Principia — Applied AI notes from Magellan PartnersField notes, teaching material, research, reproducible benchmarks and live demos from the Magellan Partners AI team.https://principia.vldv.ovh/enEleven open-weight models, one tool-calling harness, 2,200 taskshttps://principia.vldv.ovh/benchmarks/toolcalling-open-weight-2026-09/https://principia.vldv.ovh/benchmarks/toolcalling-open-weight-2026-09/The ranking is not the one you expect. Smaller models with a strict JSON schema beat larger ones with free-form tool calls — and retries, not raw capability, explain most of the gap.Wed, 02 Sep 2026 00:00:00 GMTBenchmarksAgentsEvaluationThe AI Act's GPAI obligations, read by people who have to comply with themhttps://principia.vldv.ovh/watch/ai-act-gpai-obligations/https://principia.vldv.ovh/watch/ai-act-gpai-obligations/The general-purpose AI provisions entered into application in August 2025. A year on, here is what actually changed in how we ship models for clients — and what did not.Wed, 02 Sep 2026 00:00:00 GMTWatchPolicySafetySpeculative decoding is finally boring — and that's good newshttps://principia.vldv.ovh/watch/speculative-decoding-boring/https://principia.vldv.ovh/watch/speculative-decoding-boring/Every serving engine now ships it, defaults are sane, and the 2× is real on the workloads that matter. Here is what to switch on.Tue, 01 Sep 2026 00:00:00 GMTWatchLLMMLOpsChunking, explained with a ruler and a pair of scissorshttps://principia.vldv.ovh/learn/chunking-ruler-scissors/https://principia.vldv.ovh/learn/chunking-ruler-scissors/Before embeddings, before vector databases, there is a much dumber question: where do you cut the document? It turns out the answer decides most of your retrieval quality.Fri, 28 Aug 2026 00:00:00 GMTLearnRAGCan a judge model grade its own family? Preliminary resultshttps://principia.vldv.ovh/research/judge-model-own-family/https://principia.vldv.ovh/research/judge-model-own-family/LLM-as-a-judge is everywhere. We asked whether a judge is systematically kinder to models from its own vendor — and got an answer we did not like.Tue, 25 Aug 2026 00:00:00 GMTResearchEvaluationLLMMCP one year on: what actually got adoptedhttps://principia.vldv.ovh/watch/mcp-one-year-on/https://principia.vldv.ovh/watch/mcp-one-year-on/The protocol won; most of the servers did not. A look at which patterns survived contact with production.Sat, 22 Aug 2026 00:00:00 GMTWatchAgentsToolingNew: the chunking playground now supports scanned PDFshttps://principia.vldv.ovh/demos/chunking-playground-scanned-pdfs/https://principia.vldv.ovh/demos/chunking-playground-scanned-pdfs/OCR runs before chunking, so you can compare strategies on the documents you actually have — the ugly ones.Thu, 20 Aug 2026 00:00:00 GMTDemosRAGToolingRetrieval is not the bottleneck anymore. Reranking is.https://principia.vldv.ovh/research/reranking-is-the-bottleneck/https://principia.vldv.ovh/research/reranking-is-the-bottleneck/On three client corpora, swapping the embedder moved recall@10 by 2 points; swapping the reranker moved answer accuracy by 11.Tue, 18 Aug 2026 00:00:00 GMTResearchRAGEvaluation