Re: vector search support

[Jonathan S. Katz]

Hi,

On 4/21/23 8:07 PM, Nathan Bossart wrote:

Attached is a proof-of-concept/work-in-progress patch set that adds
functions for "vectors" repreѕented with one-dimensional float8 arrays.
These functions may be used in a variety of applications, but I am
proposing them with the AI/ML use-cases in mind.  I am posting this early
in the v17 cycle in hopes of gathering feedback prior to PGCon.

Thanks for proposing this. Looking forward to discussing more in person. There's definitely demand to use PostgreSQL to store / search over vector data, and I do think we need to improve upon this in core.

With the accessibility of AI/ML tools such as large language models (LLMs),
there has been a demand for storing and manipulating high-dimensional
vectors in PostgreSQL, particularly around nearest-neighbor queries.  Many

of these vectors have more than 1500 dimensions.

1536 seems to be a popular one from LLMs, but I've been seeing much higher dimensionality (8K, 16K etc). My hunch is that at a practical level, apps are going to favor data sets / sources that use a reduced dimensionality, but I wouldn't be shocked if we see vectors of all sizes.

The cube extension [0]
provides some of the distance functionality (e.g., taxicab, Euclidean, and
Chebyshev), but it is missing some popular functions (e.g., cosine
similarity, dot product), and it is limited to 100 dimensions.  We could
extend cube to support more dimensions, but this would require reworking
its indexing code and filling in gaps between the cube data type and the
array types.  For some previous discussion about using the cube extension
for this kind of data, see [1].

I'll start making specific comments on the patches below.

I've stared at the cube code quite a bit over the past few months. There are definitely some clever methods in it for handling searches over (now) lower dimensionality data, but I generally agree we should add functionality that's specific to ARRAY types.

float8[] is well-supported and allows for effectively unlimited dimensions
of data.  float8 matches the common output format of many AI embeddings,
and it allows us or extensions to implement indexing methods around these
functions.  This patch set does not yet contain indexing support, but we
are exploring using GiST or GIN for the use-cases in question.  It might
also be desirable to add support for other linear algebra operations (e.g.,
operations on matrices).  The attached patches likely only scratch the
surface of the "vector search" use-case.

The patch set is broken up as follows:

  * 0001 does some minor refactoring of dsqrt() in preparation for 0002.

+static inline float8
+float8_sqrt(const float8 val)
+{
+       float8          result;
+
+       if (unlikely(val < 0))

Should this be:

  if (unlikely(float8_lt(val, 0))

Similarly:

+       if (unlikely(result == 0.0) && val != 0.0)

  if (unlikely(float8_eq(result,0.0)) && float8_ne(val, 0.0))

This seems pretty benign and may as well do anyway, though we may need to expand on it based on comments on next patch. Question on:

  * 0002 adds several vector-related functions, including distance functions
    and a kmeans++ implementation.

I'd suggest ensure these functions support:

* float4, float8
* int2, int4, int8





If you agree, I'd be happy to review more closely once that's implemented.

Other things:

Nice. Generally I like this patch. The functions seems to match the most commonly used vector distance functions I'm seeing, and it includes a function that can let a user specify a constraint on an ARRAY column so they can ensure it contains valid vectors. While I think supporting float8 is useful, I've been seeing a mix of data types in the different AI/ML vector embeddings, i.e. float4 / float8. Additionally, it could be helpful to support integers as well, particularly based on some of the dimensionality reduction techniques I've seen. I think this holds double true for kmeans, which is often used in those calculations. There's probably some nuance of how we document this too, given our docs[1] specify real / double precision, and smallint, int, bigint. (Separately, we mention the int2/int4/int8 aliases in [1], but not float4/float8, which seems like a small addition we should make). * kmeans -- we're using kmeans++, should the function name reflect that? Do you think we could end up with a different kmeans algo in the future? Maybe we allow the user to specify the kmeans algo from the function name (with the default / only option today being kmeans++)?

  * 0003 adds support for optionally using the OpenBLAS library, which is an
    implementation of the Basic Linear Algebra Subprograms [2]
    specification.  Basic testing with this library showed a small
    performance boost, although perhaps not enough to justify giving this
    patch serious consideration.

It'd be good to see what else we could use OpenBLAS with. Maybe that's a discussion for PGCon.

Of course, there are many open questions.  For example, should PostgreSQL

support this stuff out-of-the-box in the first place?

Yes :) One can argue an extension (and pgvector[2] already does a lot here), but I think native support would be generally helpful for users. It does remove the friction of starting out. There's also an interesting use-case downthread (I'll comment on that there) that demonstrates why it's helpful to have variability in vector size in an ARRAY column, which is an argument for supporting it there.

And should we
introduce a vector data type or SQL domains for treating float8[] as
vectors?  IMHO these vector search use-cases are an exciting opportunity
for the PostgreSQL project, so I am eager to hear what folks think.

Thanks,

Jonathan

[1] https://www.postgresql.org/docs/current/datatype-numeric.html
[2] https://github.com/pgvector/pgvector

Having a vector type could give us some advantages in how we store/search over the data. For example, we perform validation checks up front, normalize the vector, etc. and any index implementations will have less work to do on that front. We may also be able to give more options to tune how the vector is stored, e.g. perform inversion on insert/update. Again, it's a fair argument that this can be done in an extension, but historically we've seen reduced friction when we add support in core. It'd also make building additional functionality easier, whether in core or an extension.

OpenPGP_signature
Description: OpenPGP digital signature