6 articles
A document agent that decides what to search and read next (ReAct)
Plain RAG retrieves once and answers. agent-flow runs a think→tool→observe loop: it decides for itself what to search and read next, and every claim in the report carries a file:line citation. Plus a human-approval gate.
BM25 vs Embeddings: a tiny benchmark showing semantic search is brittle
Embeddings understand meaning, so they always beat keyword search — right? I built a small benchmark that proves otherwise, and shows exactly when semantic search falls apart.
I built a document search that doesn't hallucinate — here's how
A RAG that shows the file and line it pulled an answer from, instead of making things up. How the BM25 + embeddings + RRF hybrid works, and why a single method fails.
From a described process to a ready .bpmn file: a BA skill in Claude Code
You describe a process in plain words and get an AS-IS/TO-BE model in ASCII, gateway rules, edge cases, and a real .bpmn file with auto-layout that opens straight in bpmn.io. No hand-clicking XML coordinates.
What the data says on AI adoption: Poland vs the EU — a two-office SQL duel
Poland adopts AI fast, but from such a low base that the gap to the EU is widening. Hand-written SQL reconciles the same metric across GUS and Eurostat — JOIN, reconciliation, window functions.
Why your chatbot makes things up (and what RAG has to do with it)
A chatbot doesn't lie — it generates the most likely text, not the truth. Plain-language take on hallucinations and how RAG curbs them by attaching sources before the model answers.