AI Knowledge Information 5 — Questions and Answers
Question 1: What is the primary difference between 'parametric knowledge' and 'non-parametric knowledge' in AI systems?
- Parametric knowledge is stored in model weights; non-parametric knowledge is stored in external databases (Correct answer)
- Parametric knowledge requires more GPU memory to query
- Non-parametric knowledge is always more accurate than parametric knowledge
- Parametric knowledge uses symbolic rules; non-parametric uses neural networks
Correct answer: Parametric knowledge is stored in model weights; non-parametric knowledge is stored in external databases
Parametric knowledge is encoded in a model's learned weights during training, while non-parametric knowledge is stored externally (e.g., in a vector DB or document store) and retrieved at inference time.
Question 2: Which SPARQL query keyword is used to retrieve triples from an RDF knowledge graph?
- FETCH
- SELECT with WHERE patterns (Correct answer)
- QUERY FROM
- GET TRIPLES
Correct answer: SELECT with WHERE patterns
SPARQL uses SELECT queries with WHERE clauses containing triple patterns (subject, predicate, object) to match and retrieve data from RDF graphs.
Question 3: What is 'multi-hop reasoning' in knowledge graphs and QA systems?
- Running the same query multiple times for consistency checks
- Answering a question by chaining multiple inference steps across graph edges (Correct answer)
- Parallelizing knowledge retrieval across multiple servers
- Splitting long documents into multiple chunks before indexing
Correct answer: Answering a question by chaining multiple inference steps across graph edges
Multi-hop reasoning requires traversing multiple relationships in a knowledge graph or making several logical inferences to arrive at an answer that cannot be found in a single step.
Question 4: Why might an AI engineer choose hybrid retrieval (sparse + dense) over pure dense retrieval in a production RAG system?
- Dense retrieval is always slower than sparse retrieval
- Hybrid retrieval combines lexical exactness with semantic understanding, improving coverage (Correct answer)
- Sparse retrieval supports more query languages natively
- Pure dense retrieval requires labeled training data for every query type
Correct answer: Hybrid retrieval combines lexical exactness with semantic understanding, improving coverage
Hybrid retrieval leverages BM25's strength in exact keyword matching alongside dense embeddings' semantic similarity, reducing failure cases where one method alone would miss relevant documents.
Question 5: What is the role of 'reranking' in a two-stage retrieval pipeline?
- Replacing the first-stage retriever entirely with a more powerful model
- Re-scoring a small candidate set using a higher-capacity model to improve final ranking quality (Correct answer)
- Randomly shuffling retrieved documents to reduce bias
- Filtering documents that exceed the context window length
Correct answer: Re-scoring a small candidate set using a higher-capacity model to improve final ranking quality
A reranker (e.g., a cross-encoder) takes the top-k candidates from a fast first-stage retriever and applies a more computationally expensive relevance model to produce a higher-quality final ranking.
Question 6: Which approach is most effective for detecting and reducing 'factual drift' in long-context language model outputs?
- Increasing the model's temperature setting
- Applying self-consistency checks or chain-of-thought verification against source documents (Correct answer)
- Reducing the context window to force shorter responses
- Using greedy decoding exclusively to avoid sampling variance
Correct answer: Applying self-consistency checks or chain-of-thought verification against source documents
Self-consistency methods and chain-of-thought prompting that explicitly cites source passages help detect when generated claims diverge from grounded facts as output length grows.
Question 7: In the context of knowledge management for AI systems, what does 'provenance tracking' enable?
- Faster indexing of new documents into the knowledge base
- Tracing every fact back to its original source for auditability and trust assessment (Correct answer)
- Automatically generating citations for model outputs
- Compressing knowledge graphs by removing redundant triples
Correct answer: Tracing every fact back to its original source for auditability and trust assessment
Provenance tracking records the origin, lineage, and transformation history of each piece of knowledge, enabling auditors and systems to assess source reliability and update or retract facts when sources change.
AI Knowledge Information 5
What is the primary difference between 'parametric knowledge' and 'non-parametric knowledge' in AI systems?