What does a real RAG pipeline look like in production?
In this talk at @7CTOs, I walk through:
→ Chunking strategies
→ Search (BM25 + vectors)
→ Query rewriting, HyDE
→ Reranking, repacking
→ Tool calls
It’s a full-stack view of modern LLM app design.
https://t.co/NksC0APNH5
#RAG #LLM #AI #GenAI
Hot take: AI startups relying solely on OpenAI, Anthropic, or Gemini are as vulnerable as the old social startups that were entirely dependent on Facebook's API. One API change could spell disaster. Everyone in this space should experiment with running their own models.
1/ Observations of a vibe coder 🧵
I’ve been vibe coding for a while now — side projects, little work utilities, messing around. Here are some things I’ve learned. 👇
9/ If I were a Jr dev today:
I’d focus on what I learned in data structures and algorithms classes, not syntax. The skill is in applying those patterns, prompting better and reviewing the output.
8/ What I really want:
A local version of GitHub’s PR UI. Inline diffs are fine, but I want full diffs + explanations + line comments + back-and-forth refinement.
6/ Ask for smaller changes.
Asking it to “build the whole feature” rarely works. I aim for something between a loop and a full feature — and give context for where it’s going.
5/ Your code review skills matter more than ever.
You wouldn’t let a junior dev commit without a review, right? Same goes here. You should understand the diff before committing it — or before prompting again.
4/ Keep your repo clean.
Never let it work on a tree with uncommitted changes. I make tons of local branches and commit often. Makes it easy to revert if things go sideways.
3/ Orchestration matters.
Sometimes I need @claude_code to design something, @cursor_ai to refactor it, and @windsurf_ai to finish the last 10%. It’s less like “pick a model” and more like conducting a band.
2/ Tools matter.
I flip between @claude_code, @windsurf_ai, and @cursor_ai. Even when using the same model, they get stuck in different ways. Switching tools helps me stay unblocked.
Not all chunks are created equal.
In my latest blog post, I break down how Benchmate approaches chunking for RAG—from simple token splits to full semantic clustering.
“An ideal chunk is one answer to one question.”
Read more: https://t.co/MUe77Zoca3
#RAG#LLM
New post is live: “Data Sources and Retrieval-Augmented Generation (RAG)”! Dive into how high-quality data sources transform AI-driven insights in pharma research. #AI#RAG#Pharma#DrugDiscovery
https://t.co/h3nu2NmcTD
I spent a bit of time vibe coding up an app I’ve wanted to build for a long time. It took less than 30 minutes and $3 in Anthropic credits and I had a fluid dynamic simulation that runs in the browser.
As software developers we really need to expand our ideas of what we can accomplish.