AI tools: vector databases without the hype

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-tools-vector-database-fundamentals-r7a1-creators

1. What is the fundamental function of a vector database?

   1. It stores original documents and their metadata
   2. It performs nearest-neighbor search over embedded data
   3. It generates embeddings from raw text
   4. It trains machine learning models on your data

2. Which factor has the greatest impact on recall quality in a vector search system?

   1. The hardware infrastructure (GPU vs CPU)
   2. The specific vector database product chosen
   3. The embedding model and chunking strategy
   4. The amount of metadata stored

3. Before adopting a dedicated vector database, what does the lesson recommend trying first?

   1. Amazon DynamoDB with vector support
   2. MongoDB with Atlas search
   3. PostgreSQL with the pgvector extension
   4. Elasticsearch with vector plugins

4. What does 'recall@k' measure?

   1. The percentage of relevant items found within the top k results
   2. The number of vectors the system can store
   3. The time it takes to return k results
   4. The similarity score threshold used for filtering

5. Why can't a vector database improve recall when the underlying embeddings are low quality?

   1. Vector databases require perfect embeddings to function
   2. The search algorithm automatically corrects errors in embeddings
   3. Vector databases have strict size limits on embeddings
   4. The database can only work with what the embeddings represent — garbage in, garbage out

6. What problem occurs when key sentences are split across multiple chunks using fixed token count chunking?

   1. The chunks become too large to process efficiently
   2. The embedding model ignores short chunks
   3. The database automatically merges small chunks
   4. Those sentences become unfindable because context is lost

7. What chunking approach does the lesson recommend for better recall?

   1. Chunk by sentence boundaries only
   2. Fixed token count (256 tokens each)
   3. Chunk by semantic units like paragraphs or sections
   4. Fixed character count (500 characters each)

8. What is semantic similarity search?

   1. Finding documents created in the same time period
   2. Finding documents that share the same authors
   3. Finding results based on meaning rather than exact wording
   4. Finding exact keyword matches in text

9. A relevant document fails to appear in search results. What is a likely cause the vector database cannot itself explain?

   1. The document was added after the index was built
   2. The query used the wrong HTTP method
   3. The embedding model failed to capture the document's relevance to that query
   4. The database ran out of storage space

10. What capability do vector databases offer when combined with metadata filters?

    1. They can narrow results by both semantic similarity and structured criteria
    2. They can guarantee perfect recall
    3. They can convert structured filters into embeddings
    4. They can automatically generate metadata

11. What is the primary purpose of converting text into embeddings?

    1. To format text for display on websites
    2. To remove personally identifiable information
    3. To compress text for storage efficiency
    4. To enable mathematical similarity comparisons

12. How do modern vector databases handle scaling to millions of vectors?

    1. They require manual partitioning by the user
    2. They use specialized index structures for approximate nearest-neighbor search
    3. They switch to exact matching at scale
    4. They automatically delete older vectors

13. What is the benefit of overlapping chunks when chunking documents for vector search?

    1. It reduces storage requirements
    2. It eliminates the need for embedding models
    3. It automatically generates better embeddings
    4. It ensures key ideas aren't split across non-matching chunks

14. Why is the embedding model considered more important than the database choice?

    1. The database only indexes what the embedding model creates — poor embeddings can't be rescued by database features
    2. Embedding models require less maintenance
    3. Databases have no impact on results
    4. Embedding models are more expensive to run

15. What does 'nearest-neighbor search' mean in the context of vector databases?

    1. Finding the closest document by word count
    2. Finding documents created by the same author
    3. Finding the most recently added documents
    4. Finding vectors that are mathematically closest in multi-dimensional space