AI Customer Service Knowledge Base: What Goes In
The knowledge base decides most of what an AI customer service system can say. A strong model with a thin knowledge base writes polite replies that say nothing; an ordinary model with a good one answers the question. This is what the knowledge base behind our case study contains, how it is searched and how it is kept correct.
What goes in
- Product facts by model and version. Specifications differ between versions of the same product, so each fact names the version. Where versions differ, the storefront's product page is the reference.
- Published policies, word for word. Shipping, returns, warranty, invoices. The AI may quote these directly.
- Troubleshooting steps. The full sequence for each common fault: what to check, how to free a stuck part, how to use the product temporarily, what the likely cause is. These are the answers customers want first.
- Video guides. A link to a short repair or setup video often replaces three paragraphs.
- Parts and accessories. Which part fits which model, so a replacement proposal names the right item.
What stays out
Internal status names, team assignments and raw customer messages stay out. Lessons from individual tickets go into a separate library of past mistakes and their correct handling, written as general rules without names, order numbers or quotes. That library is searched by situation, so a lesson from one ticket applies to the next similar one without copying anyone's words.
Search before every draft
Do not let the model decide whether it needs the knowledge base. The system searches before every draft and passes the hits to the model; the model decides how to use them. In our tests, what decided retrieval quality was the query. One short question about one thing, phrased the way the FAQ is phrased, put the right entry first. Long queries that mixed several topics buried it.
So each email is broken into its separate questions, each question becomes one short query, and each query is searched on its own.
Never cut content while processing it
Every filter between the knowledge base and the model is a place where answers can disappear. We once added a step that removed video links from retrieved text by cutting at sentence boundaries. Because steps were written as one long sentence joined by commas, it deleted whole troubleshooting sequences, parts, causes and next steps, in a quarter of a day's drafts. The rule since: a filter removes only its target, and any change to a processing step is checked against a real knowledge entry before and after, word for word.
Correct the copy, keep the sync
The team maintains the source documents, and the system keeps a synced copy. When a source entry is wrong, waiting for the document owner can take days. The system keeps local corrections that are applied when an entry is read, so the next sync does not overwrite them. When the source document is fixed, the correction is removed.
Keep one version of each fact
When the same fact lives in several places, such as a product sheet, an FAQ, a reply guide and a translated copy for each market, the copies drift apart and the AI quotes whichever one it finds first. Keep each fact in one entry, give it the version it applies to, and let the reply step translate it for the customer. When a fact changes, one edit updates every reply in every language, and reviewers never have to ask which copy is the current one.
Signs your knowledge base is the problem
- Agents keep replacing "we have no information" with an answer they know. In one review this was 10% of all edits.
- Replies send customers to a dealer or a form instead of giving the fix.
- The same question gets different answers on different days.
For retrieval failures in more depth, see RAG not finding answers. The free plan reviews your current knowledge base against real tickets in two working days.