Deep Lake AI Knowledge Agent conducts Deep Research on your data, no matter its modality, location, or size. Deep Lake supports multi-modal retrieval from the ground up. It uses vision language models for data ingestion and retrieval so that you can connect any data (PDFs, images, videos, structured data, etc.) stored anywhere, to AI. Over time, it learns from your queries, tailoring the results to your work! Deep Lake is used by Fortune 500 companies like Bayer, Matterport, and others.
Interactive
Free Options
Launch Team / Built With
Render: Get up to $100K in startup creditsCloud infrastructure for founders who want to move fast and NOT break things
Promoted
Hi Product Hunt!
I'm Davit Buniatyan, CEO of Activeloop (YC S18). We're introducing Deep Lake AI Knowledge Agent, which conducts Deep Research on your data, no matter its modality, location, or size. Deep Lake supports multi-modal retrieval from the ground up. It uses vision language models for data ingestion and retrieval so that you can connect any data (PDFs, images, videos, structured data, etc.) stored anywhere, to AI.
Deep Lake can search your data from S3, Dropbox, and GCP. Over time, it learns from your queries, tailoring the results to your work!
Here are some example use cases:
1. Financial analysis
Connect earnings call transcripts, data from the Bloomberg terminal, and PowerPoint from the earnings reports or annual reports to analyze thousands of companies!
2. Scientific research
Connect patient EHR data, medical research (both public and internal), and lab results to discover new drugs (this is what a few Fortune 500 and leading biotech companies are doing with us already).
3. Legal stuff like patent search and generation, and contract review.
Cross-reference diagrams, scans of documents, invoices, etc. and ask questions over those! A few companies use us to search across millions of patents worldwide and generate novel, defensible patents! Examples here and here.
You may ask, what is our superpower?
There's a long answer, and a short answer (I'll let Sasun, our Director of Engineering handle the more technical part). But in short, I've spent a good chunk of my time at Princeton researching how to store complex data and connect it to AI.
TL;DR - we're multi-modal (i.e., handle all kinds of data, not just text or vectors) and highly accurate!
This is hard to achieve, but there's a way to store it in a way that's AI-native - and represent data that is unstructured in a more columnar way. At Activeloop, we figured out how to connect any data from your storage and process it in a way to extract as much information from it as possible without complicated OCR pipelines. Then, we store it in an AI-native format and can highly efficiently (more on that from Sasun) retrieve it from the storage, querying across multiple datasets at once.
With the model reasoning improving (our Agent can work with any model, closed-source or otherwise), we've unlocked the missing piece -> we provide the best-in-class AI search, while smart models help us generate answers to complex questions we find the right evidence for!
I'm very happy to share our hard work with you - please ask me anything about the product, our journey through YC to here, and more.
@david_buniatyan very exciting!
@david_buniatyan 🔥
@vahan25 indeed!
@david_buniatyan This is absolutely brilliant!
@david_buniatyan this is fantastic, really excited to see it's happening
Hi Product Hunt!
I'm Sasun, Activeloop's (YC S18) Director of Engineering. I've previously co-founded Pixomatic, one of the early successful photo-editing apps. Naturally, one of the things that excites me is how to visualize (and query) unstructured data, like images.
Except… back in the day, there was no SQL for images.
Then I met @david_buniatyan, who started Activeloop with that mission - make the way complex data - like images, videos, text, etc. stored in a more organized way - and make it easily connectible to AI (for training, and asking questions!).
This comes with a number of exciting technical challenges.
1. Unstructured data is… well… unstructured. It's hard to search across such data (imagine saying I want all the images that contain bicycles larger than 200x350 pixels, and two people in them).
Retrieval systems until Deep Lake weren't fit for that.
2. Vector Search is inaccurate.
Achieving accuracy in AI-generated insights is challenging, especially in sectors like legal and healthcare where accuracy is paramount. The issue magnifies with scale—for instance, when searching through the world’s entire scientific research corpus.
Most data lives in data lakes (S3, AWS, GCP)
3. Limited Memory
Bolting a vector index onto traditional database architectures does not provide the scalability required by AI workloads. As the scale of your dataset increases, the memory and compute requirements scale linearly. For datasets that grew past 100M, the cost becomes prohibitive to maintain the index in memory.
My team and I focused on building this as Deep Lake's ‘unfair advantage’, since we're geared more towards analytical cases where the users need to ask questions across complex, large data. As a result, we're up to 10x more efficient than in-memory approaches.
4. AI Agents can fail… spectacularly
Not claiming we've totally solved this issue, but if there's even 1% probability of failing or responding inaccurately every time, in a complex, multi-step system there will be this ‘butterfly’ effect where with every additional step, the probability to fail will increase.
So increasing retrieval accuracy is important - in critical verticals (autonomous driving, life sciences, healthcare, finance) it can be either a matter of life and death, or incalculable losses.
More on this in detail (with benchmarks here).
Feel free to ask me any technical questions on Deep Lake's capabilities, I'd be happy to answer.
Thanks for having us.
@khustup thanks for being a part of our journey and shipping an amazing product!
@mikayel_harut I'd say the most challenging part is to index the large scale data on object storage, keeping the balance between latency and scale.
This is exciting! Compelling demo!
I am curious how effective Deep Lake's integrated knowledge retrieval approach is for avoiding hallucinations and finding relevant articles not found by other tools in the same space?
@ngalstyan4 good question!
I wouldn't say it's possible to completely avoid hallucinations. Hallucinations happen for two reasons: wrong context, wrong answer by model, and right context, but still a wrong answer by a model. In the latter case, we can't do much. But we focus on making the former case obsolete!
How we do this:
Query planning and gathering context from various datasets.
Querying flexibility (choose to do hybrid, vector, keyword search, etc.)
Multi-modality (on ingestion, gaining more depth of insight into what data is about - what is contained in figures, for instance), which helps pass more imoprtant context to the model.
We also learn over time what queries you consider correct, which helps further improve search experience and increase retrieval accuracy. No other vendor can handle this, as well as #3 as well as we do!
OCR-free retrieval of documents, images, and videos? This truly feels like the next era of AI-driven data utilization! Huge congratulations on your launch! 🎉
@kay_arkain thank you so much, Kay! You're absolutely right.
@avag_simonyan hi there! Actually yes - we have a bit technical blogpost talking about diagram-related use case. This doesn't use the AI Knowledge Agent but if it's possible with just code, it would be even easier through the AI Knowledge Agent UI. Handwritten comments wouldn't be an issue (no more than for an OCR). Specifically for architecture, we haven't tested it but should be extensible from other blueprints without an issue!
Would love to chat about the use case with you.
Here are few things that I really liked in my first experiments.
1) Being able to use lots of public data sets before porting my own. Gave use good idea what the usability would be like so it was really helpful.
2) Found the onboard for data access quiet helpful, lots of things were present that were really on point, e.g. CORS configs.
3) Really liked step by step presentation of things before final output would be generated. Would want to see more control there.
Example, when doing a search it allows to see "Generating TQL", which is great, now I want to see what would happen if I were to change TQL itself. Could I see side by side data? Could I see performance metrics. Not all queries are equal in backend side so may be I want to speed things up or help things get to better results.
I can see that the platform does have the ingredients so looking to explore further and give more updates here.
The traditional stack to chat with your own multimodal data is indeed becoming too large to maintain.
Deep Lake is a game changer - it takes away all the complexity, while maintaining similar quality.
Congrats on the launch 🚀
@armzak thank you so much! Agreed.