When Accuracy Is Non-Negotiable

RAG-generated responses from your knowledge base work great for most questions. The AI reads your content, understands the question, and generates a relevant answer. But "relevant" and "exactly right" aren't the same thing. When a customer asks "how much does your premium plan cost?", a RAG response might say "Our premium plan starts at around ₱2,500 per month" — close to the real price but not exact, paraphrased from a paragraph that mentioned the price in passing. That's fine for many questions but dangerous for pricing, legal terms, and policies.

Q&A pairs solve this by creating an override layer. You define the exact question and the exact answer. When a customer asks something that semantically matches your question, the agent skips generation entirely and uses your verbatim answer. No paraphrasing, no interpretation, no "roughly around" — the exact words you wrote.

How Q&A Pairs Actually Work

Under the hood, Q&A pairs are stored in a separate namespace from general knowledge base content. When a customer asks a question, the retrieval system searches both the general knowledge base and the Q&A namespace simultaneously. If a Q&A pair matches above a confidence threshold, it takes priority — the agent uses your answer instead of generating one from general content.

The matching is semantic, not keyword-based. Your Q&A pair question might be "What is your return policy?" and a customer might ask "how do I return something I bought?" or "can I get a refund?" or "pwede ba mag-return?" — the system recognizes that all of these are asking the same thing and serves your answer. You don't need to anticipate every possible phrasing (though adding variations to your question improves match confidence).

What Deserves a Q&A Pair

Not everything needs one. Q&A pairs are precision tools, and using them for everything defeats the purpose of having a generative AI agent. Reserve them for content where the exact wording matters:

Pricing. "How much does [product/service] cost?" deserves an exact answer with all conditions, tiers, and fine print. "Our consultation fee is ₱1,500 for the initial visit and ₱800 for follow-ups. Package of 5 sessions available for ₱5,000." No room for AI interpretation.

Operating hours. "When are you open?" needs your exact schedule, including exceptions. "We're open Monday to Saturday, 9:00 AM to 6:00 PM. Closed on Sundays and holidays. Extended hours until 8:00 PM every Friday."

Policies. Return policies, cancellation terms, warranty conditions, payment terms — anything contractual. These often have specific conditions and exceptions that the AI might miss or oversimplify if generating from general content.

Contact and location. Your exact address (including landmarks for Grab/taxi directions), phone numbers, email addresses, social media handles. These are facts that must be precisely right — a wrong phone number is worse than no phone number.

Compliance-critical content. If you're in healthcare, finance, legal services, or any regulated industry, certain responses have legal requirements. A Q&A pair ensures the exact approved language is used every time.

Writing Effective Q&A Pairs

The question side should match how customers actually ask, not how you'd write it in documentation. Don't write "What are the terms and conditions of the merchandise return program?" when customers actually ask "can I return this?" Use natural language, and if possible, check your conversation history for the exact phrasing customers use.

The answer side should be complete but not encyclopedic. Include everything a customer needs to know to act on the information. If your return policy is "7 days, with receipt, unworn," don't just say "7 days" — include the conditions. But don't paste your entire Terms of Service either. The answer should be conversational, as if a helpful employee were explaining it face-to-face.

Include variations when possible. Most Q&A systems let you add alternative phrasings of the question. "What time do you close?", "What are your hours?", "Are you open on Sunday?", and "Until what time are you open?" can all be variations of the same Q&A pair, improving match accuracy.

The Practical Workflow

Start by looking at your analytics (or your chat history if you're new to AlonChat). Identify the 20 most frequently asked questions. Write Q&A pairs for each one. Test them in the playground — ask the question in different ways and verify the Q&A answer is consistently served.

Then set a quarterly review reminder. Businesses change: prices update, hours adjust, policies evolve, new products launch. Your Q&A pairs need to reflect reality. Outdated Q&A pairs are worse than no Q&A pairs because they serve wrong information with high confidence.

Twenty well-maintained Q&A pairs will handle a surprising percentage of your customer conversations with perfect accuracy. The rest can be handled by RAG from your general knowledge base. This combination — precision where it matters, generation where flexibility is fine — is what makes an AI agent genuinely reliable.

Training with Q&A Pairs: Get Perfect Answers Every Time

AlonChat Team

Ready to Build Your Own Chatbot?