Show HN: Weaviate – Build your own generative health search engine

https://github.com/weaviate/healthsearch-demo

by edichief • 3 years ago

39 26 3 years ago

We are super excited to release our latest open-source demo, Healthsearch. This demo decodes user reviews of supplements and performs semantic- and generative search on them, retrieving the most related products for specific health effects, and leveraging Large Language Models to generate product and review summaries. The demo can understand natural language queries and derive all search filters directly from the context of your query.

CShorten3• 3 years ago

Epic! Love this, one of the best Weaviate demos I've seen!

edichief5• 3 years ago

Thanks! Working on this and all the exciting features was really fun, from semantic and generative search to using Weaviate as a Semantic Cache and translating natural queries to GraphQL. The live demo was also recently updated with some good stuff! https://healthsearch-frontend.onrender.com/

robertlagrant2• 3 years ago

Epic! Love this, one of the best Weaviate demos I've seen!

Just out of interest, do you still work for Weaviate? Probably worth mentioning.

CShorten3• 3 years ago

Hey robertlagrant, my apologies -- still figuring out best practices on Hacker News. Will be more mindful of this going forward!

etiennedi3• 3 years ago

Love the visual design, too. The generative/RAG/semantic search is the interesting part, but it's also just very pleasant to look at. Which can go a long way.

EDIT: Disclaimer, I am affiliated with Weaviate.

robertlagrant2• 3 years ago

Love the visual design, too. The generative/RAG/semantic search is the interesting part, but it's also just very pleasant to look at. Which can go a long way.

Just out of interest, do you still work for Weaviate? Probably worth mentioning.

etiennedi3• 3 years ago

Fair point, edited my post. The fact that I love the visual design is my personal opinion though :-D

thatcherthorn• 3 years ago

    # Easter Egg 
     if query_text == "easteregg": 
         return JSONResponse( 
             content={ 
                 "query": " Congratulations, you rolled the demo!", 
                 "results": {}, 
                 "generative_summary": "You just got rick-rolled...", 
             } 
         )

Haha

edichief5• 3 years ago

Haha awesome you found it so quickly!

quickthrower22• 3 years ago

==

Interesting encoding/embedding and vector search you have there :-)

edichief5• 3 years ago

Haha! I initially used the rick-roll meme whenever the demo crashed, but at some point, I couldn't stand getting rick-rolled by my demo any more... :)

ke88y3• 3 years ago

One of my college room mates ended up with kidney stones at a very young age (early 20s) because of a dietary supplement. He got somewhat lucky; there have been cases of liver failure [3], damage to vision [4], and many other serious health complications. Worse still, diagnosis can be difficult because in the US dietary supplements are covered under separate rules from drugs [1], the FDA is limited to post-market enforcement [2], and ingredient labels are often incomplete or wrong [5,6,7].

If Healthsearch is based on the contents of advertisements and reviews, then it may not be much more than a search engine for at best unsubstantiated marketing claims and anecdotes about the placebo effect, and at worst summarization and promotion of dangerously inaccurate or incomplete product information.

I think this is a good example of a space where generative AI is probably likely to do far more harm than good, and where the best advice remains to seek out the advice of educated and licensed medical/dietetic professionals, where education ensures a minimum of competence and licensure ensures a minimum of skin-in-the-game.

Just remember: if it's a dietary supplement, then it's not a food or a drug. It's a third thing and mostly a wild west. You have no idea what you are putting in your body, and the words on the label are just that: mere words designed to get you to purchase something. Without ground-truth data, it's all GIGO.

[1] https://www.fda.gov/food/dietary-supplements

[2] https://www.fda.gov/food/information-consumers-using-dietary...

[3] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3076034/

[4] https://journals.sagepub.com/doi/10.1177/2474126419877567

[5] https://www.health.harvard.edu/blog/whats-in-your-supplement...

[6] https://www.gao.gov/assets/gao-19-23r.pdf

[7] https://www.gao.gov/assets/gao-19-23r.pdf

edichief5• 3 years ago

Thank you for your comment. I appreciate your thoughtful feedback and couldn't agree more with many of your points. The primary purpose of this demo isn't to encourage self-diagnosis or the unmonitored consumption of dietary supplements but to demonstrate the potential of semantic search in a field that we know is fraught with misinformation.

As you rightly pointed out, the regulatory landscape of dietary supplements is quite different from that of drugs or food, and this often leads to a host of issues, including misleading labels and unproven health claims. Our goal is not to feed into these problems but instead to use technology to parse through the vast amount of data and draw out meaningful information. Dietary supplements are not a cure-all or replacement for professional medical advice.

They can, however, be beneficial in some instances if used correctly and under the guidance of healthcare professionals. Melatonin, a hormone that our bodies naturally produce in response to darkness, plays an essential role in regulating our sleep-wake cycle. Research has shown that supplemental melatonin can benefit those with disrupted sleep patterns. For instance, a systematic review and meta-analysis published in "Sleep Medicine Reviews" found that melatonin reduces sleep onset latency (the time it takes to fall asleep), increases total sleep time, and enhances overall sleep quality [1]. Another study in the "Journal of Clinical Sleep Medicine" reported that melatonin could benefit those with Delayed Sleep-Wake Phase Disorder, a condition characterized by a significant delay in sleep onset and waking times [2]. These studies suggest that, under specific circumstances, melatonin can be an effective sleep aid, emphasizing the importance of consulting with a healthcare provider to assess individual needs and potential benefits.

[1] https://pubmed.ncbi.nlm.nih.gov/23691095/ [2] https://pubmed.ncbi.nlm.nih.gov/26414986/

ke88y3• 3 years ago

That's a nice narrative.

How is any of it relevant to a system that uses LLMs trained on user-written reviews of health supplements to recommend products based on end user's descriptions of health conditions?

anonymous3442• 3 years ago

this is bulls*it

all the smart people know that big pharma with their _symptom not the reason_ medicines are making billions and investing hundreds of millions for political influencing, internet influencing and many other illegal things. And herbs, and unpatentable things like sodium bicarbonate, borax and silverwater are the things they have been shooting down for decades successfully (sadly)

Big pharma medicine kills millions every month. Vitamin supplements maybe 3 people in the last 50 years. Think about it.

ndejaco3• 3 years ago

I wonder what Weaviate users think of the GraphQL interface. When using an LLM to do this type of query generation it needs to understand the schema and be trained to generate a semantically correct request. We found better results putting the schema itself in the prompt to help generate valid requests from English and I see here it is added in the system prompt as well. I wonder if a simpler interface might suffice so that the retrieval can be more flexible, though I don't have a better alternative at the moment :)

edichief5• 3 years ago

Great to hear that you also experimented with this! I agree that this approach is limiting and I think this can be further improved. I'm wondering whether a Retrieval Augmented Generation (RAG) could make this more flexible.

nostrebored2• 3 years ago

The near-bot responses from people who do not clearly identify that they work at this company are very frustrating. This feels like using HN for manufactured marketing campaigns.

For instance, Take CShorten's last two responses to threads that do not clearly identify that they are direct marketing for Weaviate.

https://news.ycombinator.com/item?id=36670548 Thanks so much for sharing!

https://news.ycombinator.com/item?id=36760269 Loved this paper - so much opportunity to explore Retrieval-Augmented Generation with these longer input LLMs + Vector Databases & Search!

Take another example of a user whose submission CShorten commented on: https://news.ycombinator.com/threads?id=bobvanluijt

The overwhelming majority of these comments are _spam_.

If you work at Weaviate, I'd recommend you stop doing this stuff. I'm way more likely to look at a competitor who feels less like a bot network.

freedomben• 3 years ago

IMHO, having employees (and otherwise affiliated) come to the thread can be a huge positive for the conversation, but it is indeed very important to disclose the affiliation! Don't fall into the trap of "well I'm just commenting my personal opinion, I don't need to disclose."

I think you can make a perfectly logical and reasonable argument about why it should be fine, but in the end it doesn't matter. It will backfire because it's been abused to the nth degree in the past and now people's antennae are up.

Transparency around affiliation is key. If you're commenting something of substance, then people will appreciate it and value your comment. If you're not adding anything substantive, then it's best not to comment at all.

etiennedi3• 3 years ago

Thanks, appreciate your perspective. This is much more insightful than just "affiliated==botnet". I agree that this can be a huge positive and - at least for me - that was the motivation for wanting to be active in the comments; to see if there are any challenging questions, feedback, etc.

As a HN user, I often go straight to the comments and only look at the submission if there's an interesting discussion there. I can totally see how spammy-looking comments with an undisclosed affiliation can have the exact opposite effect.

(affiliated with Weaviate)

CShorten3• 3 years ago

Hey all, my apologies for these comments! I will be more mindful of this going forward!

quickthrower22• 3 years ago

Weaviate is pretty cool IMO. It is open source and fairly easy* to get this running locally and do stuff with it. Definitely worth playing with. In the past I have done a search-type thing for a side project and used Elastic Search. While I think ES is great, it might be overkill and something like this might be easier to manage.

You can even run Weaviate as an embedded python package rather than a separate process, so compare that to running an ES cluster and I think life gets a lot easier. Although you are probably working 'closer to the metal' but really if you use ES you will be tuning the hell out of it anyway.

* depends on your experience, but if you can use docker, and know roughly what an embedding basically is a concept, then not too hard to get something up and running.

anonymous3442• 3 years ago

where can you download the database? where the health reviews are ?

Show HN: Weaviate – Build your own generative health search engine

Related Stories