Knowledge Integration and Data Handling

NCP-AAI · 20 questions

• A permit Q&A assistant must ground answers in the municipal code corpus. Which retrieval design best fits authoritative civic documents that need both topical and exact-match strength?
• Semantic-only search keeps missing queries that cite exact ordinance numbers like “12.04.210.” What change should the RAG pipeline make?
• During a flash-flood emergency, the city posts updated shelter FAQs that must reach the resident agent within minutes. What retrieval-pipeline requirement matters most?
• Clerks must audit which municipal-code section grounded a permit agent’s answer. What RAG design choice supports that audit?
• A city vector index holds policies spanning decades. How should the vector database be configured for precise permit retrieval?
• Long zoning chapters retrieve as fragmented, incoherent snippets. What vector-index optimization should the team tune first?
• p95 vector retrieval latency for the 311 agent breaches the published SLO. What is an appropriate vector-database response?
• Public website copy and confidential case notes must both support agent retrieval. How should the vector database be organized?
• A municipal agent needs current permitting, tax, and GIS data for reasoning. Which approach best integrates those heterogeneous sources?
• Assessor and utility systems use different parcel identifier formats for the same property. What ETL concern must be resolved before agent use?
• Housing-intake extracts contain Social Security numbers destined for an agent-facing store. What privacy-preserving ETL step should run first?
• The nightly ETL that refreshes fee tables fails, but a release pipeline still wants to promote a fee-quoting agent. What should happen?
• Preprocessing finds two ordinance sections with the same number but conflicting effective dates. What data-quality action should occur before embedding?
• Scanned city-council minutes enter the RAG corpus with OCR garbling of names and motion text. What should preprocessing do before indexing?
• Flood-complaint tickets are rare in the city’s training and eval sets, so the agent underperforms on storm scenarios. What judicious preprocessing step helps without inventing policy?
• A municipal agent misses parcel matches because residents write “St” while the assessor store uses “Street.” What preprocessing fix best improves civic entity matching?
• During a water-main outage, the city’s agent must reason over live work-order SQL status plus unstructured email notes from field crews. Which design best enables that?
• Transit ops wants live bus locations for the resident agent while service-policy PDFs remain in the vector store. How should knowledge access be architected?
• A public-health agent gives wrong clinic hours because it reads a stale cache. What should the city mandate for hours data?
• A court clerk agent must answer both “when is the hearing?” and “what did the opinion say?” Which knowledge pattern fits?