Embeddings: Why AI Knows Bank and Bank Are Different

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-ai-foundations-embeddings-final1-creators

1. What fundamental representation does an embedding convert text into?

   1. A vector of numbers representing semantic meaning
   2. A compressed image file
   3. A structured database table
   4. An executable binary file

2. In a vector space representation of embeddings, what does geometric closeness between two points indicate?

   1. The two texts share the exact same words
   2. The two texts are stored on the same server
   3. The two texts have similar semantic meaning
   4. The two texts were written by the same author

3. A user searches for 'unsubscribing' and the system returns results for 'how do I cancel'. What capability does this demonstrate?

   1. Case-sensitive matching failure
   2. Random result selection
   3. Semantic search that understands intent beyond keywords
   4. A bug in the search algorithm

4. Why can embeddings enable clustering of similar customer tickets without manually writing grouping rules?

   1. Embeddings automatically generate category labels
   2. Embeddings capture semantic meaning so similar tickets end up close in vector space
   3. Embeddings only work with structured data
   4. Embeddings require no computational resources

5. What silently breaks similarity comparisons when switching from one embedding model to another?

   1. The search interface crashes
   2. The database connection drops
   3. The vector space coordinates change meaning between models
   4. The document format changes

6. What information is typically lost when text is converted into embeddings?

   1. Whether the text is正面 or负面
   2. Exact wording and precise phrasing
   3. The length of the original document
   4. The approximate topic of the text

7. A query for 'shipping was slow' returns product reviews that don't contain those exact words but discuss late deliveries. Why does this work?

   1. The reviews are sorted alphabetically
   2. Embeddings capture semantic meaning so 'slow shipping' and 'late delivery' are close in vector space
   3. The system randomly selects reviews
   4. The system searches for partial matches within words

8. What does the lesson recommend when changing embedding models in a production system?

   1. Re-embed the entire corpus and version the embedding choice in metadata
   2. Switch models without any changes
   3. Delete the old vectors and continue using them
   4. Replace only the vectors that seem outdated

9. What type of content can embeddings help identify in large document collections?

   1. Emails with attachments
   2. Near-duplicate content that uses different wording
   3. Documents with the most images
   4. Files created on weekends

10. Why might embeddings struggle to distinguish between 'bank' (financial institution) and 'bank' (river edge)?

    1. Embedding models only work with sentences
    2. Embeddings cannot read words with multiple letters
    3. Context is needed to determine which meaning applies, and embeddings may conflate both
    4. The words are too short

11. What does the lesson identify as a reason embeddings might need periodic regeneration?

    1. Embedding models improve over time, causing drift in vector positions
    2. New regulations require different formats
    3. Older embeddings violate copyright
    4. Vectors degrade physically on storage

12. A support team uses embeddings to sort incoming tickets. One ticket doesn't fit any cluster. What does this likely represent?

    1. A duplicate of another ticket
    2. A ticket that was already processed
    3. A ticket with no text content
    4. An outlier document that doesn't match common patterns

13. What is a key advantage of using embeddings for recommendation systems compared to keyword matching?

    1. Recommendations can be based on meaning rather than exact keyword overlap
    2. Recommendations are always faster
    3. Recommendations require less data
    4. Recommendations never change

14. If you embed 100 product reviews and find the 5 nearest neighbors to 'shipping was slow', what would you expect to find?

    1. Reviews about delayed shipments or late deliveries that may not contain those exact words
    2. Only reviews containing the exact phrase 'shipping was slow'
    3. Reviews written most recently
    4. Reviews with the word 'slow' in them

15. What does the dimensionality of an embedding vector represent?

    1. The number of numerical features used to represent each text
    2. The length of the original text in words
    3. The file size of the original text
    4. The number of documents in the corpus