KG Pipelines

Question 1

What are the four main stages of the KG‑RAG pipeline for QA?

Answer 1

1) Query Generation 2) Knowledge Retrieval 3) Context Construction 4) Answer Generation

Question 2

Step 1: Query Generation – tasks and common approaches

Answer 2

• Tasks: Entity linking, relation extraction, SPARQL/Cypher query construction.
• Approaches:
  • Template‑based patterns
  • Neural NL→SPARQL models
  • LLM‑based query prompting
  • Hybrid combinations

Question 3

Step 2: Knowledge Retrieval – strategies and challenges

Answer 3

• Strategies:
  • Direct SPARQL/Cypher execution
  • Vector retrieval via KG embeddings
  • Path ranking for multi‑hop
  • Graph traversal for complex queries
• Challenges: Ambiguous entities, KG incompleteness, precision vs recall, query optimization

Question 4

Step 3: Context Construction – methods and formats

Answer 4

• Methods: Subgraph extraction → linearization → enrichment → prioritization
• Formats:
  • Triplet lists
  • Natural‑language verbalizations
  • Structured JSON
  • Hybrid

Question 5

Step 4: Answer Generation – LLM techniques

Answer 5

• Techniques:
  • Prompt engineering
  • In‑context examples
  • Chain‑of‑thought prompting
  • Output verification
• Advanced:
  • KG‑specific fine‑tuning
  • ReAct reasoning
  • Self‑consistency checks
  • Explainability layers

Question 6

Designing a KG‑RAG pipeline – key decision factors

Answer 6

• Query complexity (simple fact vs multi‑hop reasoning)
• KG size & latency requirements
• Number of hops needed
• Desired response style (concise fact vs conversational explanation)

Question 7

Justifying pipeline step selection

Answer 7

• Match query generation approach to NL ambiguity and data availability
• Choose retrieval method based on KG structure, completeness, and scale
• Format context to balance informativeness and prompt length for the LLM
• Apply answer generation techniques to ensure correctness and clarity

Question 8

Potential privacy issues in KG‑RAG pipelines

Answer 8

• Exposure of sensitive entities or relationships
• Leakage of user queries via logs or prompts
• Inference of personal data through embeddings
  Mitigations: Anonymization, access control, minimal context

Question 9

Potential quality issues in KG‑RAG pipelines

Answer 9

• KG incompleteness or stale data
• Retrieval precision vs recall trade‑offs
• LLM hallucinations despite correct context
• Context truncation or snippet bias
  Mitigations: KG curation, hybrid retrieval, output verification