# Response quality

> How Bookbag produces accurate answers: retrieval-augmented generation, citations, Q&A priority, and the controls you have over grounding and fallbacks.

Bookbag is built around one idea: an answer you can trust is an answer grounded in your data. This page explains how that grounding works and the levers you control.

## Retrieval-augmented generation

When a customer asks something, Bookbag doesn't send the question straight to a language model. It first **retrieves** the most relevant chunks from your indexed sources, then asks the model to answer *using only those chunks*. The model composes a natural reply, but the facts come from your data.

1. **Embed** — Every source is split into chunks and embedded into a vector index when you train it.
2. **Retrieve** — The customer's question is embedded and matched against the index to find the most relevant chunks.
3. **Ground** — Those chunks are inserted into the prompt as the only allowed source of truth.
4. **Cite** — The reply includes citations back to the sources used, so answers are auditable.

## Citations

Every grounded answer carries citations to the chunks it used. In the playground you can expand them to verify the agent is pulling from the right place — and on the widget, customers can see where an answer came from. Citations are the fastest way to debug a wrong answer: if the cited chunk is wrong or missing, fix the source.

## Q&A priority

Q&A pairs are treated as authoritative. When a question closely matches a Q&A pair, that exact answer is returned, short-circuiting paraphrase. Use Q&A for the answers that must be exact.

> **TIP:** A handful of well-chosen Q&A pairs (refund window, shipping times, warranty terms) eliminates the most damaging category of mistakes.

## Fallbacks and "I don't know"

When retrieval finds nothing relevant, a well-configured agent says so and offers to connect the customer with a human, rather than guessing. You set the fallback message in the agent's settings, and you can pair it with the [Escalate to a human](/docs/actions/escalate-to-human) action so the conversation lands in your help desk.

## Measuring quality

Two tools tell you how you're doing:

- **Analytics** — resolution rate, conversation volume, and escalation trends over time.
- **Support Audit** — paste real transcripts and get a scored report on accuracy, hallucination rate, and resolution, judged by an LLM.

## What's next

- [Support Audit](/docs/audit/overview) — Score your existing support for quality and resolution.
- [Analytics](/docs/agents/analytics) — Track resolution and volume.
- [Best practices](/docs/getting-started/best-practices) — Techniques for consistently accurate answers.