Yes, a video describing available VDBs in terms of, e.g. Open/Closed, simplicity of operation, and user interaction patterns (quality/expressiveness of API) would be great!

@ProSaladToss

11 ай бұрын

Seconded

@assethotorch2395

11 ай бұрын

kzread.info/dash/bejne/i6Nho9yPoLrYgso.html You may find it helpful to start with the time frame of the video above!!

@RomeoKienzler

10 ай бұрын

Also important how to extend the vdb with custom distance functions

I just watched an IBM explanation of vector databases and came away lost. Then I watched yours, and got it right away. Point goes to you. ;)

Let's see the more in depth comparison! Also would love to know your take on where it will go? Are they able to automatically generate vectors for your multimodal data already? Are there known companies using vector databases currently? Are there lightweight alternatives to the services you offered? (ie. a numpy verision of a vector database?)

Definitely! I'd love to see comparable benchmarks for common LLM and other tasks (i.e. transfer learning, use-cases in the context of fine-tuning, etc)

Love your work Patrick. Definitely would like to see more on vector databases, especially when you would use one over an array or other options and the pros and cons of some of the types you mentioned (I.e. Pinecone, Milvus, etc.)

The concise, high-level explainer that I needed. Thanks.

Thank you so much, Patrick. Would love to watch a video detailing and comparing all VDBs.

Yes please, more on this topic, I would appreciate it.

@HeyImAK

11 ай бұрын

👆

Nice summary on Vector databases. A comparison of Graph and Vector databases with specific use cases would also help. Thank you

Straight forward and simple. Thanks! 😊

Very useful. Now I can imagine what is a vector database. Thanks

It would be great if you explained how to use vector databases to give LLM's long term memory! 🙏

Great video, thanks! Short and exactly on point -- much appreciated. Yeah, it'd be cool to see more in-depth comparison of the dbs.

Great intro to VD! Would love to see a more in-depth video on some real-world use cases :)

Would definitely be interested in more details, especially on self hosted VDBs

Thank you for this video - just what I needed! If you haven't done one already, please do an explainer comparing. 🙏

Thanks for putting this together! :)

Could you provide an overview on the comparison of different Vector Database providers and how to decide which is better?

I would love to see a comparison of the different Vector Databases!

yup!! looking forward to a detailed analysis and comparison

Good informative video. Thanks!

Thank you, nice and short overview to get an idea of what a vector db is.

Thanks for a nice video! Would be great to learn more on how one could use Redis and PostgreSQL as vector databases. Additionally, more examples and use cases for vector databases would be cool.

This was a very clear explanation. Thank you!

It could be interesting to see a case of adding a vector dbase to an existing sql database, if it can replace it, or if a parallel approach might be interesting, using them side by side, each taking advantage of strenghts. etc.

Yes, looking forward to a more in-depth video.

Brief and to the point. Great video.

You would need to upload ur own embeddings to these db though? Or do they calculate it for you in a multimodal way? Pinecone seems like the former? If so, why not just host locally in your Postgres?

Thank You... It's a great explanation on Vector database. Please make a in depth videos on Pinecone & Redis vector databases

to the point and concise explanation !!

Thanks, this is what I needed to understand the overall idea of vector db.

I would love to see a comparison of the different VDB's and perhaps your thoughts on which one or two are the best. Thanks for a great video.

That was very helpful! Thank you!

thanks for a such a detailed and easily understandable knowledge

Simply explained. Thanks!

Perfectly clear. Thanks!

Great video thank you!

Definitely need a comparisio video and small example code for the top 3 Vector DB's used !! By the way ,Fantastic walk through of the concept !!.

Excellent overview. Many thanks!

Thanks for the Video. you are awesome and very easy to understand what they are. I think Pinecone is quite popular so if there is a video about it, it would be great. Cheers

Yes please, a VDB comparison would be great, and please include FAISS and other self-hosted options.

Thanks, describe very simply what the vector database is and its uses.🥀

Very helpful animations:) How did you do them with exalidraw, if I may ask?

Incredible video

Awesome explanation! Thank you

comparison video for the mentioned VDBs at the end would indeed be awesome!

A breakdown of differences between vector databases would be nice. But also a comparison to graph databases like neo4j and TitanDB et al would help this n00b

This is a really good explanation and visualization

Why isn't KX mentioned in this overview? They have a very strong vector database and support time-series data as well. Formula 1, manufacturing, utilities, and all the banks use them.

Thank you Patrick.

go in depth please we would like to see a video about all these technologies

Thank you, more please :)

This was a really good video! Thanks so much :)

Thanks for the video 👍

woud love to see detailed comparison of the vector databases

I love the video. One critique would be to set up further away from the background to possibly reduce the reverb you're getting

Yes please a video about that. Liked and subscribed

Great video… please go on with the next one

Super helpful!

Great content.I noticed the Elastic name is missing from the list of vector databases. Could you please include it in the list?

It would be great to see a comparison of the vector database companies

An in-depth comparison would be great!

Good explanation. Thumbs up 👍

thanks, you have a video for the comparate diferences quality between?

Can you make a video around pinecone?

Great video! thank you! A big YES for a Vector DB dedicated video Btw I am happy I have found this channel, let's subscribe !

This is a great explanation. But the indexing part is what I was looking for. Nearest neighbor search is already a hard problem in Computer Graphics and gaming (to detect collisions. E.g. if you ever play Madden and do a slow-mo replay, you'll see that the receiver never actually touches the ball. or E.g. cloth simulations for a cape often "clip" into the 3d model of the person wearing the cape).

Supabase also joined the vector DB club a while ago.

Please explain further, any one of the vector databases with an example for each Weaviate, Pinecone..

I would love to see.. what is the Best Vector database... ease of use vs performance. and why. This way we can stop guessing which one to try to use and just know this one is by Standard the best.

Vector DB’s do not get around LLM context size limitations, but it seems like that’s the hot use case for them. Embeddings are not useful until they’ve been transformed though a neural network. I keep looking at these weird use cases like Langchain and I’m baffled people accept their wide margin of failure.

This is like that scene from the Matrix where Neo stops the bullets and he sees the Matrix(humans, objects alike) as lines of code. We are now converting objects like banana and apples into a bunch of numbers which even we can no longer understand looking at them via the vector embedding.

I want to know how indexes work. How does the vector of the search prompt get mapped via index?

nice video - thanks!

helpful >>

Cool, please explain more details about each vector db thanks

Great one

Would love an explanation of indexing and how to use this with an LLM

Good topic 🎉

Can we FAISS vector store in production?

I would enjoy seeing a comparison among these different vector databases. Today I just picked the one that’s most convenient. But there’s probably a better rationale for choosing among them. The other topic I’d like to see is sustainability. For example, if I’m adding a new vector to the database once a week what will happen after 10 years? Is that a sustainable growth when I have a 1016 element vector everyweek of the year or do I need to do something to re-index the database so that my performance doesn’t drop after a number of years? The data I’m creating now would be relevant for many decades.

thank you

Yes please, i habe to decide soon which database, redisearch is cloud only, pinecone too i think

In LLM, I'm facing a token limit issue. With the vector database, will I be able to overcome token issues in llm?

I remember working on a vector database in the mid 1980s. That was a Pick system, mostly used for accounting, warehouse management and the like. Re-innovation. 😁

Yes please!

you can make a mor explication of diferences and optimitzacions cases :) thanks!

Please continue..)

🎯 Key Takeaways for quick navigation: 00:41 📊 Vector databases store Vector embeddings for fast retrieval and similarity search. 01:07 📝 Unstructured data like images, text, and audio can be challenging to store in relational databases, making vector databases valuable. 02:02 🔍 Vector embeddings allow for finding similar items by calculating distances and performing nearest neighbor searches. 03:10 🗂️ Vector databases have various use cases, including equipping language models with long-term memory, semantic search, similarity search, and recommendation engines. 03:50 💽 Examples of vector database options include Pinecone, Chroma, Redis, Cool, Trans, Milvus, and Vespa AI, each with its strengths and capabilities.

A comparison of their underlying architecture would be useful.

Which vectors, you are explaining my vectors of my matrix?

yes please!

Someone needs to make a 3D model of a LLM engine. In the document, “Attention is All You Need” the number 512 is given. Relating the number 512 to your X/Y coordinates, with the 4 quadrants: Would it be accurate for me to assume that the size of your four quadrants are each with 512 total (22.6 x 22.6)? Furthermore, given that there are 512 (22.6 x 22.6) allocated for each word at the input prompt, with 512 (22.6 x 22.6) allocated to each of the 6 of the LLM layers (processed in series). Am I correct in understanding this?

I would like to see a practical application example. Adding vector database info into a group of images and how it's searched for.

I noticed a couple weird effects. When uploading documents to free private gpt the LLM (all of them) would hallucinate when asking simple questions like list the book titles and authors. They would come up with more titles than actually uploaded and complete fantasy titles. This does not happen with ChatGPT! Where lies the problem?

What about kdb+ ?

Yes, please.

Kinda curious , will it be useful just dump all the structure data into vector database , and just ask the LLM about the data :D

Can you point to where I can learn about how the indexing is done at a mechanical/gears-level ? Not like, the state of the art version, so much as like, “here’s the naive approach, and here’s the simplest improvement on it” ?

@zackyezek3760

8 ай бұрын

Simplest way to understand this is from pure higher mathematics. Look up vector spaces, inner product spaces, and “metrics” (e.g. metric spaces). The “vector embedding” is an algorithm (function) that assigns your actual data a vector (N-element array of #s) in a mathematical “Vector Space”. Vector spaces have nice mathematical properties; these ones are usually hyperspaces with hundreds or thousands of dimensions. You can then go further and define all sorts of add-ones; a function that defines the distance between 2 vectors is a “metric”, one that maps your vectors to lower dimensional vectors is a “projection”, and so on. All these functions have to satisfy some abstract mathematical rules to be proper metrics, projections, etc. but once they do you pick up all sorts of additional nice properties for free. The “index” is generally the number(s) generated by applying 1 or more of these functions to your vector. For example, the index could be the # of nonzero indices the vector has. Or it’s length, as defined by some metric. It’s some value(s) that allow searches to quickly prune away or skip most vectors so that full checks and calculations only need to run on a much smaller subspace.

i would love to know more