Knowledge Base Intake and Training Data Gate (Onboarding)
A checklist with 7 steps: Inventory every source the client wants the bot to answer from before any upload begins.
By InnovaAI ResearchPublished
What are the steps?
Knowledge Base Intake and Training Data Gate (Onboarding)
- 01
Inventory every source the client wants the bot to answer from before any upload begins
Collect help-center URLs, PDFs, pricing sheets, and internal macros into one list, then mark each item as public, internal, or restricted. FastBots and Chatbase both train from uploaded files and links, so an unclassified source becomes a live answer the moment it is indexed.
- 02
Assign a single client-side owner who signs off on the source list
Name the person in the kickoff notes and put their approval in writing. When a bot quotes a stale price or a retired policy, the agency needs a documented chain showing who approved that document.
- 03
Strip superseded documents instead of uploading them alongside current versions
Two conflicting refund policies in the same knowledge base produce contradictory answers within the same conversation. Delete the old version at intake rather than trusting retrieval ranking to prefer the newer file.
- 04
Convert long-form documents into question-and-answer pairs where the client already has them
Support teams usually keep a macro library that maps directly to real customer phrasing. WotNot and Chatling both accept structured knowledge entries, and pre-written Q&A pairs outperform raw PDF chunks on first-contact resolution.
- 05
Define the refusal boundary for anything the bot must never answer
Legal advice, competitor comparisons, and account-specific billing questions belong on the refusal list. Chatbase supports explicit guardrails and procedures, so write the boundary as instructions rather than hoping the model infers it.
- 06
Run 25 real historical questions against the trained bot and score each answer
Pull the questions from the client's last 90 days of support tickets, not from a brainstorm. Score correct, partially correct, or wrong, and treat anything below 80 percent correct as a knowledge gap rather than a model problem.
- 07
Log the training corpus version and re-test date in the client's delivery record
Record which documents were live on the go-live date. When the client changes a policy in month three, the agency can show exactly what the bot was trained on and scope the retraining as a billable change.