This post is a part of the DP-800: Developing AI-Enabled Database Solutions Exam Prep Hub.
This topic falls under these sections:
Implement AI capabilities in database solutions (25–30%)
   --> Design and implement intelligent search
      --> Evaluate performance of vector and hybrid search

Note that there are 10 practice questions (with answers) at the end of each section to help you solidify your knowledge of the material. Also, there are 4 practice tests with 30 questions each available from the hub's main page below the exam topics section.

Introduction

Evaluating the performance of vector and hybrid search solutions is a critical responsibility when developing AI-enabled database applications. While implementing vector search is important, ensuring that the search solution consistently returns accurate, relevant, fast, and scalable results is equally essential. Database developers must understand how to measure search quality, optimize retrieval performance, balance latency with accuracy, and monitor search systems over time.

This knowledge is especially important for applications such as:

• Retrieval-Augmented Generation (RAG)
• Enterprise knowledge search
• AI-powered chatbots
• Document retrieval
• Recommendation systems
• Intelligent search applications

Why Performance Evaluation Matters

Unlike traditional SQL queries that typically return deterministic results, vector and hybrid search systems retrieve documents based on statistical similarity.

This means there is always a balance between:

• Search speed
• Search accuracy
• Resource consumption
• Scalability

A search system that responds instantly but returns irrelevant documents is not useful.

Likewise, a system that returns perfect results but requires several seconds per query may not satisfy user expectations.

The goal is to optimize the entire search experience.

Key Performance Metrics

Several metrics are commonly used to evaluate vector and hybrid search.

Query Latency

Latency measures how long a search takes to return results.

Example:

User Query
↓
120 ms
↓
Results Returned

Lower latency improves user experience.

Typical enterprise AI search systems aim for response times measured in milliseconds.

Factors affecting latency include:

• Index type
• Dataset size
• Hardware resources
• Number of search algorithms executed
• Network latency
• Number of retrieved documents

Throughput

Throughput measures the number of search requests a system can process within a given time.

Examples:

• Searches per second
• Queries per minute

Higher throughput enables more concurrent users.

Throughput depends on:

• CPU
• Memory
• Index efficiency
• Parallel processing
• Database architecture

Recall

Recall measures how many relevant documents are successfully retrieved.

Example:

Relevant documents:

A
B
C
D
E

Returned documents:

A
B
C
X
Y

Recall:

3 / 5 = 60%

Higher recall generally improves RAG quality because more relevant information is available to the large language model.

Precision

Precision measures how many returned documents are actually relevant.

Example:

Returned:

A
B
C
X
Y

Relevant:

A
B
C

Precision:

3 / 5 = 60%

High precision reduces irrelevant search results.

F1 Score

The F1 Score combines precision and recall into a single metric.

It is especially useful when both false positives and false negatives matter.

Higher F1 scores indicate a better overall balance between retrieving relevant documents and avoiding irrelevant ones.

Mean Reciprocal Rank (MRR)

MRR measures how highly the first relevant result appears in the ranked list.

Example:

Relevant document positions:

Higher MRR indicates users find relevant information more quickly.

MRR is commonly used when evaluating question-answering systems and RAG applications.

Normalized Discounted Cumulative Gain (NDCG)

NDCG measures:

• Ranking quality
• Position of relevant documents
• Graded relevance

Unlike recall, NDCG rewards placing the most relevant documents near the top.

This is especially important because users rarely read beyond the first few search results.

Evaluating Vector Search

When evaluating vector search, developers typically measure:

• Recall
• Precision
• Latency
• Index build time
• Memory usage
• Storage requirements

Index Performance

Questions include:

• How quickly are searches completed?
• How much memory does the index require?
• How long does index creation take?
• How efficiently are inserts handled?

Search Quality

Evaluate:

• Are similar documents retrieved?
• Are unrelated documents excluded?
• Are synonyms recognized?
• Does semantic similarity match user expectations?

Evaluating Hybrid Search

Hybrid search combines:

• Full-text search
• Vector search
• Metadata filtering
• Ranking algorithms such as Reciprocal Rank Fusion (RRF)
• Optional semantic reranking

Because more components participate, additional evaluation is necessary.

Ranking Quality

Developers evaluate whether:

• Exact matches appear near the top.
• Semantically relevant documents are included.
• Duplicate results are minimized.
• Ranking is consistent.

Hybrid Relevance

Example query:

“Reduce Azure costs”

Good hybrid results may include:

• Azure cost optimization
• Cloud spending reduction
• Budget management
• Reserved capacity guidance

Poor hybrid results may include unrelated Azure topics.

Measuring Retrieval Quality

Many organizations create benchmark datasets.

Example:

Question:

“How do I configure VPN access?”

Expected documents:

• VPN Setup Guide
• Remote Access Policy
• Authentication Configuration

The search system is evaluated based on whether these expected documents appear in the returned results.

Human Evaluation

Automated metrics cannot evaluate every aspect of search quality.

Organizations often perform manual reviews.

Experts examine:

• Relevance
• Completeness
• Ranking quality
• Consistency

Human evaluation is particularly valuable for RAG applications.

Offline Evaluation

Offline testing uses historical datasets.

Advantages:

• Repeatable
• Safe
• Fast
• No production impact

Developers compare:

• Multiple embedding models
• Index types
• Similarity metrics
• Ranking algorithms

Online Evaluation

Online evaluation uses live users.

Common techniques include:

A/B Testing

Group A:

Current search system

Group B:

New search implementation

Metrics compared include:

• Click-through rate
• User satisfaction
• Search success
• Session completion

User Feedback

Collect feedback such as:

• Helpful
• Not Helpful

User feedback helps improve future search tuning.

Factors Affecting Vector Search Performance

Embedding Quality

Poor embeddings reduce retrieval quality regardless of index performance.

Always choose embedding models appropriate for the domain.

Similarity Metric

Common choices:

• Cosine similarity
• Dot product
• Euclidean distance

Using the wrong metric can reduce search accuracy.

Vector Index Type

Different index types provide different tradeoffs.

Candidate Set Size

Returning more candidate documents often increases recall.

However:

• Latency increases.
• More data must be reranked.
• LLM token usage increases in RAG.

Balance is important.

Metadata Filtering

Filtering improves:

• Precision
• Latency

Example:

WHERE Department = 'Finance'

Searching fewer documents reduces processing time while improving relevance.

Evaluating Hybrid Search Components

Keyword Search

Evaluate:

• Exact matches
• Phrase matching
• Synonym handling
• Technical terminology

Vector Search

Evaluate:

• Semantic understanding
• Related concepts
• Context awareness

Reciprocal Rank Fusion (RRF)

Evaluate:

• Ranking consistency
• Combined relevance
• Candidate diversity

Semantic Reranking

Evaluate:

• Final ranking quality
• User satisfaction
• Response accuracy

Common Performance Bottlenecks

Missing Vector Index

Searching every embedding significantly increases latency.

Poor Embeddings

Weak embeddings reduce semantic quality.

Excessive Candidate Retrieval

Retrieving hundreds of documents unnecessarily increases reranking and LLM processing time.

Large Embedding Dimensions

Higher-dimensional embeddings require:

• More storage
• More memory
• More computation

Frequent Index Rebuilds

Rebuilding indexes too frequently can consume unnecessary resources.

Use incremental updates where supported.

Optimization Techniques

Choose the Correct Index

Examples:

• Small datasets → Flat
• Medium datasets → HNSW
• Very large datasets → IVF or IVF + PQ

Tune Candidate Count

Retrieve only the number of documents needed.

Use Metadata Filters

Reduce unnecessary searches.

Optimize Embeddings

Select high-quality embedding models.

Use Hybrid Search

Combining lexical and semantic search generally improves relevance.

Apply Semantic Reranking

Use reranking on a limited candidate set to improve final result quality.

Performance Monitoring

Production systems should monitor:

• Average latency
• Peak latency
• Recall
• Precision
• Throughput
• Memory usage
• Index size
• Search failures
• User satisfaction
• Search abandonment rate

Monitoring enables proactive tuning as data volumes and usage patterns evolve.

Best Practices

• Benchmark search quality before deployment.
• Measure both latency and retrieval quality.
• Use benchmark datasets with known expected results.
• Combine automated metrics with human evaluation.
• Tune candidate retrieval size based on workload.
• Select the appropriate vector index for dataset size.
• Monitor production search metrics continuously.
• Refresh embeddings when source data changes significantly.
• Evaluate hybrid search using realistic business queries.
• Test changes in a staging environment before production deployment.

DP-800 Exam Tips

Remember these key points for the exam:

• Vector search performance should be evaluated using both speed and retrieval quality metrics.
• Recall measures how many relevant documents are retrieved.
• Precision measures how many returned documents are relevant.
• MRR evaluates how quickly users encounter the first relevant result.
• NDCG evaluates the quality of document ranking.
• Hybrid search should be evaluated as a complete pipeline, including keyword search, vector search, RRF, and optional semantic reranking.
• Metadata filtering improves both precision and performance.
• Human evaluation remains important because automated metrics cannot fully measure search usefulness.
• Production systems should continuously monitor latency, recall, throughput, and user satisfaction.

Practice Exam Questions

Question 1

A database developer is evaluating a vector search solution. Which metric measures the percentage of retrieved documents that are actually relevant?

A. Recall

B. Latency

C. Precision

D. Throughput

Answer: C

Explanation:
Precision measures the proportion of retrieved documents that are relevant. High precision indicates that the search results contain few irrelevant documents.

Question 2

A Retrieval-Augmented Generation (RAG) application consistently retrieves only three of the five relevant documents for most user queries.

Which performance metric is primarily affected?

A. Recall

B. Mean Reciprocal Rank (MRR)

C. Throughput

D. Query latency

Answer: A

Explanation:
Recall measures how many relevant documents are successfully retrieved. Missing relevant documents lowers the recall score.

Question 3

Which metric evaluates how quickly users encounter the first relevant search result?

A. F1 Score

B. Mean Reciprocal Rank (MRR)

C. Precision

D. Throughput

Answer: B

Explanation:
MRR evaluates the ranking position of the first relevant result, rewarding systems that place useful documents near the top of the results list.

Question 4

A search solution returns highly relevant documents, but users complain that responses take several seconds.

Which performance metric should the development team investigate first?

A. Index build time

B. Storage utilization

C. Embedding dimension

D. Query latency

Answer: D

Explanation:
Query latency measures the time required to return search results. High latency negatively impacts the user experience, even when retrieval quality is good.

Question 5

Which statement best describes hybrid search performance evaluation?

A. Only vector search accuracy needs to be measured.

B. Only keyword search latency matters.

C. Evaluation should include keyword search, vector search, ranking quality, and overall retrieval performance.

D. Performance is determined solely by embedding size.

Answer: C

Explanation:
Hybrid search combines multiple retrieval methods, so developers should evaluate the complete search pipeline rather than a single component.

Question 6

A developer increases the number of candidate documents retrieved before semantic reranking.

What is the most likely tradeoff?

A. Lower latency and reduced memory usage

B. Higher recall but increased latency and reranking costs

C. Reduced recall with faster indexing

D. Elimination of vector indexing requirements

Answer: B

Explanation:
Retrieving more candidate documents increases the likelihood of finding relevant information but also increases processing time, reranking effort, and LLM token usage.

Question 7

Why is human evaluation still valuable when assessing AI-powered search systems?

A. Automated metrics cannot fully measure user relevance and usefulness.

B. Human evaluation eliminates the need for benchmark datasets.

C. Human reviewers create vector indexes.

D. Human evaluation replaces latency testing.

Answer: A

Explanation:
While automated metrics quantify retrieval quality, human reviewers can assess contextual relevance, completeness, and overall usefulness from a user perspective.

Question 8

Which optimization technique can improve both search precision and query performance?

A. Increasing embedding dimensions indefinitely

B. Removing vector indexes

C. Using metadata filtering to narrow the search scope

D. Returning every matching document

Answer: C

Explanation:
Metadata filters reduce the number of candidate documents that must be searched, improving both relevance and performance.

Question 9

Which performance metric evaluates the overall quality of document ranking by giving more credit when highly relevant documents appear near the top of the results?

A. Recall

B. Precision

C. Throughput

D. Normalized Discounted Cumulative Gain (NDCG)

Answer: D

Explanation:
NDCG measures ranking quality by considering both document relevance and the position of documents in the ranked results, rewarding systems that place the most relevant items first.

Question 10

A development team wants to compare two different embedding models before deploying a new search solution.

Which evaluation approach is most appropriate?

A. Online A/B testing only

B. Disable benchmarking and rely on production feedback

C. Conduct repeatable offline testing using benchmark datasets with expected search results

D. Measure only CPU utilization

Answer: C

Explanation:
Offline benchmarking with known datasets enables developers to compare embedding models, similarity metrics, and indexing strategies safely and consistently before deploying changes to production.

Go to the DP-800 Exam Prep Hub main page