Code · open source
Personal Wiki
An offline AI research assistant over my own energy research and class notes. It answers with cited sources, and nothing ever leaves my laptop.
- 8 / 8cited-answer tests passed
- 685passages indexed
- 4.5–9.6 sper answer, fully offline
How it works
A local Gemma model (4-bit, MLX) runs on Apple Silicon and loads from local files only. Retrieval is hybrid: BM25 keyword search plus bge-small embeddings, merged with reciprocal rank fusion. It indexes 41 of my notes into 685 passages and generates 60 linked wiki pages across 7 topics. I wrote 8 offline tests (cited answers and correct refusals) and iterated 8 times until all passed. The failures got fixed by better retrieval, not a bigger model. Offline operation is verified by blocking all network access.