Give meaning a shape.
The open-weight MiniLM model turns your question and each passage into vectors. It runs locally with Transformers.js and ONNX WebAssembly.
FIELD EXPERIMENT 001 · GHOST IN THE MACHINE
You remember what a letter meant, but not what it said. Light a small AI lantern and follow that feeling back to the words.
Describe the idea. The words don’t have to match.
One initial download. Inference stays in this browser. No account or API key.
FOLLOW THE LIGHT
Similarity is a ranking signal, not a probability or proof. The cabinet can retrieve a passage; it cannot know an author’s intent.
THE READING ROOM
Local imports: .txt or .md, up to 8 files, 12,000 characters each and 40,000 total. Imports replace the sample archive and disappear when you reload. They are never uploaded.
OPEN THE DRAWER
The open-weight MiniLM model turns your question and each passage into vectors. It runs locally with Transformers.js and ONNX WebAssembly.
Cosine similarity ranks passages by semantic closeness. Searching “an unfinished goodbye” can find a letter that never uses those exact words.
Each result is an exact passage with a title and character offsets. There is no generated answer, séance persona, or reconstructed author’s voice.
Your queries and imported text stay in this tab. This app has no analytics or server API. Its static host receives ordinary page and asset requests, including the initial model download. A service worker caches the app and model for offline reuse when browser storage permits.
The bundled model is English-oriented. Its rankings can be wrong or unhelpful, especially for negation, uncommon names, long passages or other languages. We split text into overlapping passages; these are retrieval results, not factual verification.
The model and runtime are pinned and served alongside the app. You can inspect the source, checksums and licenses. Browser storage may be cleared or evicted; offline availability then requires another download.