AI answers

When someone sends a message through the chat widget on your site, Chattering's AI agent reads the question and responds. That response is an AI Answer. It's not a scripted reply pulled from a menu, and it's not a word-for-word copy of something from your docs. The AI reads the sources you've connected — your website, help articles, Q&A pairs, uploaded files — and writes a response that addresses the specific question, drawing from the most relevant passages it found.

Every AI Answer includes numbered citations. The visitor sees a small reference number next to any claim that came from a specific source. Clicking the number shows the exact paragraph the AI drew from. This means visitors can check what the AI said, and you can trace every answer back to its source.

How AI Answers are generated

When a visitor sends a message, the agent does a semantic search across all of the sources you've connected. Semantic search goes beyond keyword matching — it looks for meaning, so “how do I cancel?” and “can I end my subscription?” find the same sources even though none of those exact words might appear together in your docs.

The passages with the highest relevance scores get passed to a large language model, which writes a concise reply. The model is constrained to draw from the retrieved passages, so it can't invent facts that aren't in your knowledge base. If no passage is relevant enough, the agent can fall back to a transfer to your team, or acknowledge it doesn't know and suggest the visitor contact you directly.

When AI Answers appear

An AI Answer is generated the moment a visitor submits a message in the chat panel. If the question has a clear answer in your knowledge base, the reply usually arrives in under two seconds. The panel shows a short typing indicator while the agent retrieves and writes the answer, so there's no blank screen.

If the visitor sends a follow-up question, the AI reads the whole conversation history before generating the next answer. Context carries across messages — a visitor who asked “what's in the starter plan?” followed by “does that include email support?” gets an answer to the second question that reflects what was already established in the first.

How visitors interact with AI Answers

Thumbs up / thumbs down. Beneath each AI Answer the visitor sees a small thumbs-up and thumbs-down button. When they tap thumbs down, a short prompt asks what was wrong — the answer was incorrect, incomplete, or not what they were looking for. That feedback is logged in the Conversations view in your dashboard, attached to the specific message, so you can review it alongside the source that was cited. Thumbs up is also recorded; over time, the ratio gives you a sense of how well your content covers the questions coming in.

Follow-up questions. The input field stays open after an AI Answer. Visitors can type another question, get another cited response, and keep going — the conversation accumulates just like a normal chat thread. At any point they can ask to be connected to a person, and the conversation transfers to your inbox with the full history attached.

Handing off to a human. If the AI Answer doesn't resolve the question, the agent can offer to connect the visitor to your team. Depending on how you've configured the agent, this might be a button the visitor clicks, or the agent may suggest a handoff automatically when the question is clearly outside its knowledge. Once the handoff happens, the conversation lands in your inbox alongside the full AI exchange, so your team picks up without re-asking what was already covered.

Frequently asked questions

Why does the AI Answer cite a source I didn't expect?

Chattering matches questions by meaning, not by exact keywords. If a passage in a loosely related article happens to contain the most relevant information, the AI will cite it. You can see exactly which passage was used by clicking the citation number in the widget. If a source keeps surfacing in answers where it doesn't belong, you can remove it from the agent's knowledge base or tighten the content so the relevant sections are more clearly scoped.

Can I correct a wrong AI Answer?

Yes. The fastest fix is to add a Q&A pair that pairs the exact question with the correct answer. Q&A sources have higher priority than crawled web pages, so when the same question (or something close to it) comes in again, the AI picks the authoritative pair over a less precise passage from your docs. You can add Q&A sources from the Sources tab on the agent configuration page.

How do I see which AI Answers are performing well?

Open your agent's Analytics view. It shows thumbs-up/down ratios per time period and the questions that triggered the most negative feedback. The Conversations tab lets you filter by “AI resolved” to see every conversation that ended without a human handoff — that list is the clearest signal that your knowledge base is covering your customers' questions.

Does the AI answer every question, or only some?

By default, the agent attempts to answer every question. If no passage in your knowledge base scores above the confidence threshold, the agent falls back to whatever you've configured — a prompt to contact you by email, a suggestion to browse the help center, or an immediate handoff to your team. You set the fallback behavior in the agent settings; the default is to suggest emailing your support address.

What happens if the AI gets something wrong?

The citation system makes errors visible quickly — a visitor can see exactly which passage the AI drew from, and if that passage is incorrect or outdated, you can update the source directly. Correcting the source updates the knowledge base for all future conversations. For questions where precision really matters, a Q&A pair with the exact correct wording is the most direct control you have over what the AI says.