Exams / AIP-C01 / Study guide / Choosing and generating embeddings (Titan, Cohere, Nova multimodal)
Choosing and generating embeddings (Titan, Cohere, Nova multimodal): AIP-C01 study notes and practice questions
What to know about Choosing and generating embeddings (Titan, Cohere, Nova multimodal) for the AIP-C01 exam, from the Foundation Model Integration, Data Management, and Compliance domain (31% of the exam guide), then 1 free question to test it. Part of the AIP-C01 study guide.
AIP-C01 asks which embedding model fits the language and media of a corpus, how dimension trades cost against accuracy, and how to migrate a corpus safely.
Choosing a model
Need
Choose
General English or multilingual text in a Bedrock knowledge base
Amazon Titan Text Embeddings V2
Multilingual corpus with a model built for many languages
Cohere Embed Multilingual
English-focused text
Cohere Embed English
Image and text search, or video and audio mixed with text
Amazon Nova Multimodal Embeddings
Titan Text Embeddings V2 lets you pick 256, 512 or 1024 dimensions and optionally normalize the output.
Dimension, storage and accuracy
Each vector stores one number per dimension, so a 1024-dimension index costs about four times the storage of a 256-dimension one. Smaller dimensions cut storage and search cost, and a lower dimension is often enough for well-separated content. Choose the smallest dimension that still meets a recall target on your own test questions. Titan Text Embeddings V2 and Cohere Embed can also return binary embeddings, which cut storage much further at some cost in accuracy. The vector store must support binary vectors, as OpenSearch does.
Distance metric and normalization
Normalized vectors have length 1, so cosine similarity and dot product give the same ranking. When the model or index expects normalized input, enable normalization rather than mixing raw and normalized vectors in one index.
Changing the model
Vectors from different embedding models live in different spaces and cannot be compared. Changing the embedding model means re-embedding the whole corpus and building a new index, then switching traffic over. A multimodal model still needs every image and text item re-embedded if you move to it from a text-only model.
For a large backfill, run a throttled job with retries and exponential backoff instead of firing one call per document in a tight loop. This keeps you under throttling limits and makes a failed batch easy to resume.
How to choose
Scenario wording
Choose
“Search product photos and their descriptions together”
Nova Multimodal Embeddings
“Reduce vector storage cost, accuracy still acceptable”
Smaller Titan V2 dimension
“Documents in many languages, one index”
Multilingual embedding model
“Team wants a better model for existing data”
Re-embed everything into a new index
“Vectors from the old model and new model must match”
Not possible; re-embed all content
Common traps
Mixing vectors from two models in one index and expecting good scores.
Re-embedding only new documents after a model change, which leaves old vectors in an incompatible space.
Maximum dimension by default, which raises storage and latency without a measured accuracy gain.
Text-only embeddings for image search, where the model cannot see the pixels.
Assuming a different distance metric fixes bad results. The problem is usually the model or chunking.
A travel marketplace indexes 60 million short text passages in English, Spanish, Portuguese and French for a search feature. Storage and memory for the vector index are now the largest part of the search bill.
The team can accept a small loss in retrieval quality, but the search must still return the right passages for at least 98 percent of the queries in its own benchmark set, which it can run before any change goes live.
Reindexing the whole corpus is acceptable once, during a planned maintenance window.
Which approach meets these requirements at the LOWEST index cost?
Choose one answer.
AA. Keep the English-only Cohere embedding model at 1024 dimensions for all four languages, because its high dimension count gives the strongest recall across the corpus.
BB. Use Amazon Titan Text Embeddings V2 at 1024 dimensions, the largest available size, because larger vectors give better recall at the same cost per query.
CC. Switch to Titan Text Embeddings V2 at 256 dimensions without running any benchmark, because smaller vectors cost less to store and search, and accept whatever recall results.
DD. Use Titan Text Embeddings V2 at 512 dimensions, after benchmarking it against the 1024-dimension baseline on the team's own queries, and reindex the corpus once.
Show the answer and explanation
Answer: D
A. Wrong: The English-only model is weak on Spanish, Portuguese and French passages, so recall would fall for most of the corpus. The high dimension count also keeps the index as large as the cost requirement forbids.
B. Wrong: Larger vectors raise the storage and memory that drive the bill, so the cost of the index grows with the dimension count. The claim that the cost stays the same is wrong, so this option misses the cost requirement.
C. Wrong: The smallest size does reduce index cost, but skipping the benchmark means the 98 percent recall target cannot be shown. The team needs measured quality before the change goes live.
D. Right: Halving the dimensions roughly halves the vector storage while the benchmark confirms whether recall stays at the target. Titan V2 is multilingual, and the single planned reindex matches the stated maintenance window. Learn more: Amazon Titan Text Embeddings models (opens in a new tab)
Explanation
Titan Text Embeddings V2 is multilingual and offers smaller dimension sizes, so choosing 512 dimensions cuts index cost while a benchmark confirms the recall target. The English-only model, the largest size and an unmeasured smallest size each fail a stated requirement.
Want more on Choosing and generating embeddings (Titan, Cohere, Nova multimodal)? The 5 full-length practice sets have 6 more Choosing and generating embeddings (Titan, Cohere, Nova multimodal) questions, 375 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets →
Study Choosing and generating embeddings (Titan, Cohere, Nova multimodal)