RAG means retrieval-augmented generation: retrieve relevant material, place selected context into an AI request, and let a generative model produce a response. A local RAG desktop workflow keeps preparation and retrieval on your own computer by default. It does not make the generation step local unless the separately connected model also runs locally.
A packaged desktop app can reduce setup work compared with assembling parsers, embedding models, vector storage, search tools and connectors independently. The tradeoff is that the buyer must validate the product's supported platforms, file behavior, update path, licensing connection and integration boundaries rather than assuming every RAG component is interchangeable.
Which parts of the retrieval workflow run locally by default?
Preparing files, indexing, semantic and exact search, and memory run without a cloud AI provider, and in the default workflow files, prepared output, vectors and memory stay on storage you choose.
Central Brain prepares PDF, DOCX, TXT, Markdown, CSV and XLSX files, with optional OCR for scans and photos when it is included in the build. It keeps the source path, prepared-file path, page and chunk with results where available. These capabilities form the retrieval side of a RAG workflow.
The license is re-checked online every 14 days (Team and Enterprise can get a build with no online check), and optional cloud connections and cloud-hosted AI clients need internet. Keygen receives licensing and device data, not documents. Optional cloud embeddings or synced storage add further boundaries, and when a connected AI tool retrieves context, that provider receives the selected context and produces the response.
What should Windows and Mac users verify before adopting it?
Verify the operating system and processor, the installed build, project size, file mix, storage capacity, OCR needs and the AI tool you intend to connect.
Central Brain is built for Windows and Apple silicon Macs. A universal RAM, disk or operating-system-version minimum is not stated because needs vary with project size, file types, OCR and release. Keep enough storage for originals, prepared output and the local index, and test with the machine and files that will actually be used.
Supported AI clients differ by platform: Claude Desktop, Claude Code, GitHub Copilot and the ChatGPT desktop app (Chat, Work and Codex modes) on Windows and Apple silicon Macs. Capabilities vary by client. Use the canonical getting-started guide for current setup information, and do not assume that a connection demonstrated on one platform behaves identically on another.
How is a desktop retrieval app different from an AI model?
The retrieval app finds and supplies project context; the AI model interprets that context and writes the response.
This separation helps diagnose problems. If the expected passage is absent, investigate the scope of the indexed files, preparation, indexing and the search query. If the correct passage is present but the response is wrong, investigate the prompt, selected context, model behavior and review process. Treating both stages as one black box makes failures harder to correct.
Central Brain also keeps project memory in five note types (decision, fact, question, todo and summary) plus a project brief, stored as readable local files. Those records support continuity, but a connected tool must call the memory tools, and people should review important records.
What does a useful desktop RAG acceptance test include?
Test installation, preparation, changed files, semantic and exact retrieval, source opening, memory continuity, connection boundaries, backup and recovery.
Prepare a representative project containing supported text documents, a spreadsheet, a scan if OCR matters, and at least one updated file. Ask known-answer questions and confirm the expected sources. Stop and restart the app, then verify that the project and index remain usable. Save and recall a non-sensitive test decision in a new session.
Next, disconnect the network and confirm which actions still work: preparing files, indexing, search and memory run without a cloud AI provider, while licensing, optional cloud connections and cloud-hosted AI clients need a connection. Reconnect only the intended services. Back up the project locations, restore them in a controlled test, and record the recovery procedure. A successful first query is only one part of operational readiness.
When is Central Brain a reasonable fit?
It is a reasonable candidate when project knowledge lives in files, local-first preparation matters, source inspection is required, and a supported desktop workflow is acceptable.
It may fit an individual or team that repeatedly searches the same project files and wants retrieval plus durable project notes. It may not fit a browser-only environment, an unsupported processor or operating system, a workflow that depends on an unverified integration, or a deployment whose governance requirements have not been evaluated.
The product page remains the canonical description of current capabilities. This article is a selection framework. For a document-level test, continue to the source-traceable retrieval guide.
Validation checklist
What should you verify before relying on the workflow?
- Confirm Windows or Apple silicon Mac support for the installed build.
- Measure storage needs using a representative project and its prepared output.
- Test preparation, exact search, semantic search, source traceability and updates.
- Separate retrieval failures from connected-model response failures.
- Document network, licensing, optional-connection, backup and recovery boundaries.
Common questions
Does Central Brain run on Windows and Mac?
Central Brain is a desktop app for Windows and Apple silicon Macs. Supported AI clients: Claude Desktop, Claude Code, GitHub Copilot and the ChatGPT desktop app (Chat, Work and Codex modes) on Windows and Apple silicon Macs. Capabilities vary by client.
Is Central Brain itself a generative AI model?
No. It prepares, indexes, searches and supplies project context. A connected AI tool generates the response.
Can the default retrieval workflow run without a cloud AI provider?
Yes. Central Brain runs air-gapped: preparing files, indexing, search and memory need no network connection. The license is re-checked online every 14 days, and Team and Enterprise plans can get a build with no online license check. Optional cloud connections and cloud-hosted AI clients need internet.
Does local RAG guarantee a correct answer?
No. Retrieval can miss or mis-rank material, source content can be wrong, and a connected AI tool can misinterpret context. Important outputs require source review.
Next steps
- Getting started: current installation and connection guidance
- Central Brain product page: the canonical product and feature description