Aleksei Kokorev

Chrome extension · Python backend · Retrieval-augmented generation

What’s new in this article?

My contribution. I built a Chrome extension in JavaScript and a Python API that processes articles, retrieves relevant passages from the reader’s history, and compares them with the current article.

Reader outcome. The extension shows a short summary in Chrome and separates points covered in earlier reading from what the current article adds.

The problem

When you follow a topic across several publications, a new article often spends most of its time on familiar ground. A summary tells you what the article says. As a returning reader, I wanted to answer a more specific question: What does this add to what I’ve already read?

How the comparison works

Comparing the current article with earlier reading
Previously read articles The Python backend splits the text into passages.
↓ Encode each passage with Sentence Transformers
ChromaDB · Vector store Index the passages from the reader’s history.
↓ Search the index using passages from the current article in Chrome
Python API Retrieve relevant earlier passages and compare them with the current article text.
↓ Return the comparison
Chrome extension Article summary · Already covered · What this article adds

Searching reading history in ChromaDB

The Python backend splits previously read articles into passages, turns each passage into an embedding with Sentence Transformers, and indexes those embeddings in ChromaDB. When the reader opens a new article, the backend searches that index using passages from the current article. It uses the matching passages as context for the comparison.

Passage-level retrieval matters because two articles can cover the same subject while differing on one important detail. For example, an earlier article might announce a software release, while a follow-up repeats the announcement and adds a migration deadline. The comparison should surface the deadline even when most of the follow-up sounds familiar.

What “new” means

The reference point is the reader’s history. If the backend finds a related passage, it can compare what the two articles actually say. If it finds no matching passage, the narrower conclusion is “I couldn’t find this in your earlier reading.” That distinction matters because retrieval can miss a relevant passage.