RAG that doesn’t hallucinate

Visitors do not forgive wrong answers on support chat. Retrieval-augmented generation (RAG) helps because the model answers from your chunks instead of inventing policy from memory-but RAG alone is not a guarantee. You still need clear content, processing discipline, and fallback behavior when nothing relevant is found.

What Conversa does under the hood

When you add sources, Conversa extracts text, splits it into chunks, and creates embeddings for search. Each visitor question triggers a retrieval step: the system finds the closest chunks and passes them to the model with instructions to stay grounded. That is why uploading clean FAQs and official docs beats dumping random exports.

Write sources for retrieval

Use descriptive source names (“Returns policy - 2025”, “Pricing FAQ”). Break huge manuals into several sources so processing stays reliable. Remove outdated PDFs instead of stacking contradictions. If two documents disagree, the bot may blend them-fix the source of truth in your files first.

Tune behavior in the dashboard

Set a fallback message that honestly says when the bot does not know, rather than guessing. Adjust temperature conservatively for support use cases. Review Conversations weekly: repeated “I don’t know” on the same topic means you are missing a source, not that AI failed.

Domain and widget hygiene

Limit allowed domains to production and staging hosts you control. That reduces stray embeds and keeps test traffic separate. Pausing a bot is instant if you need to stop public answers during a content migration.

When to escalate to humans

RAG works best for repeatable questions: shipping times, account steps, product specs. Keep billing disputes, legal threats, and medical advice on a human path. Tell visitors how to reach your team in the fallback or welcome message.

Measuring quality

Track which pages drive chats, which questions get weak answers, and which sources were cited. Add a short internal note each time sales or support corrects the bot in email-that correction should become a new or updated source.

The goal is simple: answers that match your docs. Conversa gives you the pipeline; your team maintains the knowledge base like any other critical channel.