Hello, world: a terminal that knows me
This site is a terminal. You type /projects or you just ask a question, and an agent answers —
about me, in the third person, in English or Serbian. This first post is about the one
architectural decision that shaped everything else: there is no RAG here.
The default recipe for "chat over my content" in 2026 is still retrieval: chunk the documents, embed them, stand up a vector store, retrieve top-k at question time, hope the right chunk came back. For a corpus like mine, that recipe is solving a problem I do not have. My entire professional life — bio, roles, projects, FAQ, even post summaries — packs into roughly 20–25 thousand tokens. Modern models take ten times that in a single request without blinking. So this site does the simple thing: full-context injection. Every question ships with the whole knowledge base as a prefix. The agent never wonders whether the retriever found the right chunk, because there is no retriever. Recall is 100% by construction.
The objection is cost, and it is a real one — naively, resending ~25K tokens per question would be expensive. Prompt caching is what makes the naive thing viable. The knowledge base is packed deterministically: same files in, byte-identical prompt out, every time, with anything per-request kept strictly after the static prefix. The provider then charges the cached prefix at a tenth of the price, and a periodic warmer keeps the cache alive across a low-traffic day. One byte of drift in the prefix — a timestamp, a shuffled ordering, a "helpful" random example — forks the cache and multiplies the cost by six. Byte-stability is not a nicety; it is the economics.
Full-context also changed how I write the content. Chunk-optimised writing wants every paragraph self-contained, because a chunk may arrive alone. Whole-corpus writing wants the opposite discipline: front-load the facts, use absolute dates ("since March 2024", never "for two years"), give explicit numbers, and keep it short — because every token in the corpus is paid on every single question. Editing my bio now feels like editing a config file that has a monthly price.
The rest of the system exists to keep this honest: a hard token budget on the packed corpus with a CI gate, a daily cost breaker so a hostile traffic spike degrades the agent instead of my wallet, and a bilingual golden eval set — including prompt-injection canaries planted inside the documents themselves — that runs before releases, not nightly.
Would I recommend this over RAG? For a personal site, a product doc set, anything under ~50K tokens that one person curates: yes, without hesitation. Retrieval earns its complexity when the corpus outgrows the context window or changes faster than a cache can live. Mine does neither. The degradation path is written down for the day that changes — and step one is "summarise the posts", not "add a vector database".