What it is, and how it differs from a help center
An AI knowledge base is the set of sources a retrieval system searches before a model answers. IBM describes the RAG pattern as retrieving relevant passages for a question and providing them to the LLM. OpenAI’s retrieval guide says you pass the search results and the original query to a model yourself.
A help center is pages a customer browses; a knowledge base is a chosen source list that can include playbooks and contract text. In a shared Slack channel the bot answers in front of the customer, so if a file would embarrass you quoted back, keep it out of that bot’s index.
Scope, conflicts and response-time questions
Keep one index per customer for contract text and one shared index for public docs, and query only those two. Metadata filters exist, but a separate index fails safe where a forgotten filter does not.
Precedence. The signed agreement wins for that customer. Between public documents, the later effective date wins. If the agreement is silent, the newest public policy applies. If the newest public policy and the agreement disagree, the bot does not choose. It posts “I don’t have a confirmed answer for that, so I’ve passed it to our support team” and a person answers.
Response times. Index one line per customer, for example “Effective date: [date]. Example Co first response target: 4 business hours” (example figures, not a benchmark). If the retrieved text has no figure for that account, hand off. Never estimate from the public help page.
Updates, chunking and testing
Replace a changed file instead of adding a copy. OpenAI’s retrieval guide says removal is eventually consistent, with no timeline given, so test removals as well. After a refund window moves from 30 days to 14, ask “Can I still get a refund after 20 days?” right away and the next day. Restore the previous file if a bad version gets indexed.
As of November 2026, OpenAI’s retrieval guide sets defaults of 800 tokens per chunk and 400 tokens of overlap. A fixed window can split a rule from its exception, so keep them in one section and flatten tables to one sentence per row. Anthropic chunks plain text and PDFs into sentences, and citing images from PDFs is not currently supported.
Test five questions per policy area, including one the sources cannot answer and one where two sources conflict. Rerun the set after each change, and sample 20 answers a week for correct and grounded rates per source document.
Citations and ownership
Anthropic’s citations docs say plain text citations carry a character index range, 0-indexed with exclusive end indices. Citations need document blocks with citations enabled, on all or none of a request’s documents, so check each response for citation blocks. OpenAI’s file search results are not returned by default; add file_search_call.results to the include parameter to keep them.
Review rule. If the cited passage lacks the number or condition in the answer, the bot does not post it. It posts the handoff line and sends the draft to an internal channel for a person. Name one owner who approves additions before indexing and reviews the weekly sample. Assume every guest can read what the bot posts, file titles included.