Embedding search in practice

Embedding search maps an input into a vector and retrieves nearby vectors. The formula is simple; the production contract is not.

A vector has an identity

Dimension alone does not define compatibility. Two 512-dimensional embeddings from different models are normally in different vector spaces and cannot be compared meaningfully.

A usable vector contract includes:

• model and exact version;
• preprocessing steps;
• output dimension;
• normalization policy;
• similarity function;
• entity represented by the vector;
• migration or re-index version.

Treat these fields like a database schema. A silent model change without re-indexing is a data compatibility bug.

Retrieval is a pipeline

Query
  │
  ├─ encode with the matching model
  │
  ├─ apply tenant, site, time, and entity filters
  │
  ├─ retrieve Top K candidates
  │
  ├─ rerank with richer signals
  │
  └─ group, paginate, and explain results

The order changes semantics. Global Top K followed by grouping is not equivalent to Top K per group. A strict filter before retrieval can improve precision and latency; the same filter can destroy recall if its source metadata is incomplete.

Cosine similarity

For non-zero vectors $x$ and $y$:

$$ \operatorname{cosine}(x,y)=\frac{x\cdot y}{\lVert x\rVert\lVert y\rVert} $$

If all vectors are normalized to unit length, cosine similarity becomes a dot product. The implementation must still use the same normalization assumption for indexed and query vectors.

A threshold is not portable across models or datasets. Calibrate it from labeled examples, not from intuition about what “0.8” should mean.

Candidate retrieval and final ranking

Approximate nearest-neighbor indexes trade exactness for speed. They should produce a candidate set large enough for the final ranking stage—not necessarily the final product order.

Reranking can combine:

• vector similarity;
• lexical match;
• freshness;
• entity confidence;
• business rules;
• a cross-encoder or multimodal model.

Keep the raw retrieval score and final ranking score distinguishable in logs. Otherwise a ranking regression looks like an indexing problem.

Evaluation by failure type

Use a labeled query set and measure the stage that failed:

Failure	Question
Representation	Are relevant items close in this vector space?
Filtering	Was the correct item removed before retrieval?
Candidate recall	Did Top K contain a relevant item?
Ranking	Was the relevant candidate ordered high enough?
Presentation	Was a correct result grouped or paginated away?

Aggregate precision is useful, but stage-level labels tell you what to change.

Production checklist

• Store model and preprocessing versions with every vector.
• Reject dimension or space mismatches explicitly.
• Define whether Top K is global or per group.
• Log filters, candidate counts, and ranking stages.
• Keep a stable evaluation set with difficult negative examples.
• Re-index intentionally when the vector contract changes.
• Compare latency and quality at realistic collection sizes.

Embedding search becomes manageable once each stage has an explicit contract and its own evidence.