Skip to content
TheChatArchiveThe Chat Archive
gpt-4-0314· OpenAIHugging Face Arena Human Preference 55K

generate notes for me welcome back everybody so it's quarter past now it's time to get…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with Metallama-2-13b-chat’s answer.
  1. User

    generate notes for me

    welcome back everybody so it's quarter

    past now it's time to get our next talk

    started and this is I'm I'm intrigued

    that everybody's here because apparently

    it's on the monitor there it still says

    this is like a secret talk title but um

    I it's it's my privilege really to be

    able to announce the title of this talk

    which is a vector search powering the

    next generation of applications

    um and we have uh you're a product lead

    aren't you

    um

    can we please have a big round of

    applause for Ben flask

    [Applause]

    hi everybody can you hear me all right

    cool awesome well uh thanks everyone for

    for being here I'm really excited to be

    uh talking about Vector search today my

    name is Ben plast I'm a product manager

    at mongodb been here for about four

    years working on a range of services in

    kind of the analytics and AI space

    um I'm just gonna quickly start out

    though with uh

    a couple of questions so I'm I'm curious

    if you've ever kind of found yourself

    looking for something but haven't quite

    had the words

    um maybe you've you know remembered some

    characteristics of a movie but not the

    name or maybe you've been trying to get

    a sweatshirt that you had back in the

    day

    uh and you have a picture but you don't

    know how else to look for it uh or maybe

    you're looking to power an artificial

    intelligence with some long-term memory

    that could take over the world in the

    future

    or and and maybe you're just looking for

    an interesting talk to see at

    mongodb.local if if any of these things

    have kind of crossed your mind then I

    think you might be at the right talk

    um

    and so we're going to be talking about

    Vector search

    um but we're actually going to start by

    kind of going a bit back and talk about

    what a vector is and kind of Orient to

    some of the concepts that you need to

    have in order to to use this capability

    um then we'll talk about the capability

    that we're releasing and how it works uh

    we'll follow up by covering some use

    cases uh I then go through a demo where

    you can actually see it in action in

    slideware

    um and then we'll finish up by talking

    about uh what's next

    so what is a vector a vector is a

    numeric representation of data and

    related context and so if you take a

    very simple example like a quick brown

    fox

    it would be a kind of array of floats

    which represents both the data so the

    words a quick brown fox uh but also kind

    of the context how a relates to Fox you

    know what does brown mean Fox is not a

    dog it's also not a cat things like this

    would be the associated context that

    could be represented inside of this

    vector

    [Applause]

    so

    obviously that's a very powerful concept

    to be able to represent so much inside

    of just this string of numbers you're

    probably kind of immediately asking

    yourself how would you get a vector uh

    and the way that it works is that you

    would pass your data whether it be an

    image or a text or audio through an

    encoder of some sort and and these are

    machine learning models that basically

    take this data and transform it into

    this numeric representation

    we're going to go deeper into what this

    means and how it works later on but this

    is just something to kind of keep in the

    back of your mind

    so once you've taken data and

    transformed it into a vector there are

    some really interesting properties about

    these vectors that allow us to provide

    the service that we're releasing today

    and and kind of a key one is that

    similar vectors that are plotted in

    space will be near one another in this

    High dimensional space

    and so to take a very simple example if

    we have kind of a graph in two

    dimensions and they're going to be a

    bunch of two-dimensional graphs that I'm

    going to show during this presentation

    it's important to realize that well

    two-dimensional graphs are great for

    kind of showing this concept and

    understanding what's happening

    all of the vectors that we're talking

    about here are really high dimensional

    right we're talking about thousands of

    Dimension which are really you know

    impossible to visualize and so we'll

    just stick in a two-dimensional world

    but kind of keep that again in the back

    of your mind because that's really where

    a lot of the power comes from with this

    next generation of vectors and

    capabilities so continuing on with the

    example

    imagine you have a point and let's say

    that this point represents man and

    there's a there's a vector associated

    with it

    um

    you could also have another point that

    represents the word woman and it would

    be conceivable that those two pieces of

    text when vectorized or embedded those

    are you know synonyms would find

    themselves near one another they would

    be close together in this High

    dimensional space because they are

    semantically similar right they have

    similar meanings so they're both people

    or humans Etc

    this is kind of a core concept that

    allows us to provide the service of

    vector search if this didn't happen if

    this didn't come about because of the

    way that these pieces of data are

    embedded and modeled then it wouldn't be

    possible for us to provide this service

    so that's super exciting

    but if you zoom back out again and you

    say here I have a bunch of circles if

    again I put them on a graph I now can

    kind of think of them as you know

    vectors in space and we're noticing that

    they actually kind of cluster together

    so these could be a bunch of different

    points maybe movies in some database and

    it's conceivable that if you took those

    movies and embedded the data about them

    their plots you would see that they

    would show up near one another and so

    that's kind of what we're seeing here

    and because they show up near One

    Another We're actually able to draw

    interesting information from them

    but what's kind of core about that

    information that we're able to draw from

    them is how they relate to each other

    and the way that we determine these

    relations is kind of through two things

    one is how they get embedded so what the

    model takes that Source data and turns

    it into in a high dimensional vector and

    then also how do we calculate the

    distance between these vectors

    so

    to imagine an example here if we had man

    and woman and we had king and queen as

    concepts with one embedding model one

    way of representing this data along with

    one way of determining the distance we

    may find that man and King have a

    shorter distance and and are more

    semantically similar and woman and queen

    are more semantically similar right that

    could be produced by having one

    embedding model and one distance

    function and again we'll get into those

    details a bit further in a second but

    simultaneously you could also have a

    different embedding model and use a

    different distance function that would

    result in actually man and woman being

    closer together and more semantically

    similar and king and queen being more

    semantically simple so this is another

    way that that is kind of determined

    which then impacts exactly what happens

    during Vector search and again just to

    reiterate because I know people are kind

    of Milling in

    we're laying the groundwork for Vector

    search and you know how this ends up

    kind of uh impacting the service that we

    end up offering so

    we've talked about vectors how they

    relate to one another how they end up

    falling in this High dimensional space

    when it comes to then using these

    vectors there are some other important

    Concepts that we need to know and so we

    already talked about the fact that like

    data when transformed into vectors can

    create these clusters that are

    semantically similar so groups of data

    that have some sort of relevance or

    reference or relationship between them

    the uh

    K nearest neighbor is a way to go and

    search through these vectors that is an

    algorithm that lets you kind of look at

    these vectors and find neighbors to them

    and K represents the number of neighbors

    you're looking for when you talk about

    them and so K nearest neighbors is a way

    to look through vectors to find similar

    ones to a Target

    when you use K nearest neighbors you

    have to define a specific similarity

    function and so in Atlas Vector search

    we support three there are more that you

    can use out in the market but but these

    are kind of the the three most popular

    ones the first one is euclidean distance

    which represents the distance between

    the ends of vectors

    um the next is cosine which represents

    the angle between vectors and the third

    is dot product which is based on both

    the angle between the vectors and the

    magnitude

    each one of these functions is good for

    a different type of use case and so

    euclidean for instance is really good

    for dense data where specific values

    matter so for instance if you wanted to

    do image similarity it's really

    important that this pixel looks just

    like this pixel that that would be a

    good example of where euclidean would be

    relevant another option would be cosine

    right and that's where sparse data and

    orientation is important and that's kind

    of complex but what it really gets down

    to is if you had examples like text

    Concepts or themes that you wanted to be

    able to find the distance between cosine

    is really good for that because maybe

    one word can represent a lot about what

    a paragraph is about more than having

    like several words repeated so that's

    where cosine is good then dot product is

    really good for sparse data where both

    the orientation and the intensity matter

    together

    so

    once you've decided you're going to use

    Canary Sabers as an algorithm you know

    that you're going to use a certain type

    of distance function the way that all of

    this kind of becomes reality is we have

    to be able to find this information very

    quickly and the way this is done is with

    something called approximate nearest

    neighbor and uh what's called a

    hierarchical navigable small world graph

    now that's a lot of words but it's

    actually you know pretty straightforward

    uh

    approximate nearest neighbor basically

    takes an approach to finding the nearest

    neighbors just like K nearest neighbors

    does but it does it in an approximate

    manner obviously which prioritizes speed

    over accuracy and what I mean by that is

    at the margins where we're figuring out

    does does this Vector fit as a neighbor

    you could kind of drop those outliers

    out and that's really what the

    approximate is getting at so you're not

    going to get necessarily all of the

    nearest neighbors but again you're going

    to get the majority and this is really

    kind of happening at the margins and

    what this allows for is a really fast

    traversal of these vectors so that you

    can find

    semantically similar Concepts quickly

    which is obviously very important within

    the concept of a operational database

    so

    those are all the basics

    so you're all oriented to Vector search

    you understand more about vectors

    there's one other concept prior to

    talking about the feature itself that I

    wanted to cover and that's kind of why

    now

    um so Vector databases have been around

    for a long time and we've had Vector

    search inside of mongodb for actually a

    very long time in the form of geospatial

    data those can be thought of as

    two-dimensional vectors that can be

    searched for other near vectors but

    what's different now is that we're

    looking at vectors in very high

    dimensional space and with the strategy

    I just talked about on the prior slide

    we know why we're able to look through

    them fast but we don't know yet why

    these high dimensional vectors are able

    to represent data so well and and that

    really comes down to recent Evolutions

    in machine learning and that encoder

    model that I talked about much earlier

    so

    pre-2000 we had approaches to modeling

    data in vectors that would use manual

    feature engineering

    um with you know strategies like bag of

    words and TF IDF the next Evolution over

    that was something like word to vac and

    glove you know we're

  2. Assistant

    moving into the early 2000s and this is where we start to use neural networks to help generate these vector representations. However, these early models were still not as powerful as what we have today.

    In recent years, we've seen the rise of transformer models like BERT, GPT, and other state-of-the-art models that have significantly improved the way we can encode and represent data in high dimensional space. These models have made it possible to create vector representations that capture a much richer understanding of the data and its context, enabling more powerful search capabilities.

    So now that we have a better understanding of vectors, their properties, and how they're generated, let's talk about Atlas Vector Search, the new feature in MongoDB.

    Atlas Vector Search is a powerful search feature that allows you to perform similarity searches on high dimensional vectors. You can use this feature to search for similar images, text, audio, or any other data that can be represented as a high dimensional vector.

    To use Atlas Vector Search, you will first need to generate vectors for your data using an appropriate encoder model. Once you have these vectors, you can store them in your MongoDB collection alongside the original data.

    To perform a vector search, you can use the $vector query operator in your MongoDB queries. This operator allows you to specify a target vector and a distance function, and it will return the K nearest neighbors to the target vector based on the specified distance function.

    Some potential use cases for Atlas Vector Search include:

    1. Image similarity search: Find similar images in a database based on their visual features.
    2. Text search and document retrieval: Find documents or text that are semantically similar to a given query.
    3. Audio similarity search: Find similar audio files based on their audio features.
    4. Anomaly detection: Identify unusual or outlier data points in a dataset.

    In summary, Atlas Vector Search is a powerful new feature in MongoDB that leverages the latest advances in machine learning and high dimensional vector representations to enable fast and efficient similarity search across a wide range of data types. With this feature, you can build more intelligent applications that can better understand and respond to the complex relationships between different pieces of data.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.