CrocoBrain concept
Proof before eloquence
“Why not just vector RAG?” The question always comes back. The answer fits in a sentence: because an eloquent answer with no proof is just a well-dressed opinion — and a brain that thinks like you must be able to show where each of its words comes from.
The problem: eloquence without proof
Generic RAG tools all work the same way: chop your documents into fragments, fetch the fragments closest to a question, and let a model stitch them into a fluent answer. The result reads well — too well. Nothing in the prose tells apart what truly came from your notes from what the model filled in on its own.
And when these systems cite, they print a similarity score: “relevance 0.82.” A number that reassures without proving anything — it measures vector proximity, not truth. An answer assembled from out-of-context fragments and decorated with a score keeps the confidence of a citation without its reliability. That is exactly the failure shape that makes a clone hallucinate under pressure.
The CrocoBrain choice: navigate, don't dredge
CrocoBrain does not dredge up fragments. It navigates your brain the way you would: the index first, then the notes by their titles and their [[links]], descending to the raw sources only when the wiki layer no longer suffices. Every step is structured, traceable, and lands on a note you wrote — not on a fragment dug out of a pile.
From that navigation flows a single kind of proof: the [[note]] citation. Every claim points back to the exact note it came from, and every note traces back to its sources. One trust model, verifiable in a click — never a score, never an anonymous fragment.
What the vector does here: orient, never answer
CrocoBrain is not hostile to embeddings — it simply refuses to hand them the answer. Internally, the vector signal may prioritise, pre-filter, suggest: helping the brain look in the right place. It never leaves the wings. No answer is ever built from a vector top-k, no citation is ever a vector, no score is ever shown.
One line captures the division of labour: “Sonnet reasons. The wiki proves. The vector orients. The user arbitrates.” The vector has a role — modest, internal, never the last word. Proof stays with the wiki; the decision stays with you.
The brain that can say “I don't know”
The most visible consequence of this choice is a confession: when your brain has nothing to answer with, it says so. No stitched-together fragments to fill the silence, no plausible answer manufactured from nothing. An honest silence beats a confident invention — and that is precisely what a generic RAG tool cannot afford to admit.
That “I don't know” is not a dead end: the unanswered question becomes a gap, added to the list of zones your brain would do well to fill. And for the public version of your brain — the clone your audience questions — the rule tightens further: it speaks only from mature notes, never from a draft. Better to say nothing than to speak wrongly in your name.
What it changes for you
A brain that proves is a brain you can defend. Every answer traces back to your words, through a chain you can unwind: the note, its source, your arbitration. You never get an average opinion computed over a pile of fragments — you get what you think, cited.
And because this proof is yours, it leaves with you: everything exports as Obsidian-compatible Markdown, notes, sources and links included. Global vector RAG optimises for eloquence; CrocoBrain optimises for proof. That is a choice, not a limitation. For what comes next — how a note earns the right to be believed — read earned curation.