Dimensionality Reduction

In the case of embedding models trained on large bodies of text, most of the concepts they learn will be unused when dealing with any single piece of text. For collections of documents that deal with specific topics, only a fraction of the language models learned associations will be relevant. Dimensionality reduction is an important technique to improve performance on your documents, both in terms of quality and latency for embedding recall using nearest neighbor search.

Why Dimensionality Reduction?

  • Improved Performance: Reducing the number of dimensions can significantly improve the computational efficiency of machine learning algorithms.
  • Reduced Storage: Lower-dimensional data requires less storage space.
  • Enhanced Visualization: It is easier to visualize data in two or three dimensions.

What is Matrix Decomposition?

Dimensionality reduction is a key technique in machine learning and data analysis, particularly when dealing with high-dimensional data such as embeddings. A table full of embeddings can be considered a matrix, aka a 2-dimensional array with rows and columns, where the embedding dimensions are the columns. We can use matrix decomposition methods, such as Principal Component Analysis (PCA) and Singular Value Decomposition (SVD), to reduce the dimensionality of embeddings.

Matrix decomposition involves breaking down a matrix into simpler, constituent matrices. The most common decomposition techniques for this purpose are:

  • Principal Component Analysis (PCA): Reduces dimensionality by projecting data onto a lower-dimensional subspace that captures the most variance.
  • Singular Value Decomposition (SVD): Factorizes a matrix into three matrices, capturing the essential features in a reduced form.

Dimensionality Reduction with PostgresML

PostgresML allows in-database execution of matrix decomposition techniques, enabling efficient dimensionality reduction directly within the database environment.

Step-by-Step Guide to Using Matrix Decomposition

Preparing the data

We'll create a set of embeddings using modern embedding model with 384 dimensions.

CREATE TABLE documents_with_embeddings
id serial PRIMARY KEY,
body text,
embedding float[] GENERATED ALWAYS AS (pgml.normalize_l2(pgml.embed('intfloat/e5-small-v2', body))) STORED
INSERT INTO documents_with_embeddings (body)
VALUES -- embedding vectors are automatically generated
('Example text data'),
('Another example document'),
('Some other thing'),
('We need a few more documents'),
('At least as many documents as dimensions in the reduction'),
('Which normally isn''t a problem'),
('Unless you''re typing out a bunch of demo data');
timer 46.823
CREATE VIEW just_embeddings AS
SELECT embedding
FROM documents_with_embeddings;
timer 14.259ms


Models can be trained using pgml.train on unlabeled data to identify important features within the data. To decompose a dataset into it's principal components, we can use the table or a view. Since decomposition is an unsupervised algorithm, we don't need a column that represents a label as one of the inputs to pgml.train.

Train a simple model to find reduce dimensions for 384, to the 3:

FROM pgml.train('Embedding Components', 'decomposition', 'just_embeddings', hyperparams => '{"n_components": 3}');
timer 48.087 ms
INFO: Metrics: {"cumulative_explained_variance": 0.69496775, "fit_time": 0.008234134, "score_time": 0.001717504}
INFO: Deploying model id: 2
project | task | algorithm | deployed
Embedding Components | decomposition | pca | t

Note that the input vectors have been reduced from 384 dimensions to 3 that explain 69% of the variance across all samples. That's a more than 100x size reduction, while preserving 69% of the information. These 3 dimensions may be plenty for a course grained first pass ranking with a vector database distance function, like cosine similarity. You can then choose to use the full embeddings, or some other reduction, or the raw text with a reranker model to improve final relevance over the baseline with all the extra time you have now that you've reduced the cost of initial nearest neighbor recall 100x.

You can check out the components for any vector in this space using the reduction model:

SELECT pgml.decompose('Embedding Components', embedding) AS pca
FROM just_embeddings
timer 14.259ms

Exercise for the reader: Where is the sweet spot for number of dimensions, yet preserving say, 99% of the relevance data? How much of the cumulative explained variance do you need to preserve 100% to return the top N results for the reranker, if you feed the reranker top K using cosine similarity or another vector distance function?