ReplyShelf: an open model finds the reply you already wrote

Description and use

ReplyShelf is a browser app that uses an open‑weight sentence‑embedding model to retrieve previously written support replies for paraphrased questions, enabling agents to quickly find, review, and copy trusted answers.

Problem

Support agents often need to reuse previously written replies for paraphrased questions, but finding the right reply becomes a separate task.

Outcome

The demo achieved 6 out of 8 correct matches on synthetic paraphrases, with semantic similarity outperforming simple Jaccard overlap in ambiguous cases. The app runs locally in a browser, requiring ~46 MB of assets and a modern WebAssembly‑capable browser.

Workflow

• Store a reviewed question‑answer card in the shelf
• Ask a new question in the app
• The model computes a semantic similarity score between the new question and all stored questions
• If the highest score exceeds 0.42, the app displays the corresponding saved reply and its source question
• If the top two scores differ by less than 0.06, the app flags ambiguity and shows alternatives
• If no score exceeds 0.42, the app withholds a ready‑to‑copy reply
• Agent reviews the displayed reply, copies it, and personalises it for the customer
• Agent can add new cards, export a JSON backup, or run a local comparison in the "Behind the shelf" view

Evidence

ReplyShelf gives those answers a second life.

Source

The core is Xenova/all-MiniLM-L6-v2, an Apache-2.0 open-weight sentence embedding model, running through Transformers.js 3.8.1 and ONNX Runtime’s WebAssembly backend in a Web Worker.

Source

I tested eight fixed synthetic paraphrases against the same eight-card shelf and compared the model with simple Jaccard word overlap. Both methods retrieved the expected top card 6 out of 8 times.

Source

The app withholds a ready-to-copy answer because the shelf has no close match.

Source