Two ways to answer a question
Same three sources, same question. Toggle the strategy, then press ask.
Pay the synthesis cost once
Classic RAG is a promise that similarity search plus a large context window equals understanding. Sometimes it does. But you're re-paying the reading cost on every question, the model free-associates over raw text it has never vetted, and two identical questions can get two different answers depending on which chunks surfaced that day.
The LLM-Wiki idea flips the cost curve: when a source arrives, an agent reads it once, properly — writes a cited Source Summary, folds new facts into the concept pages they belong to, and proposes a couple of atomic notes for the human's inbox. Questions are then answered from the compiled layer: small, curated, cited. The 2am answer above didn't touch the 31-page PDF; it read one section of one page that was written to be read at 2am.
That compile step is hosted now, not aspirational. vault_ingest_source accepts text, HTML,
and PDF (extracted with unpdf); Workers AI returns schema-constrained content channels while
code constructs the Source Summary, EXTRACTED bullets, citations, inferred marks, directories, and two
or three 0-Inbox/ seedlings. Bad model output gets one corrective retry and then fails
visibly. The model supplies content; it never gets to invent the trust structure.
Every compiled claim is marked EXTRACTED (traceable to a source) or
*(inferred)* (the agent's synthesis). Synthesis is welcome — disguising it as sourced fact
is the failure mode this whole system exists to prevent.
A format the machine can hold you to
Wikis rot when pages are freeform. OKF ("open knowledge format") is deliberately tiny — a YAML header
every page must carry, and a validate command that machine-checks the whole vault. Hover
each line:
validate() can enforce it vault-wide.Your thinking stays yours
A vault is not only the wiki. It's your PARA structure — inbox, projects, areas, resources, archive — plus your atomic Zettel notes, with the wiki layered beside them. The division of labor: the human layer is thinking (ideas in your own words); the wiki is the compiled reference that absorbs the bookkeeping — summaries, cross-references, currency — so you can keep writing instead of maintaining.
Code before the model
The retrieval hot path contains zero model calls. Extract keywords (lowercase,
stopwords out, dedupe). Score the catalog — plus three per hit on an index line, plus two on a
filename, plus one on a directory segment — without opening a single page. Take the top page,
split on ## headings, return the best section. Done.
That's the whole trick, and it's the one that compounds: retrieval becomes reproducible (same question, same evidence, forever), debuggable (the scores are right there), free (it's string matching), and biased toward the curated layer — the model never free-associates over the corpus, because it's never shown the corpus. It's shown one section of one page and told to cite it.
Try it yourself — the shipped algorithm runs live in chapter 01's query engine.
Agents propose; you promote
The last discipline is the social one, enforced in code twice over: every agent-written page enters as a seedling, and only a human hand moves it forward. Click promote:
Four ideas, one system: compile once, hold pages to a contract, keep the human layer human, and let code do the retrieving. Helios is these ideas with tenancy, auth, and a deploy pipeline wrapped around them — field note 2 walks you from an empty vault to a working one.