Semantica

A browser extension and digital literacy tool to help you scrutinise and understand the media you consume better.

Implementation plan

Features

• News sources: using a local-based LLM (maybe Ollama?) to highlight phrases that suggest, imply, or carry additional meaning beyond their literal wording (e.g. emotive language)
• News sources: fact-checking?
• Support for thinking models — toggleable?
  • Ollama might return errors if we ask for thinking on non-thinking models; I’ll need to figure out a solution to either identify the model used or gracefully handle errors

Devlog

9 June 2025

• Started on the pop-up UI for displaying analyses
• Identified an issue where the analysis displayed on the extension popup is essentially a race — the tab that gets loaded last will have its analysis (i.e. the latest analysis) being displayed
  • I entirely forgot that I haven’t implemented any caching for the analyses. In other words, if you have multiple tabs open, the tab that loads the latest will override all other tabs to be analysed
  • Implemented temporary caching of analyses under a Record, but more work needs to be done to actually properly save the analyses under local storage
• Identified an issue where the popup content won’t be changed if the popup is open; the popup has to be re-opened for the updated content to populate

5 June 2025

• Trying to understand the core concepts of browser extensions that seem to be applicable to both Chrome and Firefox:
  • Pop-ups (the easiest to understand): HTML pages that show when you click on the extension icon in the extensions menu (usually the top-right corner of the browser)
  • Content scripts: files that run in the context of web pages — have the ability to interact, read, and alter the DOM
  • Service workers (also background scripts): ephemeral (loaded when needed, unloaded when dormant) scripts that run; has no access to the DOM, but can use messaging to communicate with other components (i.e. content scripts) if needed
• Came across a common problem when attempting to scrape news sites: how do we only get the main content, instead of everything else (e.g., recommendations, ads, whitespaces) in the way?
  • Experimented with @mozilla/readability — had the thought that since WXT already uses Node, this may be used
  • @mozilla/readability works! All hail open source
• Turns out that Hugging Face Transformers works! Came across an issue with authentication with a private model (Gemma 3) though, so will KIV for now and stick with local models with Ollama
• CORS once again comes to ruin my life :’)
  • Ollama by default limits the “Access-Control-Allow-Headers” header which might result in a failed connection because of CORS. The solution is to set these two environment variables in the system you have Ollama installed in:
    • OLLAMA_HOST to 0.0.0.0
    • OLLAMA_ORIGINS to * (or more specific if you’d like)
• Prompt engineering has been super crucial in ensuring that the model understands its responsibilities, though I’m concerned that more distilled models (like Gemma 3 8B) might not entirely understand them
  • From experiments so far, thinking models (like Qwen 3 8B) seem to perform a lot better — the thinking process likely meant that these models could understand the instructions provided to them deeper
  • Ollama’s API accepts a format value that specifies the expected JSON format the model should return. This blog post about format helps a lot, and it’s even more helpful that Zod can be used to define a schema and convert that to the JSON schema format that Ollama expects

2 June 2025

• Started checking out WXT — seems like a pretty cool batteries-in-box solution for creating browser extensions
• Tested if Ollama (at port 11434) can be reached from the extension, and so far seems successful
• To consider: should we just make things easy and use commercial LLMs (e.g., GPT, Gemini, or Claude)?