Overview

Following Previous Post: #7, this is the 8th cycle. It covers two URL classifier quality fixes and one taxonomy-path bug found while running the multilingual blog. Three small fixes, but all three touch surfaces a user notices instantly — classification accuracy and build success.

URL Classifier — Perplexity Computer + Billing Page Filter

Background

Perplexity’s newly launched Computer Agent URLs were appearing in extracted history, but the classifier was dropping them into generic web_page. At the same time, RunPod billing/account pages were leaking personal account data while still being classified as a “tech URL.”

Implementation

Two changes to url_classifier.py’s dispatch table:

• Added Perplexity Computer agent pattern → ai_chat_perplexity
• Moved explicit billing patterns (/billing, /account) into the default-filter set

Result

Billing pages no longer surface in history extracts, and Perplexity Computer results now classify correctly into the AI chat group.

URL Classifier — YouTube Shorts and GitHub Accuracy

Background

The classifier didn’t detect YouTube Shorts (youtube.com/shorts/...) and dropped them into generic web_page. Separately, GitHub URLs containing /blob/, /tree/, or /commit/ were misclassified as github_repo, causing the fetcher to try and pull a README that wasn’t there.

Implementation

# Add YouTube Shorts pattern
re.compile(r"youtube\.com/shorts/"): UrlType.YOUTUBE,
# Strengthen GitHub repo pattern — split blob/tree/commit into github_other

Result

YouTube Shorts are now valid transcript fetch targets, and GitHub /blob/ URLs are split into a separate category (github_other), eliminating fetcher failures.

Multilingual Taxonomy Path Bug

Background

After the blog moved to bilingual (Korean/English), creating a new tag wrote _index.md to content/tags/{tag}/_index.md. But in multilingual mode, Hugo expects content/{lang}/tags/{tag}/_index.md. Result: every new tag 404’d on the English homepage.

Implementation

graph LR
    A[Old: content/tags/{tag}/] --> B[Hugo i18n <br/> cannot resolve]
    A --> C[New: content/{lang}/tags/{tag}/]
    C --> D[Routes correctly]

The taxonomy _index writer in image_handler.py now accepts the --language flag and writes under language_content_dirs[lang]. When the publish command runs both languages, both _index files are created.

Result

The previously-broken tag index on the English blog now works, confirmed live after a cache bust.

Commit Log

Insights

All three fixes were of the form “what assumption did we wrongly model about another component (Chrome history DB, Hugo i18n)?” The URL classifier is essentially a set of assumptions about external site URL schemas — when a new pattern like Perplexity Computer ships, those assumptions go stale immediately. The multilingual taxonomy bug was more interesting: Hugo’s multilingual mode keeps the default content directory in place, but taxonomy indexes have to move under the language root. That asymmetry is exactly the kind of thing that breaks first at the single-language → multilingual transition.