We fed our whole knowledge base into a 1M-token prompt. Results:
1. Answers slower
2. 6x the cost
3. Missed facts buried mid-context
Retrieval still wins on speed, cost, accuracy.
Curious how we wired it? Happy to compare notes.
#RAG#LLM
When a RAG bot 'hallucinates,' it's often quoting a 4-month-old copy of your own docs. Rarely a chunking or prompt issue. Nobody re-indexed after the content changed. Our fix: per-source sync timestamps + a nightly diff re-embedding what changed. Building this? Door's open. #RAG
IBM's 'Bob' writes enterprise code in minutes.
That was never the hard part.
The hard part is owning the architecture when it breaks 18 months later. Generating code got cheap. Owning the decisions didn't.
Ever inherited an app nobody could explain?
AI writes code 10x faster. Enterprise teams still ship late.
The bottleneck moved. Typing was never the slow part. Deciding what to build is.
Week one of a client build, we halve the feature list. Fast code on the wrong thing is quicker waste.
Where do you start cutting?
Everyone's reading the Virgin Media £28m fine as 'don't trap customers.'
Wrong lesson.
Your cancel button is marketing. The 2-click exit is why they come back and refer you.
We build cancel flows we'd be proud to screenshot.
What's the worst one you've hit?