never had an account and i dont click on twitter links, since there is a login wall most of the time.

Im pretty sure im not the only one.


Twitter now demands log in to view tweets in many cases. This immediately turns off people who aren't regular twitter users - something to be aware of.


I can't open Twitter/X on Safari iOS because of some "privacy extension".

Make things as easy as possible to view for potential users (aka customers). Don't make them open another app/site just to view some demo.


A significant number of people avoid what Twitter has become, or downright boycott it.

Putting a screenshot on your web page would work a lot better.


Unrolled thread: https://unrollnow.com/status/1815053542610067487


I would really like this to work - lots of public health applications monitoring for events - but it didnt work for the query below - any idea why not, I thought the prompt would be fairly simple given the amount of news coverage on this issue, the result was: There have been no reported events of animal to human transmission of avian influenza H5N1 in the USA since March 2024. Additionally, searches for general news and historical cases related to H5N1 also yielded no results. Therefore, it can be concluded that there have been no significant events or cases reported during the specified timeframe.

Source:

Bing search results for "events of animal to human transmission of avian influenza H5N1 in the USA since March 2024" Bing search results for "avian influenza H5N1 USA news July 2024" Bing search results for "historical cases of avian influenza H5N1 animal to human transmission USA" Bing search results for "CDC avian influenza H5N1 updates" Bing search results for "WHO avian influenza H5N1 updates" Date Time: July 29, 2024, 5:49 a.m.


To clarify, my prompt was: "Please search for events of animal to human transmission of avian influenza H5N1 in the USA since March 2024." Same verbatim search in Google - top result was: "Since April 2024, 10 human cases of avian influenza A(H5) infection have been reported in the United States. Four of these cases were associated with exposure to sick dairy cows and six were associated with exposure to avian influenza A(H5N1)-infected poultry. - 19 July 2024" Happy to continue to test and help you optimise. Would like to get an alert with each new case.


Awesome!!

Kindly try now with same question. I made some changes. I use Bing API as a initial resource selector. there were some issues. fixed it.


Neat. Such services ("delta filters" in information filtering, a sub-part of information retrieval, "diff alerts" in Web mining, another sub-part of information retrieval) are useful to implement anything from price comparison to legal canaries (cf. https://en.wikipedia.org/wiki/Warrant_canary).

It would be useful to have a language to specify exactly what kind of event triggers an alert (e.g. regex, presence/absence of keyword or phrase, or domain from SERP).


Currently, I compare the previous & current answers (with clue) and send both to the LLM to determine if an alert is required for the given alert-task (question + Clue). I should admit that in some corner cases, such as when comparing list page answers, it is not working as expected. I suppose to parse those list as pre-processing step. The alert is triggered on the clue that you give while creating alert.


Interesting idea, how do you decide if the answer has changed? Or to put it differently, which types of questions do you think this works best? eg stock price changes every second VS presidential elections are every few years


Not suitable for real-time data changes. Although I can schedule the scraper to run every minute, this is not an ideal use case for such a setup. This approach is more suitable for web page data that changes every few hours.

Instead of asking simple questions like "Who is the current NBA champion?", I would like to approach this project as a complex data collector.

For example:

I live in Seattle and want to buy a MacBook Pro with the following specifications: - 8-Core CPU - 10-Core GPU - 16GB Unified Memory - 1TB SSD Storage

Can you compare online prices and provide me with a report?

Alert Setup: Please alert me if there are any price drops in any store.


A baby custobot!


This is more or less why I created Huginn. Excited to see more people thinking about systems that let you set complex triggers to know when the world changes.


Huginn is a great product, though I never used, I am aware of the value it provides. Alertfor is not similar to Huginn. my product is part of something else, I extracted and unleased as separate product. Will go in different direction.


Love seeing these creative ideas using LLMs to solve existing problems - however small the problem may be.

I also have many ideas similar to this using LLMs. They're not money makers. But they do solve personal problems for me. However, my ideas are usually bottlenecked by context size, $/token, or models not being capable enough. Luckily, all 3 are improving at an incredible pace.


Could you share your ideas? I've had a few that I don't think are monetizable either but help solve personal problems, like something to help normalize recipes for cooking and to help remember things around home/health maintenance. I'm interested in hearing what others have come up with.


"Love seeing these creative ideas using LLMs to solve existing problems" - yeah, since web wasn't improve as semantic web, I have to use LLM to traverse, understand and collect the web data.


Great app! I have several uses for this.

Some sort of organization of queries into categories would be nice.

Most of my queries don't need to be checked every 6 hours. Some are fine being checked weekly.

You could have a free tier with limited numbers of queries and 24-hour interval to attract new customers.

Someone else mentioned public quereis being free. If I can still get a personal notification for those, then you only spend on sending notifications and not re-running the query for many users. If my needs are served by public queries alone I may not subscribe. I'd weight the risks.

Different methods of receiving notifications are usually only needed for enterprises. I'd weight the benefits of doing that for non-enterprise customers because the cost of debugging is high (think hooking that up to AWS SQS or, worse, some proprietary webhook).


It's a surprise to see weekly queries. If you don't mind, could you please share one or two examples of these queries?

I've added all suggestions from entire thread to my log and will prioritize implementing them.

Regarding pricing, I don't know offering a free tier is right approach or not .. I believe that the queries are complex and long. not sure they could be shared among users. Anyways, for next 2 or 3 months its going to be beta. my current focus is adding value, will ask the users about pricing later. Note that, for each question, it crawls 10-15 pages, or even more if you ask to enrich a table. there is LLM cost, now a days webpages are really huge.

Thank you again for your input and support. Please let me know if you have any other thoughts or questions.


> It's a surprise to see weekly queries. If you don't mind, could you please share one or two examples of these queries?

I think the one in your demo about the MacBook prices is a good one. Something that fluctuates not every hour/day or at least the user doesn't care because it's going to stay the same for a few days at least.


> Someone else mentioned public quereis being free. If I can still get a personal notification for those, then you only spend on sending notifications and not re-running the query for many users. If my needs are served by public queries alone I may not subscribe. I'd weight the risks.

At a large enough scale I imagine you'd see a long tail effect where most of the queries will be then same and then there will be many queries by one or a few users.

Having some kind of dashboard for common queries would be cool.


Yes, totally agree. But from a business perspective, even though on the backend it's the same-ish query (if they are using a semantic cache), it could impact number of service signups and people who subscribe to paid plans. Just saying this with my amateur entrepreneur hat on.


To clarify, my app is the equivalent of SearchGPT + web alerts, using in-app notifications rather than email alerts.


An interesting idea! So - is my query dependant on the latest scrape date for <whatever> AI? If yes, then this has an obvious problem in that cadence for AI scraping is vague at best. If the answer is 'no' -- where are your sourcing the answer to the question I am asking?


The answers provided by the app are sourced directly from the web, and the LLM is used as 1. tool learner/selector 2. query interpreter 3. page understanding and response generation. The answer includes citations from the sources used.

For example. Query: What is the temperature in Seattle? Answer: 17°C Alert Setup: Alert me if it reaches 23°C.

The app continuously monitors the source for updates on the queried information. In this example, the temperature data is regularly checked against the alert condition set by the user.

The real-time aspect depends on how frequently we scrape data using scheduled jobs. Currently, the app checks the data at set intervals (6 hours) to ensure timely notifications.


Using shorter intervals could be a premium feature.


This is an interesting experiment in truth-seeking. People often ask "which media outlet/truth source should I trust", as if there was a single point of contact for facts. In reality, it is more of a moving investigation. It is better to assess over time than to conclude in an instant.


This is a very cool idea!

Obviously, it needs some work (especially judging from all the comments here). You already added an example to the homepage, but my feedback would be:

1. Add more examples of possible alerts. You can honestly just look at the Zapier homepage and copy how they show examples.

2. Include pricing information. I see the "create account" button, but I don't know how much you'll charge (If you don't charge, let me know if I need to add my API keys).

Still, it's a cool idea. It reminds me of existing tools that alert you to flight price changes, Amazon product price changes, etc... There's definitely something here.


Cool.

1. Currently, it's a proof of concept only. In the next release, I will make sure to include some use cases/examples.

2. I haven’t finalized the pricing yet. For now, it's free. When I release, the starting price will be around $7-10 per ~200 runs/questions. I’m also considering offering a free tier; I’ll work that out later. Thank you.


I think this is a good idea for a LLM wrapper app.

I've heard the leaders of the foundational models say a similar thing: bet on the technology getting better. That is, if your business idea becomes less valuable given a smarter model then it isn't a good business idea. Alternatively, if your business idea becomes more valuable given a smarter model then it is a good business idea. (Not that I totally trust them, but this does seem like good advice)

So, even if your current product has some issues now with the diff provided by current models - consider that it will only get better as the models get better.

You are going to see a lot of competition in this space.


I slept, sorry for the delay reply.

The quality of LLM does not improve my product much, I am facing extraction issues, will have to improve continually.

"You are going to see a lot of competition in this space." I welcome them all :)


Feedback: I like the idea and wanted to check out the demo, but the link took me to Twitter/x and I don't have an account there. I could only see one post right until "Here are some use cases I tested . . .". Only on the second visit I noticed the Vimeo embed.


Checked your site. It is pretty cool. What AI models are you working with to find answers in real time? And what's the frequency of re-running these searches for updating the answers? Are you doing it once daily, weekly or monthly?


I use LLM+Agentic framework, it runs every 6 hours. In the next release I will provide UI to customize it (daily, weekly, 6 hours).


Which Agent are you using? And which LLM - is there any easier way to get started?


Extended version Autogen. Added 8 new tools and modified some prompts. I never expected this much quality. linking my tiny benchmarking test in a tweet. https://x.com/saran945/status/1815441284812738734


I'm just not getting it at all. The video is silent so you have to wrap your head around what is happening. I would strongly suggest you try explaining the concept better.


The use cases for your product are too broad. Focus on a promissing niche that has a problem and are willing to spend money on it. Specialize the product and communication to it.


This is a best advice, repeatedly hearing it.

focus on a promising niche that has a problem and is willing to spend money on it is often regarded as best practice. many successful companies have been built on this principle. Unfortunately, this approach has never worked for me. I have worked on many products and followed these best practices, but I have yet to see success. Over the past 12+ years, I have experimented these best practices but nothing has worked. It would work for founders who has good networking /connections, living in CA, best in marketing, have good followers in Social media etc. Founder like me have no such assets. this is a big chicken or egg problem.

Instead, I decided to take a different approach and I create imaginative products based on ideas I'm passionate about and publish them continuously. If there is an interest, I then optimize the product for the interested customers, taking their feedback into account and continuing development accordingly.


How is this different from RAG and Langchain? Sorry for the ignorant question, I recently understood what these concept mean in the scope of LLM.


RAG is to enhance the generation capabilities of a language model by integrating external knowledge sources (DB, KB etc). This is done by retrieving relevant information from a knowledge base or document store and using it to inform or augment the model's responses.

Langchain is a library specifically designed to facilitate the development of applications using LLMs. It provides tools and utilities to build complex NLP pipelines with ease.

Alertfor is a SaaS product that combines LLMs with agents to provide automated search.

Does this answer your question ?


Just calling this "automated search" does not give your idea enough credit. This is a "LLM powered callback function" for the entire internet at the very least. Personally, I think it's brilliant. Please pursue this, because it has tremendous potential.


I disagree lol. Who cares what he uses on the backend whether it’s llamaindex, langchain, etc. Do you discount a news outlet for using Postgres and Ruby because it’s not a novel stack? He’s selling a product with a clear intention.


sure. I do not stop.

I have designed the product for consumers and professionals. Since it's a new kind of product (?), I'm unsure what to call it. The agents continuously search the web on behalf of the user and create a feed for each query. Let me think more . . .


>Alertfor finds the most relevant answer on the web.

Are you intentionally being vague about HOW it finds/decides on the best answer?


I don't understand your question. . .

Do you need more details about How it finds the answer? I already shared most of the part in this thread.

The home page was done in 60 minutes, the whole product was done in 3 days. I was testing and releasing without any intention.


I absolutely love this and I have a use case right now. In terms of pricing, what are you thinking?. Wow, this is very cool.


its free during beta (may be 2 or 3 months). pricing will be around $10 for basic version (200 job runs or questions per month), my current focus is collecting all the feedbacks in my log to fix it. . .


My use case failed

Asked if there were any available 3 bedroom units currently for rent in my building

There aren’t in reality

But service came back saying there were 4


It is an expected behavior. context is missing. if you know any websites that has your request then you may mention it in your query with context. so that you can force the agent to look the info you need.

for e.g any available 3 bedroom units currently for rent in Seattle use these resources: Zillow, redfin, realtors.


Adam, did you expect that behavior?


LLM hallucination is not something the author can solve with a simple hack.

That being said, the LLM can inspect the prompt and give feedback (such as "not enough context, provide some pointers to help research").


I tried to sign up but never got a confirmation email.


this is strange. I saw emails signups and sending emails recently. kindly try with other email. if the problem persists please communicate me saran@tealme.com


What did you use to make the demo video?


I made using https://www.screen.studio/


Good and useful tool




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

Search: