Embedding Quality Metrics

The Problem

Cannot measure embedding quality, making it impossible to know if embeddings accurately capture document semantics or compare embedding models.

Symptoms

• ❌ Don't know if embeddings are good

• ❌ Cannot compare embedding models

• ❌ No metrics for domain adaptation quality

• ❌ Embedding drift undetected

• ❌ Cannot validate fine-tuning improvements

Real-World Example

Company considers switching:
→ Current: OpenAI text-embedding-ada-002
→ Candidate: Cohere embed-english-v3.0

Questions:
→ Which performs better for our domain?
→ How much better (quantify)?
→ Worth the migration cost?

No metrics:
→ Cannot answer objectively
→ Guessing based on vibes

---

Deep Technical Analysis

Intrinsic Metrics

Embedding Dimensionality:

Model A: 768 dimensions
Model B: 1536 dimensions

Higher dimensions:
→ More expressive (can capture nuance)
→ But: Slower search, more storage
→ Not always better

Need: Quality per dimension

Intra-Cluster vs Inter-Cluster Distance:

Good embeddings:
→ Similar docs close together (high intra-cluster similarity)
→ Dissimilar docs far apart (low inter-cluster similarity)

Measure:
→ Silhouette score
→ Davies-Bouldin index

Higher = better embedding quality

Extrinsic Metrics (Task-Based)

Retrieval Accuracy:

Gold standard test set:
→ Query: "API authentication"
→ Relevant docs: [doc_5, doc_12, doc_23]

Measure:
→ Precision@K: Are top-K results relevant?
→ Recall@K: Did we find all relevant docs?
→ NDCG: Ranking quality

Compare models on same test set

Example Evaluation:

Model A (OpenAI):
→ Precision@5: 0.85
→ Recall@10: 0.78
→ NDCG@10: 0.82

Model B (Cohere):
→ Precision@5: 0.88 (+3pp)
→ Recall@10: 0.82 (+4pp)
→ NDCG@10: 0.86 (+4pp)

Model B objectively better

Domain Adaptation Testing

Out-of-Domain Detection:

Test on domain-specific terms:
→ Your jargon: "Redwood Strategy", "Cedar mode"

Generic model:
→ May not understand these terms
→ Low similarity to related concepts

Domain-adapted model:
→ Should recognize: "Redwood" ~ "RAG approach"
→ Higher semantic accuracy

Semantic Similarity Tests:

Pairs that SHOULD be similar:
→ ("login", "authentication") = High similarity expected
→ ("API key", "access token") = High similarity expected

Pairs that SHOULD be dissimilar:
→ ("login", "pricing") = Low similarity expected

Measure:
→ % correct classifications
→ Higher = better embedding quality

Embedding Drift Detection

Temporal Consistency:

Track embedding quality over time:
→ Month 1: NDCG = 0.85
→ Month 6: NDCG = 0.78 (dropped)

Possible causes:
→ Knowledge base changed (new jargon)
→ User queries evolved
→ Embedding model updated

Alerts: Quality degradation

Model Version Comparison:

OpenAI updates model:
→ text-embedding-ada-002 v1
→ text-embedding-ada-002 v2 (silently updated)

Re-evaluate:
→ Did metrics change?
→ Need re-embedding?

Track model versions explicitly

Fine-Tuning Validation

Before vs After:

Base model:
→ NDCG: 0.75

After fine-tuning on your docs:
→ NDCG: 0.82 (+7pp)

Validates fine-tuning effort:
→ Quantifies improvement
→ Justifies cost

---

How to Solve

Create golden evaluation dataset (query + relevant docs) + measure Precision@K, Recall@K, NDCG@K on eval set + compare models using same metrics + track metrics over time (detect drift) + test semantic similarity on domain-specific term pairs + evaluate fine-tuned models against baseline + monitor embedding model version changes + automate eval runs on each knowledge base update. See Embedding Metrics.