Inside BrainBox.
The components, setup choices, and ways to improve your local workflows.
Back to BrainBox ↗What’s in the box.
The software behind your local workspace. Your knowledge stays separate from the app, and you choose which models to download.
- gbrain Knowledge engine
- Imports and indexes your knowledge. PGLite stores the local search index, vectors, and document records.
- Present in the local Mac development bundle; the source installer includes an installation step.
- Embeddings Search by meaning
- A local embedding model turns passages into numerical representations for retrieval alongside keyword search.
- Default: nomic-embed-text. Model downloaded during setup.
- Ollama Local model runtime
- Runs the models used for embeddings and answers. Choose an answer model that fits your hardware.
- Runtime present in the local Mac development bundle. Models are downloaded separately.
- Web UI Your workspace
- Import files, search, ask questions, inspect sources, and explore a graph of saved document connections.
- Included. Opens locally in your browser or the Mac development shell.
- BrainBox harness Connects the pieces
- The local server connects retrieval, model calls, source links, reusable instructions, and saved feedback.
- Included. The current harness supports retrieval and Q&A; general workflow automation is planned.
- Skills & improvement Adapt to your work
- Save answer preferences and review feedback. Evaluate examples before deciding whether a model needs fine-tuning.
- Two instruction skills and feedback capture available. Automated evaluation, auto-research, and training runners are planned.
Add modalities as needed. PDF extraction, audio transcription, and image understanding need additional tools or models. The Mac development bundle does not include PDF or audio helpers.
Hardware and release status.
The connected app has been tested on an Apple silicon Mac. The native Mac shell is a development preview, not a signed public release. Linux, Windows, Raspberry Pi, and DGX compatibility has not been established by a complete installation test.
Make every answer
more useful.
Build repeatable workflows around sources you know. Review the results, refine the instructions, and check whether the next run improves.
- Ask or searchStart with a real question.
- Check the sourcesSee where the answer came from.
- Add your feedbackMark what helped or needs work.
- Refine skills & searchImprove the instructions and sources.
- Ask againCheck whether the answer improved.
Make the answer
work for you.
Lead with what matters, connect the answer to its sources, and explain the details that help you act.
Check the original source
“Start with the key point. Cite the original material. Include helpful detail when needed, and check the source before acting.”
Fictional note for this example.
Your feedback gives you a place to start the next review.
Save the way
you like to work.
A skill is a reusable instruction. Try these two preferences and see how the same sourced answer changes.
The source stays the same. Your instructions change how the answer is put together.
Explore the knowledge workspaceInteractive example. Feedback and preferences stay in this preview; nothing is saved or trained.
A guided path
to fine-tuning.
Start with better sources and reusable skills. If the model still needs to learn a consistent pattern, review a training plan with clear examples and checks.
Choose examples worth learning from.
Review the question, the source, and the answer you would want. Include only examples you approve. A thumbs-up can flag an example for review; it does not train the model.
This walkthrough previews the decisions. It does not run training or change a model. Get specialist help with evaluation and fine-tuning.
A few good
questions.
The idea behind BrainBox.
What is BrainBox?
BrainBox brings your specialized knowledge and local models into one workspace. Search, chat, and explore connections in your sources. The current harness supports retrieval and Q&A. General workflow execution is planned.
What can I try here?
This website is an interactive design preview. Explore how your documents connect to local AI, try feedback and skills, open the example source, or open the app preview. The app uses fictional documents, and the improvement tools are illustrative.
Does it run without the internet?
The intended core experience keeps ingestion, indexing, questions, and citations local. Offline behavior needs to be verified on the actual machine; this design preview does not demonstrate a live local model.
Where does my knowledge go?
The core idea is that your original files, their index, and your model stay on hardware you own. The app makes the selected runtime and privacy policy visible.