What Is RAG and Why Your Support Bot Shouldn't Guess

A customer asks your chatbot whether you refund annual plans after 30 days. A guessing bot invents a generous "yes, anytime" policy. A RAG (Retrieval-Augmented Generation) bot searches your billing page first, finds "annual plans are non-refundable after 14 days," and answers accurately—with a link to the source.

Customers can tell when a bot makes things up. One wrong refund promise costs more than a year of chat software. RAG retrieves facts from your indexed content, then generates a natural-language answer grounded in those facts—not from memory of the whole internet.

How RAG works in plain English

1. A visitor asks a question in natural language.
2. The system searches indexed website pages and uploaded files (PDFs, docs).
3. Relevant snippets are passed to the model as context.
4. The model writes an answer constrained by that context.
5. If nothing relevant is found, the bot should escalate—not guess.

Before and after: the same question

Why rule-based bots failed—and what RAG still cannot do

Strengths and human handoff

• RAG handles paraphrases: "return window" vs "how long to send it back."
• Multi-part questions: shipping + duties + processing in one message.
• Still needs humans for account-specific lookups, billing disputes, and safety issues.

Common RAG mistakes (and fixes)

1. Stale index after pricing changes → recrawl after every launch.
2. Missing PDFs not on public HTML → upload rate cards and policy docs.
3. No escalation when confidence is low → dispatch to WhatsApp/SMS/Telegram.
4. Over-broad crawl pulling blog opinions → crawl canonical policy pages only.
5. Voice ignored on mobile → enable browser voice on the same RAG index.

How Chirps implements RAG

Crawl your domain into a vector index. Upload PDFs for off-site content. Customize instructions for tone and boundaries. When retrieval confidence is low or the user asks for a human, dispatch fires to WhatsApp, SMS, Telegram, or Discord with the full transcript. Recrawl after pricing, shipping, or terms updates—and monthly for active catalogs.

RAG testing checklist

1. Ask the ten most common tickets—verbatim and paraphrased.
2. Ask a recently changed policy—confirm recrawl picked it up.
3. Ask something not in docs—bot should IDK or escalate, not invent.
4. Ask a trick question with a false premise ("lifetime refunds?").
5. Test mobile browser voice; verify GDPR consent for EU traffic.
6. Trigger escalation; confirm dispatch arrives with transcript context.