Evidence Desk

Find the passage. Keep the source.

Search six fictional work notes about equipment, bookings and refunds. Each result opens in the note it came from.

Story LabSee why the source matters
47 / 800

Optional model download: about 38 MB. Searches run in your browser.

Example loaded. Open the highlighted passage or ask your own question.

Passages

3 ranked passages
Text matching

Ranks passages by shared words. Lower results may be off topic.

  1. 01Word overlap 0.239
    “If a borrowed camera is damaged, take photographs of the damage and label the kit before asking the equipment coordinator to inspect it. Do not return it to the available shelf.”
    Camera returns2026-01camera:1
  2. 02Word overlap 0.200
    “An organiser may transfer a ticket to another named attendee after recording the original and new names. This is a ticket change, not a refund.”
    Ticket refunds: current note2026-02 currentrefund-current:2
  3. 03Word overlap 0.177
    “A projector that shows no picture should first be checked for a loose video cable and the selected input. If neither fixes the picture, stop the setup and contact the room technician.”
    Projector setup2026-01projector:1

The file includes your question, matches and their source versions.

Original note

Full text
SOURCE 012026-01

Camera returns

camera

camera:1If a borrowed camera is damaged, take photographs of the damage and label the kit before asking the equipment coordinator to inspect it. Do not return it to the available shelf.

camera:2A working camera can be checked back in after its battery, lens cap and carry bag have been counted. Record any missing item in the return note.

Memory only · clears on reload
Search methods

How the desk finds a match

Start with text matching. Try search by meaning when your question uses different words from the notes.

How the two methods performed +

On 13 answerable held-out questions, semantic search accepted 7 and abstained on 6. On 3 no-answer questions it returned 0 false positives and 3 abstentions. Text matching accepted all 13 answerable questions and all 3 no-answer questions.

Ungated rankings · before the threshold
MethodRecall@3MRR
Semantic0.9230.827
Text matching0.5380.515

Recall@3 is the mean fraction of labelled relevant passages found in the first three; MRR uses the first relevant passage in the full ranking. Both exclude no-answer questions and do not imply a question was accepted. These are Node CPU evaluation results, not browser performance measurements.

The 0.5773 threshold was selected on 8 separate calibration questions, never the held-out set. Six fictional sources and 12 passages are a tiny test, not evidence of general accuracy.

Download raw evaluation ↗ · Frozen test set ↗
What stays in your browser +

Notes, questions and vectors live in memory. They are not sent to a service, logged or saved in application storage. Reset clears the workspace; reloading clears your edits. Public model files download from this same site only when requested. Your browser may cache those public files.

Export deliberately creates a file containing your question and retrieved quotations. Keep it wherever you normally keep those notes.

Model details and note versions +

MiniLM-L6-v2, q8 weights, 384 dimensions, mean pooling and normalized vectors. Transformers.js 4.3.0 runs ONNX WASM in a dedicated worker with one thread. Every question and passage is checked against the model’s 512-token limit; nothing is silently shortened.

Archived and current versions stay separate. Retrieval does not resolve contradictions or establish which statement is true. Editing sources invalidates results and unloads the model to clear pending work.

Pinned asset hashes ↗ · Architecture ↗