Knowledge Information Flashcards
7 cards from real AI practice questions. Tap to flip, then mark Knew It or Still Learning — missed cards come back until you master them.
Read the first 7 Knowledge Information flashcards as text
In information theory, what does 'entropy' measure in the context of a knowledge source?
Answer: The average number of bits needed to encode information from a source
Shannon entropy quantifies the average uncertainty or information content of a random variable, measuring how many bits are needed on average to represent outcomes.
Which technique is used to evaluate whether a language model's generated answer is faithful to a retrieved source document?
Answer: Faithfulness evaluation (e.g., using NLI models)
Faithfulness evaluation uses natural language inference (NLI) models or dedicated metrics to check whether generated claims are logically entailed by the source document.
What is 'sparse retrieval' in the context of document search for AI systems?
Answer: Retrieving documents using term-frequency-based methods like BM25
Sparse retrieval relies on lexical matching using bag-of-words representations and algorithms like TF-IDF or BM25, contrasted with dense retrieval using neural embeddings.
What challenge does 'knowledge staleness' pose for AI engineers deploying production systems?
Answer: The model's internal knowledge diverges from current real-world facts over time
Knowledge staleness means a deployed model's parametric knowledge becomes outdated as the world changes, requiring mitigation via RAG, fine-tuning updates, or knowledge editing.
Which approach best mitigates conflicting information when merging multiple knowledge sources?
Answer: Apply source authority scoring and conflict resolution rules
Conflict resolution strategies assign trust scores to sources based on authority or provenance and apply rules (e.g., majority vote, recency weighting) to resolve contradictions.
What is 'contextual compression' in RAG pipelines?
Answer: Extracting only the relevant portions of retrieved documents before passing them to the LLM
Contextual compression post-processes retrieved documents to strip irrelevant content, passing only the most pertinent excerpts to the language model to reduce noise and token cost.
In knowledge graph embedding models like TransE, how are relationships represented?
Answer: As translations in vector space such that head + relation ≈ tail
TransE models relationships as vector translations, enforcing that the embedding of the head entity plus the relation vector approximates the embedding of the tail entity.