A memory is more
than its missing words.
Find a passage. Read what was written. Leave room for what wasn’t.
Reading three fictional letters. Each nonempty line becomes a passage.
This demo uses three invented letters, not a real family archive. Searches run in your browser with MiniLM. On first search, library and model files download from jsDelivr and Hugging Face (approximately 50 MB). Questions are not sent to an inference API. No generated answers or voices.
Search to discover related passages.
Possible connections
Read the source
Select a passage to see its context.
What this search can — and cannot — tell you
Semantic similarity can point to related language. It cannot establish intent, historical truth, or whether a question has an answer. Even the highest-ranked passage may be irrelevant. The provisional 0.4 threshold is not a calibrated confidence score. Low scores do not prove that something is absent.
The “Waiting at home” example deliberately preserves a known retrieval failure. Read the alternative matches rather than trusting the first.
Model: Xenova/all-MiniLM-L6-v2, the quantized ONNX conversion of sentence-transformers/all-MiniLM-L6-v2 (Apache-2.0). English prototype. Questions remain in browser memory. The hosting provider receives page requests; model hosts receive file-download requests. Browser cache eviction can require another download. Offline reload is not guaranteed.