Knowledge Base Intake and Grounding Audit (Onboarding)
A sequence with 7 steps: Inventory every source the client wants the bot to answer from, then rank each by authority.
By InnovaAI ResearchPublished
What are the steps?
Knowledge Base Intake and Grounding Audit (Onboarding)
- 01
Inventory every source the client wants the bot to answer from, then rank each by authority
Separate canonical documents (pricing sheets, policy PDFs, signed SOWs) from marketing pages and stale help articles. Anything older than 12 months gets flagged for client sign-off before it enters the training set.
- 02
Test the raw source material against 20 real customer questions pulled from the client's ticket history
If the source set cannot answer 15 of 20 verbatim, the gap is a content problem, not a model problem. Document the misses and send them to the client as a content request before configuration begins.
- 03
Upload and index the approved corpus, then verify retrieval on the same 20 questions
Platforms like Chatling and FastBots auto-sync from a website URL, which is fast but pulls navigation text and blog filler. Where the client has a clean document set, prefer file upload over URL crawling.
- 04
Write refusal and fallback language for every question the corpus cannot answer
A bot that guesses on pricing or compliance questions creates liability the retainer does not cover. The fallback should name the human channel and the expected response window.
- 05
Run a hallucination sweep with 10 adversarial prompts designed to pull answers outside the corpus
Ask about competitors, unlisted services, and hypothetical discounts. Log every fabricated answer with the exact prompt so the client sees the failure mode rather than a claim that it was tested.
- 06
Get written client sign-off on the corpus version, the fallback copy, and the known gaps
Record the date and the document hash. When the client later asks why the bot did not know about a new product, the sign-off log is the answer.
- 07
Schedule the first corpus refresh and assign an owner on both sides
Monthly is the practical floor for clients with active pricing or policy changes. Put the refresh in the retainer scope so it is billable work rather than an unbilled favor.