1/π§΅ Just finished another Agentic AI class at Johns Hopkins! π Key learnings: Agentic RAG vs normal RAG.
Data chunking: *metadata filtering* becomes crucial. A "self-querying tool" separates filters from the query to retrieve relevant chunks β great for date-based filtering!
Chunking is the hidden trap βοΈ
Too small = sharper search, but the chunk loses context β wrong info lands in the prompt.
Fix: chunk at natural breaks, search small, retrieve the BIG parent chunk for the model. And LLMs forget the MIDDLE β so keep key info at the end. β
Week 1 of the Agentic AI Program at Johns Hopkins π
(Just finished the Generative AI Program β )
How AI got here:
1οΈβ£ A rule for everything β impossible
2οΈβ£ ML: patterns + probabilities (1990s)
3οΈβ£ LLMs: a brain in a jar β clever, but passive
The plot twist: you don't need the strongest model if your retrieval (RAG) is great.
RAG = 3 parts:
π Knowledge base
π Retriever (the search)
π‘ Generator (the LLM)
I like the point from @steipete that running agents is easier for an engineer who has been a founder before given they don't really write code anymore. But then also the point that sometimes you need to go 1:1 with an agent. Also the point on nobody valuing AI slop and "if you want attention, be human" and if you want a better chance of having your PR accepted into Openclaw "have some typos in your PR or submit your logs to prove effort". https://t.co/6d3HlLEcHA
Great article on what's behind Loveables credit costs. Main takeaway is writing the app requirements in PRD can save 30-45% of costs. Plan mode also no brainer at 1 credit a message regardless. https://t.co/V02uBtVSO3
Oh and lastly Openrouter has a cool feature to run a classifier on all your input prompts and tag them accordingly so I've added this so I can see what I'm actually using OpenClaw for