Artificial intelligence
Post LinkedIn lead magnet · Enterprise software
I've spent the last 2 years of building the most sophisticated retrieval engine for company knowledge I shared it all in a webinar yesterday 1/ Why early RAG broke and got such bad press In 2023, teams were jamming entire documentation into GPT-3.5 and getting garbage back. The models were weak and got even weaker with large messy context. Now GPT-5 and Opus 4.6 have gotten 10x smarter and can handle much more complex retrievals and non-deterministic tasks. 2/ There are many R(etrievals), and you need them all → your good old keyword search (find "Project X" or close) → semantic (finds "churn" when you search "customers leaving") → structured (SQL queries) → API/MCP retrieval (live data from Salesforce) → Hybrid ("All tickets in Intercom in last 2 weeks mentioning pricing issues") We demoed all four running together in Super. One search bar, four systems working in parallel. 3/ There are 4 levels of RAG sophistication. Naive is embed, match, generate. That's where most teams stop. Multi-loop re-ranks results across multiple passes. Agentic means the AI picks which tools to query (we showed this live with MCP). Self-correcting retries when it fails and builds memory. 4/ The build vs buy question comes down to permissions and data size. → If your data fits in a context window, just paste it into Claude. → If you know where the data is, use API or MCPs and automations. But if the answer is "somewhere", scattered across 50 tools with different access rules, you're about to spend 8 months building permission-aware retrieval. 90% of our customers tried building it before coming to Super. 5/ Native AI connectors lack semantic + company understanding ChatGPT's Slack integration can't do semantic search. It doesn't know that your team uses "P0" to mean urgent or that sales decks live in a specific Drive folder structure. It just searches filenames. Same with Claude's Drive connector. They're built for personal use, not enterprise knowledge. 6/ The part that got the most questions was loose orchestration. Instead of hardcoding "search Slite, then check Slack, then query the CRM," you give the AI dynamic tools it can call and say "find out why this customer churned." The AI figures out the steps. One person asked if we had a flowchart. We don't. That's the point. Let me know if you'd like the recording and I'll send it! And of course, thanks for having me Alexandre Kantjas, Pierre-Yves Garcia and 9x!😉
Mécanisme lead magnet
Let me know if you'd like the recording and I'll send it!
