We gave @cfosilvia a filesystem instead of a vector database. No embeddings, no chunking, no vector index. It runs grep. Here's why: Almost every AI product does memory the same way. Chunk the document, embed the chunks, return whatever sits closest to the question. That fails in specific, expensive ways when the documents are financial. A cost basis table gets split across a chunk boundary. Retrieval returns the header without the lots, or the lots without the account. The model doesn't know it's holding half a table. It does the math anyway. Then there's exact match. Ask about section 1202 and similarity search hands you a passage about qualified small business stock that never names the section, while missing the one that does. Embeddings map things that mean the same thing to the same place. In finance, the things that must be exact are the things that get smoothed away. And you only get one guess. If the right passage ranked eleventh and you asked for ten, nothing tells you it's missing. The answer arrives with the same confidence either way. RAG was built for a world where the model was weak, and the window was small. You had to guess the right paragraphs up front because that's all you could fit, and you got one attempt. The chunking strategy, the reranker, the hybrid search. All of it exists to make one guess as good as possible. A modern agent doesn't work that way. It can search, read what came back, decide that wasn't it, and search again with a better term. Once the thing can look around, you don't need to guess right the first time. You need to give it a room to look in. So we stopped building a better guesser and built a better filing cabinet. There's a second choice underneath that one: why the filesystem doesn't live inside the sandbox. That's the part that decides whether any of this is affordable to run. Wrote it up here:

How @cfosilvia routes models: a small classifier reads the first few hundred characters of your message and predicts, in under a millisecond, how much work the whole task will take. Learn more from Silvia Labs:

When you ask an AI a hard tax question, it answers with total confidence. The problem is that confidence and correctness aren't the same thing, and in tax, a wrong answer that sounds right is worse than no answer at all. So we built something different, and tested it against everyone. We evaluated 7 of the most popular AI products across 10 of the hardest tax scenarios, covering federal law and five state tax codes. Every answer was graded 1 to 10 for factual accuracy. @cfosilvia scored 8.73, the highest of any product tested. We open-sourced everything. Full breakdown, with the benchmark and research paper linked: