Week 11: Vector Databases

DSAN 6000: Big Data and Cloud Computing
Fall 2025

Class Sessions

Author

Affiliation

Amit Arora and Jeff Jacobs

jj1088@georgetown.edu

Published

Monday, November 10, 2025

Open slides in new tab →

Vector Spaces: The Geometry of AI!

Why vector spaces over keyword matching?
What if words have different meanings in different contexts?

Word Counts: Good Enough?

Just four keywords per section:

Section	Keywords
U.S.	`state`, `court`, `federal`, `republican`
World	`government`, `country`, `officials`, `minister`
Arts	`music`, `show`, `art`, `dance`
Sports	`game`, `league`, `team`, `coach`
Real Estate	`home`, `bedrooms`, `bathrooms`, `building`

For each article, vote for section with highest keyword count:

	Arts	Real Estate	Sports	U.S. News	World News	Total
Correct	3020	690	4860	1330	1730	11630
Incorrect	750	60	370	1100	590	2870
Accuracy	0.801	0.920	0.929	0.547	0.746	0.802

What’s Missing? Context!

You shall know a word by the company it keeps. (Firth 1957)

Article A (World News):

President Bush said he was trying to convince President Vladimir Putin of Russia that cooperation was “in Russia’s security interests,” even though Defense Secretary Robert M. Gates did not win Mr. Putin’s support during a trip to Moscow last week.

“Bush Steps Up Effort to Persuade Putin on Missile Defense Plan”, NYTimes, 1 May 2007.

Article B (U.S. News):

President Bush began his day yesterday at dawn on the golf course. He began Saturday on the golf course, too. A weekend earlier, the president played two rounds of 18 holes on the course at Andrews Air Force Base just outside Washington.

“White House Letter; Bush Makes Quick Work of Relaxing”, NYTimes, 5 August 2002.

Basic Context Engineering

What exactly do we mean when we say “context”?
Different training objectives give different types of contexts!
Example: Closest vectors to Turing, trained on Wikipedia in two ways:

Example from Uber AI Labs’ amazing Piero Molina
	Dependency Tree	5-Word Window
Same category \(\rightarrow\)	Pauling	nondeterministic	\(\leftarrow\) Topically related
(Paradigmatic)	Hotelling	computability	(Syntagmatic)
	Lessing	deterministic
	Hamming	finite-state

Pretentious But Helpful Terms!

Paradigmatic \(\approx\) Grammatical “substitutes”, drawn from same category/type
Syntagmatic \(\approx\) semantic relationships between words in a sentence

*(Both terms from Saussure (1916))*
	Paradigmatic ↓		Paradigmatic ↓
	`apples`	← Syntagmatic →	`sweet`
	`oranges`	← Syntagmatic →	`sour`
These	\(\underline{\hspace{32mm}}\)	taste very	\(\underline{\hspace{32mm}}\)	!

How Do We Actually Get These Vectors?

High-level goal: Retain information about word-context relationships while reducing the \(M\)-dimensional representations of each word down to 3 dimensions.
Low-level goal: Generate rank-\(K\) matrix \(\mathbf{W}\) which best approximates distances between words in \(M\)-dimensional space (rows in \(\mathbf{X}\))