See the work. Learn the whole product.

Begin with a verified local setup.

The current public path is source installation. Light desktop packages for macOS arm64 and Windows x64 are implemented, but remain private candidates until platform signing and clean-machine acceptance pass.

1. 1
   Install the current prerequisites.

   Use Node.js 24 or newer and Ollama for the source-development flow, then clone the public repository.

2. 2
   Let setup guide the model choice.

   Run npm install, then npm run setup. Downloads require an explicit confirmation.

3. 3
   Check before launching.

   Run npm run doctor. Fix a reported model or storage problem instead of guessing.

4. 4
   Open the private launch URL.

   Run npm run dev and use the one-launch loopback URL printed locally. Do not share it.

Choose a model that fits the machine and the job.

Source installs use the documented Ollama setup and choose the active chat model with OLLAMA_MODEL. The desktop candidate adds an in-app Model Manager that distinguishes reviewed chat models, Knowledge embeddings and unqualified installations.

1. 1
   Choose a reviewed source model.

   Use npm run setup, or set OLLAMA_MODEL in .env.local when you already know the reviewed model you intend to run.

2. 2
   Verify before launching.

   Run npm run doctor, then restart the source server after changing the active model.

3. 3
   In the desktop candidate, open Model Manager.

   Select the model status to compare intended use, approximate download size, memory guidance and review state.

4. 4
   Reuse weights or download deliberately.

   The candidate can use reviewed Ollama weights in place—without copying them—or install a reviewed model only after your explicit click.

Ask normally. Use the mode only when the boundary changes.

Most work belongs in Smart mode: it stays local and can retrieve relevant Knowledge Vault material. Local only avoids retrieval; Teacher requires vault-grounded teaching. Codex is displayed but cloud handoff is disabled.

1. 1
   Start with the outcome.

   Say what you need, who it is for, and any format or length constraint. The current turn outranks older context.

2. 2
   Use a starter when useful.

   Explore an idea, Build something, Write an email or Create a document pre-fills a strong opening—not a hidden workflow.

3. 3
   Keep continuity explicit.

   Use Reply when a follow-up depends on one earlier message. Reopen a local chat when the whole conversation matters.

4. 4
   Stop work you no longer need.

   The Send control becomes Stop during generation. Cancelled or failed partial text stays out of later prompts.

Make the resting screen yours without creating a secret profile.

Preferences controls the optional greeting name, fresh-chat content, appearance and colour palette. These choices stay separate from reusable model memory.

1. 1
   Open Preferences in the header.

   Add a preferred name only if you want it used in the greeting.

2. 2
   Choose one calm welcome mode.

   Use Mix, Quotes, Jokes, Thoughts or My books. The offline collection avoids the recent 60 items.

3. 3
   Set appearance separately.

   Choose light or dark and one of the maintained colour environments, including the canonical Rangabot palette.

4. 4
   Save once.

   The next fresh chat rotates the greeting and welcome line without adding controls to the home state.

Keep the sidebar small, then retrieve work when you need it.

Chats remain local and searchable. Projects scope new conversations; pinning keeps an important title close; Markdown import and export provide a readable transfer format.

1. 1
   Create a project before the work begins.

   Enter a project name in the sidebar, select it, then choose New chat. That new conversation is project-scoped.

2. 2
   Return through titles, not folders.

   Search all local conversation titles and content, or select a project to filter the list.

3. 3
   Pin sparingly.

   Pin only active reference chats; reveal delete and pin actions when needed so the title stays primary.

4. 4
   Transfer readable conversations.

   Export the open chat as Markdown. Import a valid file into the active project when you need text and reply references elsewhere.

Approve what should persist. Correct it in one place.

Memory is an explicit local list. Rangabot selects only relevant titles for the current request and excludes saved preferences that conflict with what you just asked for.

1. 1
   Open Tools → Memory.

   Choose Preference, Fact I provided or Standing instruction.

2. 2
   Write one atomic item.

   “Prefer concise answers with one practical example” is easier to review than a mixed personal biography.

3. 3
   Approve and inspect.

   Edit or delete an item whenever it changes. Nothing is inferred into this list from normal conversation.

4. 4
   Move memory carefully.

   Export JSON for portability. Import first creates a preview of new items, duplicates and conflicts; existing items win unless you explicitly replace them.

Turn local documents into cited, inspectable teaching.

The 4 GB Knowledge Vault accepts text-based PDF, DOCX, HTML, Markdown and text files. Smart mode can retrieve automatically; Teacher mode requires a stricter vault-grounded answer.

1. 1
   Add material through the current source flow.

   Place compatible files in data/knowledge/inbox/, then run npm run knowledge:ingest. This step is not yet an in-app file picker.

2. 2
   Check Vault before asking.

   Open Brief → Vault to confirm document, passage, storage and incompatible-file status.

3. 3
   Ask for the evidence shape.

   Request a comparison, explanation or synthesis and tell Rangabot whether to stay strictly inside your documents.

4. 4
   Inspect the answer boundary.

   Look for Hybrid or Keyword only status, passage citations, and explicit separation between vault claims and model background.

5. 5
   Use Knowledge Brief for orientation.

   Open Discover for surfaced ideas, This week or This month for bounded recaps, Vault for library health, and Rangabot updates for the local product changelog.

Let a normal question use a calculation—without hiding the query.

Rangabot supports explicitly approved CSV, Parquet and DuckDB files up to 100 MB. The chat can propose bounded read-only SQL, or you can inspect and run one exact query manually.

1. 1
   Open Tools → Analyze in the desktop app.

   Use the native picker to choose a CSV, Parquet or DuckDB file, approve it, select the dataset, and choose Use selected data in chat. The source-only browser panel cannot select a new local file.

2. 2
   Add context only when it helps.

   Teach Ranga about this data can save local descriptions, aliases and confirmed joins against the exact dataset identity and revision. It is optional and grants no new access or operation.

3. 3
   Ask the business question.

   Use ordinary language. Non-analytical messages do not touch the attached data.

4. 4
   Review the proposed calculation.

   For manual SQL, choose Review query and inspect read-only access, row/time limits and the dataset fingerprint before Run once.

5. 5
   Open “How this was calculated.”

   Check the query, dataset, returned rows, duration and fingerprints; do not accept a plausible sentence in place of a correct grain or denominator.

Attach narrow code context—or turn the conversation into a real Word file.

Builder does not crawl a machine in the background. Repository access is explicit and revocable; Word creation gathers a brief in chat and validates the resulting local file.

Inspect progress without inflated claims.

Path to Mastery separates code that exists, criteria that passed and a capability that has actually been earned. Version 1.2.0 also returns eligible failed requests as reviewable drafts after their local bindings revalidate.

1. 1
   Open Tools → Mastery.

   Read each capability state, dependencies and criterion-level evidence instead of treating a merge as proof of quality.

2. 2
   Open one capability audit.

   Check the current score, named evidence, open criteria and the exact work required to advance.

3. 3
   Use Evidence for exact limits.

   Check the named suite, model, build and denominator before repeating a result as a product claim.

4. 4
   Recover deliberately.

   Choose Restore request only when useful, review the draft, then press Send yourself. Failed partial output is excluded; changed resources are rejected; nothing reruns automatically.

5. 5
   Rate only on an eligible build.

   Helpful and Needs improvement are optional, reversible and local. Ordinary or mixed source builds keep the control disabled.