The message

I like the shopper thing — I'm not building it, but a structure for something different. A mystery-shopper price check is very similar, so well done, that's a perfect use case. The hardest thing is setting up things — you can do OTP but then you run out, can't do it all the time, especially with the same companies because their systems block it, especially ones using Salesforce and HubSpot. The other thing — you can use agentmail, and your orchestrator agent can set up another dummy email. That can be done. Good one, I like that.

The RAG stuff — I understand what you're trying to do. Just remember: vector stores by themselves, RAG as you know it, is limited. RAG was unreal two years ago because it wasn't there before. Now we use a lot of different things. I've got a whole ingestion platform — bring it in, store it all locally, not in vector stores, all coded for LLMs, it all links up, effectively semantic + graph, many layers deep so the LLM follows a road to real data. Then it's about making sure it doesn't make it up on the way out. Very hard to have industry-specific — if it's about an industry you then have to rebuild it for a company.

Even for what we're doing with legal, we have three layers: the compliance/regulatory stuff (massive) — the LLM refers to it but it's also a guardrail (we've got 14). It has more knowledge than a compliance officer at a tier-one firm. Then the firm layer — history, what it does, USPs, SOPs — lighter and focused, for conversation/brand ID. I have a completely different JSONB file for context profiles per client. Layer upon layer. Your knowledge base should always be layered — never one RAG. In my view only some information should be ragged. We keep a context profile on every contact in structured JSONB the LLM reads in a heartbeat — a low-grade LLM, doesn't need a frontier model.

Lots of things, mate, but you're getting there. I do like the insurance one, well done. Just build solutions — productize it. I might show you what I'm doing next week — you might just copy me. Maybe I won't. Have a good weekend.

What it changes