Will the AI make things up or answer outside my information?
Last updated
Plans: Free: not included · Pro: included · Scale: included
Replavia’s AI is told to answer only from the facts you give it, and to write nothing when a question is not covered. By default, you also approve every reply before it posts.
Open Web knowledge →Why this matters
A wrong price or a made-up promise under your post is worse than no reply. People can screenshot it, and you then have to correct it in public. So Replavia’s AI is built to stay inside what you told it.
Where its facts come from
The AI answers comments from two places, and you control both:
- Your knowledge library. Web pages and documents you add on the Web knowledge and Documents tabs of AI Replies. Replavia saves short facts from each one. See the knowledge library.
- Your AI Agent settings. The Business name, What you offer, Tone, Do's and don'ts and FAQ the agent can answer fields. Open Lead Inbox, then the AI Agent tab.
Comment replies use those settings even when the AI DM assistant is switched off.
For each comment, Replavia picks the saved facts whose words best match the comment. Only those facts go to the AI, together with the comment.
What happens when the answer is not in your facts
The AI has a clear rule. If the comment asks something your facts do not answer, it must skip. It may not answer from general knowledge. It may not invent prices, links, stock or claims.
When it skips:
- No reply is written, so nothing reaches your review queue and nothing posts.
- The comment appears as Failed on the Recent activity tab, with a note that nothing was posted.
- You can answer it yourself on Instagram if you want to.
Retrieval, not training
The AI does not “learn” your business. It is handed your facts each time it writes a reply. Nothing you add is used to train an AI model.
If you remove a source, future replies stop using its facts straight away.
Extra checks on every reply
- Links, email addresses, @mentions and hashtags are removed from every AI reply.
- Replies are short, at most 280 characters.
- The reply is written in the same language as the comment.
- Every reply passes a brand-safety check before it can post. See safety checks.
How to give it good facts
Write facts the way you would answer a customer. Plain statements with real numbers work best.
Good examples:
- “UK delivery takes 2 to 3 working days and costs £3.99.”
- “Classes run Monday to Thursday at 7pm. A drop-in class costs $15.”
- “We do not ship outside the EU.”
Saying what you do not do is useful too. It lets the AI give a clear “no” instead of skipping.
Common mistakes:
- Linking a home page with little text. Link the page that holds the prices or policies.
- Uploading a scanned PDF. It has no text the AI can read.
- Leaving old prices on a page. Click Refresh on that source after you update it.
- Two sources that disagree. Remove the old one so the AI never sees both.
Good to know
- No AI is perfect. Strict rules make mistakes rarer, not impossible. That is why every reply waits for your approval by default. See review modes.
- A question the AI reads but cannot answer still uses one draft from your monthly allowance. See AI allowance and usage.
- Knowledge sources belong to one Instagram account. Settings you leave blank for an account use your shared settings, so give each brand its own offer and FAQ. See AI for several brands.
- AI replies are on Pro and Scale. The Free plan has no AI features.