liveBook · Manning

Grounded answer

As explained in Effective Conversational AI

A grounded response is an answer whose claims are supported by retrieved source evidence.

A grounded answer supports its claims with retrieved evidence instead of filling gaps with plausible-sounding guesses. When evidence is missing or weak, the safer response is to acknowledge uncertainty, defer, or connect the user with a human specialist. section 6.2.2

When the evidence does not answer the question

A user asks whether a booster could worsen dysphagia associated with achalasia. If the retrieved resources do not clearly answer this, a responsible system says it could not find a clear answer and offers connection to a specialist. This avoids presenting an unsupported medical claim as fact. section 6.2.2

Evaluate a response by checking whether it:

  • Uses retrieved passages that support its claims. section 6.2.2
  • Deflects or returns an explicit fallback when retrieval finds nothing. section 6.2.2
  • Escalates to a human when generated content has poor overlap with the evidence. section 6.2.2
Table 6.1 A comparison of answers for intents and RAG
Interaction with an intent-based bot
Interaction with a RAG-integrated bot
User: What are the current COVID-19 guidelines for restaurants in New York?
PharmaBot: Current COVID-19 guidelines for restaurant dining include limited indoor capacity, mask requirements for staff and patrons when not seated, and enhanced cleaning protocols. Please check local regulations for the most up-to-date information.
User: What are the current COVID-19 guidelines for restaurants in New York?
PharmaBot: COVID-19 remains a significant concern for New Yorkers. Make sure you are up to date on the updated COVID-19 vaccines for the 2023-2024 season that offer protection against emerging variants and enhance overall immunity. Masks continue to be effective tools in mitigating the spread of COVID-19 and other respiratory viruses, reducing the transmission of droplets during various activities. To minimize the risk of contracting COVID-19, individuals are advised to wear masks in crowded indoor public settings.
Source: NYC Health

Compare how intent-based and retrieval-augmented responses handle the same changing-information question. section 6.2.2

How Effective Conversational AI covers it

How deep it goes
Explains evidence comparison, empty retrieval, fallback responses, and human escalation without detailing implementation.
Examples
Uses a medical question about achalasia and dysphagia, plus a comparison of intent-based and retrieval-augmented chatbot answers.
What you can do after
Evaluate whether an answer is supported and choose deflection, fallback wording, or specialist escalation when evidence is insufficient.
How it connects
Connects retrieval-augmented generation with large language models, semantic overlap, fallback responses, and human-agent escalation.
Cover of Effective Conversational AI
liveBookmanning.com
Read Effective Conversational AI on liveBook Get the book at manning.com

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Effective Conversational AI. The section links point to the parts of the book it draws on.