A developer explores using text and visual embeddings of their Medium blog posts to build a resume reference system for job hunting. Two embedding approaches are compared: extracting raw HTML text and using Gemma-4 31B to generate visual summaries from browser-rendered page images, then passing both through Gemini Embedding v2. UMAP visualization and cosine similarity analysis reveal the two approaches produce noticeably different distributions (similarity range 0.58–0.83), with shorter articles showing closer agreement. The author hypothesizes that lossy summarization of long articles, ignored code snippets, and distracting images contribute to the divergence, and concludes that a more sophisticated index-query pipeline may be needed.
Table of contents
BackgroundWhat this is and what this is notThe setupThe embedding processThe quality / agreement of the 2 ways of embeddingGet Stephen Cow Chau ’s stories in your inboxConclusion394 Impressions