Some vector indexes return squared L2 distance because it preserves the same nearest-neighbor ranking as ordinary L2 distance while avoiding the square root calculation. For search, the ranking is usually what matters most.
Squared L2 distance is not a different idea from Euclidean distance. It is the Euclidean distance before the final square root is applied.
Short Answer
Vector indexes often use squared L2 distance because:
- it is faster to compute than ordinary L2 distance
- it keeps the same ordering of nearest results for a fixed query
- it avoids unnecessary square root operations during search
- it is convenient for indexing and optimization
- it exposes the raw distance used by the index
The important trade-off is that squared L2 scores are not on the same numeric scale as ordinary L2 scores.
L2 Distance vs Squared L2 Distance
Ordinary L2 distance, or Euclidean distance, is:
L2(a, b) = sqrt(sum((a_i - b_i)^2))Squared L2 distance is:
squared_L2(a, b) = sum((a_i - b_i)^2)The only difference is the square root.
Example
Suppose two vectors differ by 3 in one dimension and 4 in another:
a = [0, 0]
b = [3, 4]The squared L2 distance is:
3^2 + 4^2 = 9 + 16 = 25The ordinary L2 distance is:
sqrt(25) = 5Both describe the same pair of vectors, but the returned score is different.
Why Ranking Stays the Same
The square root function is monotonic for non-negative numbers. That means if one squared distance is smaller than another, its square root will also be smaller.
4 < 9
sqrt(4) < sqrt(9)
2 < 3So for nearest-neighbor ranking, the square root is not needed.
If candidate A has a smaller squared L2 distance than candidate B, candidate A will also have a smaller ordinary L2 distance than candidate B.
Why Avoid the Square Root?
Vector search may compare a query vector against many candidate vectors. Even approximate indexes still perform many distance calculations while building, updating, and searching the index.
A square root operation adds work. If the index only needs to rank candidates, it can skip that work and use squared L2 directly.
This is a small saving for one comparison, but it can matter at scale.
Why Indexes Expose the Raw Metric
Some systems expose the same distance value used internally by the index. If the configured metric is squared L2, the returned value is the squared distance.
That makes the score honest and predictable: the returned distance is exactly the metric used to rank results.
It also avoids pretending that the system used ordinary L2 when it actually used squared L2.
Lower Still Means Closer
Squared L2 is a distance metric in the search-result sense: lower values mean closer vectors.
A distance of 0 means the vectors are identical under that metric. Larger values mean the vectors are farther apart.
Do not sort squared L2 scores as if higher is better.
Do Squared L2 and L2 Always Produce the Same Top-K?
For exact comparisons against the same query and the same candidate vectors, squared L2 and ordinary L2 preserve the same ordering.
However, real vector search systems may also include approximate indexing, compression, quantization, filters, rescoring, or tie-breaking. Those features can affect final results independently of the L2-versus-squared-L2 distinction.
The basic mathematical point remains: removing the square root does not change exact L2 ordering by itself.
Thresholds Are Different
This is the most common practical mistake.
A threshold for ordinary L2 cannot be reused directly as a threshold for squared L2.
If your L2 threshold is:
L2 <= 5then the equivalent squared L2 threshold is:
squared_L2 <= 25If you switch between L2 and squared L2, update your thresholds.
Scores Look Larger
Squared L2 scores often look larger than ordinary L2 scores because the final square root has not compressed the value.
For example:
L2 = 10
squared_L2 = 100This does not mean the search got worse. It means the score is on a different scale.
Impact on RAG Systems
In a RAG pipeline, squared L2 distance can be used safely for ranking if the embedding model and index are designed for L2-style comparison.
The main care point is thresholding. If you use a distance cutoff to decide whether retrieved context is relevant enough, that cutoff must be calibrated for squared L2, not ordinary L2.
Bad thresholds can cause good chunks to be rejected or weak chunks to be included.
Impact on Score Display
If you show scores to users or downstream systems, label them clearly.
Say squared L2 distance, not just distance, when the distinction matters.
This helps avoid confusion when someone expects ordinary Euclidean distance values.
When Squared L2 Is a Good Fit
Squared L2 is a good fit when:
- the embedding model expects L2-style comparison
- the vector index supports or prefers squared L2
- you mainly need nearest-neighbor ranking
- you can calibrate thresholds on squared distances
- you want distance values that match the index’s raw metric
Common Mistakes
Common mistakes include:
- assuming squared L2 scores are ordinary L2 scores
- copying thresholds from L2 to squared L2
- thinking larger squared L2 means more similar
- comparing squared L2 scores with cosine scores
- displaying raw squared distances without labeling them
- changing the distance metric without revalidating retrieval quality
Summary
Some vector indexes return squared L2 distance because the square root is unnecessary for ranking nearest neighbors. Squared L2 preserves exact L2 ordering while reducing computation and exposing the raw metric used by the index.
The ranking intuition stays simple: lower squared L2 means closer. The practical difference is score scale. Thresholds, logs, and displayed scores must be calibrated for squared L2 rather than ordinary L2.