@desto6:

destokEla
destokEla
Open In TikTok:
Region: CM
Friday 17 July 2026 20:37:25 GMT
7
0
0
2

Music

Download

Comments

There are no more comments for this video.
To see more videos from user @desto6, please go to the Tikwm homepage.

Other Videos

RAG that doesn't hallucinate — the myth, the real failure, and the fix in order 📖 Everyone says
RAG that doesn't hallucinate — the myth, the real failure, and the fix in order 📖 Everyone says "just add RAG and your AI stops making things up." It's not true. Paid, "hallucination-free" legal AI tools still invented court cases — somewhere between one in six and one in three answers. So bolting retrieval onto your model does NOT end hallucination, and if you don't know why, you're fixing the wrong end. Quick version of what RAG is: a normal AI answers like a student taking a closed-book exam — from memory, and sometimes it confidently misremembers. RAG (retrieval-augmented generation) turns it into an open-book exam: before it answers, the system finds the exact pages that matter and hands only those over. Why? The model's memory is frozen at training — it's never seen your files, your pricing, last week's policy. Here's the part nobody tells you: the #1 way RAG fails isn't that it couldn't find the answer. It finds the right page, puts it on the desk… and answers from memory anyway. Researchers call this evidence override, and it happens far more often than coming up empty — on some tasks, four in ten answers ignored a correct page sitting right in front of the model. So the fix isn't "prompt better." It's a pipeline, in order: Chunk on sentences and paragraphs — not blindly by character count — with a little overlap. Search meaning AND keywords (hybrid search), so one different word can't make the lookup miss. Give every chunk its context before you file it. Add a reranker — pull a wide stack, keep the best few. Build the prompt on purpose — documents on top, tagged, capped. Force grounding — answer only from what you gave it, quote the line first, and let it say "I don't know." Then measure whether the answer actually stuck to the pages. Save the fancy stuff — agentic RAG, knowledge graphs — for last. That's context engineering: the difference between an AI that sounds confident and one that's actually right. Want the whole path, from zero to actually building this — AI, RAG, multi-agent, the full stack? I put together seven courses that take you there. It's all in my profile. #rag #aiagent

About