Grounded customer answers from your knowledge base, using RAG

Customers ask the same questions your help center already answers; they just cannot find the article. Retrieval-augmented generation finds the relevant passages in your own content, answers in plain language with links, and hands the conversation to a person when the content runs out.

01the problem and who owns it

Help centers grow to hundreds of articles written by different people at different times. Search matches keywords, customers use different words, and they give up and open a ticket for something already documented. Agents spend their day pasting article links.

A general chatbot makes this worse if it answers from model memory instead of your content: it will describe features you do not have and policies you never wrote. The knowledge manager has to be able to see exactly which passage produced each answer.

02what the AI does, step by step

1. Inventory and clean the sourcesHelp center articles, product docs, policy pages, and approved macros are collected. Outdated, duplicate, and internal-only content is removed or marked, because retrieval will faithfully surface whatever is in the index.
2. Chunk and indexContent is split into passages that keep headings and context, embedded, and stored in a search index that supports both semantic and keyword matching. Each passage keeps its source URL, product area, and last-updated date.
3. Retrieve for each questionThe customer's question, plus relevant account facts like plan or product, is used to fetch the best passages. Results are re-ranked so the most specific, current passage wins over a general overview.
4. Answer only from what was retrievedThe model writes a short answer using the retrieved passages, cites the articles it used, and is instructed to say it does not know when the passages do not cover the question rather than filling gaps.
5. Hand off with contextWhen the answer is missing, the customer is frustrated, or the question involves their account specifically, the conversation goes to a human with the transcript and the passages already tried.
6. Evaluate continuouslyA test set of real questions with known correct answers runs on every content or prompt change. Production conversations are sampled weekly and graded for correctness, grounding, and tone.

03systems it connects to

• Content sources. Help center platforms such as Zendesk Guide, Intercom Articles, or Confluence, plus product docs in a docs site or repository.
• Search index. A vector database or search service with hybrid keyword and semantic retrieval.
• Chat surface. Your existing chat widget, help center search box, or in-app help.
• Helpdesk. For human handoff, with the conversation attached as a ticket.

04human checkpoints

• Content approval for indexing. The knowledge manager decides which sources are in the index. Draft and internal articles stay out.
• Launch gate. The test set must pass an agreed bar before the assistant answers customers, and again before any major change goes live.
• Policy topics. Refunds, legal terms, and pricing exceptions are answered only with approved wording or handed to a person.

05what to measure

• Resolution without handoff. Conversations the customer marks resolved, validated by sampling, since self-reported resolution overstates.
• Grounded answer rate. Share of sampled answers fully supported by the cited passage.
• Repeat contact. Customers who open a ticket on the same topic shortly after a bot answer.
• Unanswerable questions. Questions with no good passage, which become the content backlog.

06risks and guardrails

• Hallucinated policy. An invented refund promise can bind you or anger customers. Restrict policy answers to approved text and test them specifically.
• Stale content. Retrieval surfaces old articles as confidently as new ones. Re-index on publish and prefer recently updated passages.
• Prompt injection and data leaks. Customers will try to make the bot ignore its rules. Keep account data access narrow and never index content customers should not see.

07build vs buy

Most helpdesks now sell an AI agent that answers from their own article base, and for a single-product company with a tidy help center it is often the quickest win.

A custom RAG build earns its keep when answers must draw on several sources (docs repos, policy PDFs, product data), when account-specific context matters, or when you need your own evaluation harness and control over which model runs.

08related playbooks

Browse every customer support playbook or the full library.