Smart Sentence Retriever

STAR Summary

Situation

I wanted a fast way to retrieve sentences that answer a question even when the wording doesn't match exactly.

Task

Build a semantic-search demo that retrieves and ranks relevant sentences even when a question uses different wording.

Action

• Cleaned Project Gutenberg text, split it into sentences, and precomputed embeddings for the fixed corpus.
• Compared several embedding models on a sample subset and balanced clustering quality against model size and serverless deployment cost.
• Deployed the selected model behind an immutable AWS Lambda alias and a same-origin IAM proxy, with a browser demo that checks `/health` before ranking top-k matches by cosine similarity.

Result

• Shipped a working semantic-search demo for Alice in Wonderland that returns ranked sentence matches instead of keyword hits.
• Kept the site embed usable with warm-up status, inline similarity bars, and an open-in-new-tab fallback for a larger view.