The bug is the tutorial's first line: "set up your vector database"

The load-bearing evidence in the piece is also its thinnest. The claim that ripping out an embedding pipeline and replacing it with grep outperformed the vector setup, by a lot, arrives as a widely-cited case with no codebase, no corpus size and no numbers attached ^[4]. Treat it as a hypothesis about your own data rather than a result you can cite. If keyword search across your files is competitive, that is something you can establish and write down, and the post's own instruction is to establish it before provisioning anything ^[6].

What makes the tutorial opening line a defect rather than a stylistic habit is ordering. "First, set up your vector database" puts a provisioning step ahead of problem definition, so teams reach for Pinecone or Chroma before confirming they have retrieval that needs embeddings at all ^[2]. The post keeps a legitimate case for the index: large, stable knowledge bases such as product docs, FAQs and glossaries, paired with a good reranker ^[5]. Note what that exemption is keyed on. It is corpus size and rate of change, not the ambition of the application.

The same shape repeats in the other layers. On agents, the test offered is whether you can say what each one does that a single call could not; if you cannot, you have one prompt wearing several hats and billing for each ^[8]. On fine-tuning, the post's position is that weights are for tone, format and task alignment ^[11], and a bad place for facts, since models memorise them poorly and forget them unpredictably, while a retrieval layer can be corrected in seconds ^[10]. On memory, the machinery has to decide what to keep, what to age out and what to resurface, which is overhead in an app that answers a question and moves on ^[12]. Across all four, the prescribed remedy is identical: build the version without the layer, and add it only when a specific measured shortfall forces it back in ^[14].

That remedy has one property the architecture diagram does not. It leaves a record. The failure the post describes is not enthusiasm for scaffolding, it is the absence of a baseline: layers landed before anyone could name the problem each was meant to solve, so nobody can say afterwards what any of them improved ^[1]. The framing follows from that. Starting too big is the mistake, not starting too small ^[13]. The cheapest artifact in an AI build is the plain-search baseline you kept and the number it produced, because it is the only thing that makes the next layer arguable instead of assumed.