Query Clarification

The system must first recognize that a query is underspecified or ambiguous before initiating clarification. This is typically triggered by:

• High entropy in intent classification scores (multiple intents with similar probability)
• Faceted queries where a single term maps to multiple distinct entities (e.g., 'Mercury' as planet, element, or car brand)
• Overly broad queries lacking sufficient constraints for precise retrieval (e.g., 'best practices' without domain context)
• Low confidence in the primary retrieval path, often measured via query performance prediction (QPP) metrics

Once ambiguity is detected, the system formulates a targeted clarifying question designed to maximally reduce uncertainty with minimal user friction:

• Slot-filling questions ask for a specific missing parameter: 'Did you mean Mercury the planet or Mercury the element?'
• Facet-selection prompts present distinct interpretations as tappable options rather than open-ended questions
• Constraint-elicitation questions narrow scope: 'Are you interested in Python for web development or data science?'
• Effective questions are non-redundant—they never ask for information already present in the query or conversation history

Clarification is a stateful dialogue act that temporarily suspends the primary retrieval pipeline:

• The system must persist the original query and any partial interpretations while awaiting user response
• A clarification context is pushed onto the conversation stack, distinct from the main conversational context
• Upon receiving the user's disambiguating response, the system merges the clarification with the original query to form a fully specified retrieval request
• Timeout handling is essential—if the user abandons the clarification, the system may fall back to the most probable interpretation or return diverse results

Clarification strategies differ by when they intervene in the retrieval pipeline:

• Reactive clarification occurs after an initial retrieval attempt returns low-quality or ambiguous results, using post-retrieval signals like low confidence scores
• Proactive clarification intercepts the query before retrieval, using pre-retrieval intent and entity analysis to predict ambiguity
• Hybrid approaches perform a lightweight initial retrieval to inform clarification question generation without fully executing the pipeline
• Proactive clarification reduces latency for ambiguous queries but risks over-clarifying queries that might have succeeded on first pass

In Retrieval-Augmented Generation (RAG) systems, clarification serves a dual purpose:

• It prevents hallucination by ensuring the retrieval corpus is queried with precise, unambiguous terms
• It improves citation accuracy by narrowing the retrieved document set to highly relevant sources
• Clarification questions can be generated by the same LLM used for answer synthesis, leveraging its understanding of knowledge gaps
• The retrieved documents from a clarified query provide factual grounding that is more tightly scoped and verifiable

Measuring clarification effectiveness requires specialized metrics beyond standard retrieval benchmarks:

• Clarification acceptance rate: the percentage of times users engage with and respond to a clarifying question
• Resolution efficiency: the number of turns required to reach a fully specified query (optimal is 1)
• Post-clarification precision: the improvement in retrieval precision after clarification compared to the ambiguous baseline
• Task completion rate: whether the clarified query ultimately leads to a successful user outcome
• Over-clarification rate: instances where clarification was unnecessary because the original query would have succeeded