Free · open · no sign-up
Grounded AI Consultant: how to stop a store assistant from inventing products
A catalog bot that invents a SKU or a price is not a cute model glitch. Grounding is a constraint: retrieval, a runtime guard, and a human handoff.
The model predicts the next span of text. It has no stockroom and no habit of saying it does not know. Asking it to sound warmer makes invention more likely, not less.
Nine modules in the player. The matching delivery is AI consultant and chatbots. The matching case is the grounded catalog consultant.
For whom
- Catalog owners who will put a consultant on the site, not a widget from a builder
- Operators who need a checklist to brief or to audit a contractor
- Not a sales page for the consultant service — that lives under Services
Modules
- Why an LLM lies
- Product vs commodity box
- Harness architecture
- Prompt composition
- Grounding and retrieval
- Runtime guard and fallback
- Shield · not replacement
- Verification gate
- Deploy · economics · support
After the course
- Explain why an LLM invents catalog facts, and what grounding actually means
- Sell a done-for-you consultant, not a box
- Build a thin kernel plus instances: a new client is a persona swap
- Compose the prompt from blocks, not a monolith
- Run deterministic retrieval of candidates before the model speaks
- Set guard, retry, and a closed fallback into lead capture
- Treat the handoff as a conversion event
- Tell golden tests from adversarial ones
- Keep production drift in view after deploy