Quiz: Embeddings Deep Dive
Test your knowledge about embeddings, embedding models, and selection criteria.
30 questions - Click on each question to reveal the correct answer.
Question 1
What are embeddings in the context of NLP?
a) HTML elements embedded in pages
b) Dense vector representations of text
c) Database indexes
d) Encryption keys
Show Answer
Correct Answer: b
Embeddings are dense vector representations of text that capture semantic meaning, allowing computers to compare text by meaning.
Question 2
What does cosine similarity measure?
a) The length of vectors
b) The angle between two vectors
c) The color of text
d) The file size
Show Answer
Correct Answer: b
Cosine similarity measures the cosine of the angle between two vectors, indicating how similar their directions are (and thus semantic similarity).
Question 3
What is the typical dimension range for modern embedding models?
a) 1-10
b) 50-100
c) 384-3072
d) 10000+
Show Answer
Correct Answer: c
Modern embedding models typically produce vectors with 384 to 3072 dimensions.
Question 4
Which OpenAI model offers the best embedding quality?
a) text-embedding-ada-002
b) text-embedding-3-small
c) text-embedding-3-large
d) GPT-4
Show Answer
Correct Answer: c
text-embedding-3-large offers the highest quality embeddings from OpenAI with 3072 dimensions.
Question 5
What does MTEB stand for?
a) Machine Text Embedding Benchmark
b) Massive Text Embedding Benchmark
c) Multi-Task Embedding Base
d) Model Training Evaluation Benchmark
Show Answer
Correct Answer: b
MTEB stands for Massive Text Embedding Benchmark, the standard benchmark for comparing embedding models.
Question 6
What is the cosine similarity range?
a) 0 to 1
b) -1 to 1
c) 0 to 100
d) -100 to 100
Show Answer
Correct Answer: b
Cosine similarity ranges from -1 (opposite) to 1 (identical), with 0 meaning orthogonal/unrelated.
Question 7
Which is NOT a benefit of self-hosting embedding models?
a) No per-token cost
b) Lower latency
c) Automatic updates
d) Data stays internal
Show Answer
Correct Answer: c
Self-hosted models require manual updates. API services provide automatic updates to the latest models.
Question 8
What special feature does Cohere embed-v3 offer?
a) Free unlimited use
b) Different input types for documents vs queries
c) 10000 dimensions
d) Built-in translation
Show Answer
Correct Answer: b
Cohere embed-v3 has different input_type options (search_document vs search_query) optimized for their respective uses.
Question 9
How much does text-embedding-3-small cost per million tokens?
a) $0.002
b) $0.02
c) $0.20
d) $2.00
Show Answer
Correct Answer: b
text-embedding-3-small costs $0.02 per million tokens, making it very cost-effective.
Question 10
Which open-source model family is from BAAI?
a) Sentence Transformers
b) E5
c) BGE
d) BERT
Show Answer
Correct Answer: c
BGE (Beijing Academy of Artificial Intelligence General Embedding) models are from BAAI.
Question 11
What happens to semantically similar texts when embedded?
a) They produce very different vectors
b) They produce similar vectors (close in space)
c) They produce identical vectors
d) They cannot be embedded
Show Answer
Correct Answer: b
Semantically similar texts produce vectors that are close to each other in the embedding space.
Question 12
What is the max token limit for OpenAI embedding models?
a) 512
b) 2048
c) 8191
d) 32000
Show Answer
Correct Answer: c
OpenAI's text-embedding-3 models support up to 8191 tokens per input.
Question 13
Why might you reduce embedding dimensions?
a) To improve quality
b) To reduce storage and improve search speed
c) To add more languages
d) To encrypt the vectors
Show Answer
Correct Answer: b
Reducing dimensions decreases storage requirements and speeds up similarity searches, with acceptable quality trade-off.
Question 14
Which library provides Sentence Transformers models?
a) TensorFlow
b) PyTorch only
c) HuggingFace / sentence-transformers
d) OpenAI
Show Answer
Correct Answer: c
Sentence Transformers are available through the sentence-transformers library and HuggingFace.
Question 15
What is a key advantage of API-based embeddings?
a) Free usage
b) No internet required
c) No infrastructure maintenance
d) Unlimited dimensions
Show Answer
Correct Answer: c
API-based embeddings don't require infrastructure setup or maintenance - you just call the API.
Question 16
What prefix does BGE recommend for queries?
a) No prefix needed
b) "query: "
c) "Represent this sentence for searching: "
d) "Question: "
Show Answer
Correct Answer: c
BGE models recommend using instruction prefixes like "Represent this sentence for searching:" for best results.
Question 17
How many dimensions does all-MiniLM-L6-v2 produce?
a) 128
b) 384
c) 768
d) 1024
Show Answer
Correct Answer: b
all-MiniLM-L6-v2 produces 384-dimensional embeddings, making it compact and fast.
Question 18
What is the main trade-off with higher dimensional embeddings?
a) Lower quality
b) Higher storage and slower search
c) Fewer languages supported
d) Less accurate
Show Answer
Correct Answer: b
Higher dimensions provide better quality but require more storage and slower similarity searches.
Question 19
For a startup MVP, which embedding model is recommended?
a) Custom trained model
b) text-embedding-3-large
c) text-embedding-3-small
d) Self-hosted BGE
Show Answer
Correct Answer: c
text-embedding-3-small offers a good balance of quality and cost for startups and MVPs.
Question 20
What is batch embedding useful for?
a) Reducing quality
b) Processing many texts efficiently
c) Encrypting data
d) Reducing dimensions
Show Answer
Correct Answer: b
Batch embedding processes multiple texts in a single API call, reducing overhead and improving efficiency.
Question 21
Which metric is preferred when vectors are normalized?
a) Euclidean distance only
b) Dot product (equivalent to cosine)
c) Manhattan distance
d) Hamming distance
Show Answer
Correct Answer: b
For normalized vectors, dot product is equivalent to cosine similarity and can be faster to compute.
Question 22
What does E5 stand for?
a) Embedding 5
b) EmbEddings from bidirEctional Encoder rEpresentations
c) Enterprise Embedding 5
d) Efficient Embedding Engine 5
Show Answer
Correct Answer: b
E5 stands for EmbEddings from bidirEctional Encoder rEpresentations (Microsoft).
Question 23
When should you consider self-hosting embeddings?
a) Never
b) When processing >10M documents monthly
c) For prototyping only
d) When quality doesn't matter
Show Answer
Correct Answer: b
Self-hosting becomes cost-effective at high volumes (>10M documents) or when data privacy is critical.
Question 24
What does multilingual embedding support enable?
a) Translation
b) Cross-lingual semantic search
c) Grammar correction
d) Text generation
Show Answer
Correct Answer: b
Multilingual embeddings enable searching across languages - a query in English can find relevant French documents.
Question 25
What is typical API latency for embedding generation?
a) 1-5ms
b) 50-150ms
c) 1-5 seconds
d) 30+ seconds
Show Answer
Correct Answer: b
API-based embeddings typically have 50-150ms latency, compared to 10-30ms for self-hosted GPU inference.
Question 26
Which factor is MOST important for real-time applications?
a) MTEB score
b) Latency
c) Dimensions
d) Price
Show Answer
Correct Answer: b
For real-time applications, latency is the most critical factor - users expect quick responses.
Question 27
How can text-embedding-3-large dimensions be reduced?
a) It cannot be reduced
b) Using the dimensions parameter in the API
c) Post-processing only
d) Retraining required
Show Answer
Correct Answer: b
text-embedding-3-large supports a dimensions parameter (256, 1024, etc.) to reduce output size.
Question 28
What happens if you use different embedding models for indexing and querying?
a) Works fine
b) Retrieval will fail or be poor
c) Results are doubled
d) Speed improves
Show Answer
Correct Answer: b
You must use the same embedding model for indexing and querying - different models produce incompatible vector spaces.
Question 29
Which is the best model for edge/offline deployment?
a) text-embedding-3-large
b) Cohere embed-v3
c) all-MiniLM-L6-v2
d) GPT-4
Show Answer
Correct Answer: c
all-MiniLM-L6-v2 is compact (384 dims, ~80MB) and fast, ideal for edge and offline deployments.
Question 30
What is the primary purpose of embeddings in RAG?
a) Text generation
b) Enabling semantic similarity search
c) Data compression
d) Encryption
Show Answer
Correct Answer: b
Embeddings enable semantic similarity search - finding documents by meaning rather than exact keywords.
Results
| Score | Level |
|---|---|
| 27-30 | Excellent - Strong understanding of embeddings |
| 21-26 | Good - Some review recommended |
| 15-20 | Average - Review the chapters again |
| Below 15 | Needs comprehensive review |