context engineering, completed:
entity memory as a permanent fact store decoupled from recency
LLM-based entity extraction and its cost mitigations
context overflow and token budget management
context conflict resolution via source hierarchy and labeling
Went deep on hybrid search for RAG:
BM25 + semantic search, combined via Reciprocal Rank Fusion (rank fusion beats score fusion since BM25/cosine scores aren't comparable)
Cross-encoders vs bi-encoders for reranking -- why joint attention catches what cosine similarity misses
so when im bored i go to Pinterest or substack or goodreads to see the edits or quote from one of the book im reading or like any random book.
But yk I want smth where i can read all the quotes from the very first start page to the last.
So im taking matters in my own hand.
I got claude pro ahahah screaming
How do I use it to its actual potential. Please no comments like build fable or claude itself. Something that made your life better?!?!
#claude#questiontime
Introducing Claude Fable 5: a Mythos-class model that weβve made safe for general use.
Its capabilities exceed those of any model weβve ever made generally available.
Today I finally got around to exploring LangSmith.
Connected it to my AI application, traced agent runs, used an LLM-as-a-judge evaluator to measure response quality.
Seeing the entire execution flow and evaluation results made debugging much easier than relying on intuition.