The traditional way LLMs edit a doc: read the whole thing, rewrite the whole thing, hope the parts you didn't ask about come back unchanged. That hope is usually misplaced. Daneshbase doesn't ask you to trust it.
Find/replace, not rewrite
When the agent refines a section, it doesn't ship a new full-doc body. It submits a small find/replace patch — "match this exact text, replace with this new text". The server only applies the change if the find string matches exactly once in the current doc. If the model paraphrased the find string, the patch fails fast; if the section the agent meant to edit no longer exists, the patch fails fast.
Either way, the rest of your doc cannot change. By construction.
End-of-doc appends as their own op
For "add a new paragraph at the end" the agent uses a dedicated append op. No find string to get wrong — just the new content tacked onto the doc body. The same protection applies: nothing else can shift.
The review queue shows the patch, not a diff
When a patched edit waits in review, the queue renders it as a compact find/replace card — the matched span on one side, the new content on the other. You see exactly what's about to change. No line-noise diff to read; the targeted-ness of the edit shows on the page.
Full rewrites still possible, when you want them
For a wholesale rewrite — a doc that's growing into two — the agent uses the regular update op with a complete new body. That option's there when you ask for it. The agent uses the surgical patches by default because they fit how editing actually works: a small change, somewhere specific.