A sanitation chatbot’s fluency scores look fine while groundedness collapses on pickup-schedule facts. What should ops prioritize before promotion?
Select an answer to reveal the explanation.
Short Explanation
Pretty sentences that invent pickup days still fail residents. Fix groundedness first; fluency can wait when facts are wrong.
Full Explanation
Prioritizing groundedness recovery over fluency polish for factual civic answers is correct because groundedness measures whether generative responses stay faithful to retrieved or provided sources. When a sanitation chatbot’s fluency scores remain fine while groundedness collapses on pickup-schedule facts, residents can receive eloquent but wrong collection days, routes, or holiday exceptions. For factual municipal operations, ops must restore source-faithful answers first; polishing already-fluent prose does not fix invented schedules.
Increasing sampling temperature to strengthen groundedness of civic schedule answers fails because higher temperature only increases randomness and hallucination risk; it does not restore source-faithful collection days, routes, or holiday exceptions. Grounding pickup-schedule answers with general web search results for broader coverage fails because broad web results can inject other cities’ schedules or outdated pages; it does not ensure faithfulness to the authoritative sanitation source for collection days, routes, or holiday exceptions. Prioritizing similarity scoring over groundedness validation for civic schedule answers fails because similarity only measures surface overlap and can score eloquent but wrong pickup days highly; it cannot detect invented collection days, routes, or holiday exceptions the way source-faithfulness checks do.
Exam caveat: when fluency is healthy and groundedness collapses, the exam expects groundedness (and retrieval/citation) remediation before promotion—not fluency tuning. Factual civic domains (schedules, fees, eligibility) weight groundedness especially highly.
Operational check: sample failing pickup-schedule traces, inspect retrieved chunks and citations, repair index freshness, chunking, or prompt grounding instructions, re-run groundedness evals until they meet the sanitation threshold, and only then spend cycles on fluency or tone polish for promotion.
Prefer to study offline? Printable PDF study book + test engine ($7.50, unofficial). These questions stay free here.