Instruction precedence
Safety and the current request always outrank stale context, memory and model defaults.
Rangabot coordinates private models, approved knowledge, selective memory and bounded local tools—then exposes enough of the process to earn trust.
A model-independent control plane compiles the current request, recent context, relevant approved memory and capability boundaries into one inspectable answer contract.
Safety and the current request always outrank stale context, memory and model defaults.
Only useful recent conversation is carried forward, avoiding unlimited prompt growth.
Unavailable actions and live information are stated honestly instead of being simulated.
Memories are explicit local records with origin, confidence and controls. A normal conversation does not silently become a permanent fact.
The user decides what becomes reusable memory.
Unrelated personal facts stay out of the prompt and response.
Review, edit, export, supersede or delete local memory.
The Knowledge Vault extracts supported documents locally, combines keyword and embedding search, and gives the model bounded source passages for teaching and synthesis.
Keyword precision and semantic similarity work together with a visible fallback.
Answers cite the document, passage and available page or section context.
Vault-supported claims remain distinct from downloaded-model background.
With explicit read-only access, Rangabot identifies when computation is required, compiles a constrained analytical plan and explains the verified local result.
A file must be allowlisted and attached to the current conversation.
Only validated read-only SQL runs inside local DuckDB resource limits.
Dataset, query, row count, timing and fingerprints stay available for review.
Rangabot searches only approved folders, skips common secret and dependency paths, and attaches bounded file previews to local-model chat.
Approvals are explicit, reversible and stored only on the local device.
There is no hidden background repository index.
Word creation gathers requirements in chat, validates output and can render local previews.
Available, experimental and planned abilities are labelled separately. Visit Evidence for exact results or Path to Mastery for criterion-level progress.