Search Your Japanese Brain in Romaji

The Problem

A developer on a Japanese team stores lessons in Japanese every day:

learn_from_attempts:
  topic: "deploy:k8s"
  whatWorked: "readinessProbeのfailureThresholdを10に増やすことで解決"

Three days later, they try to recall it. They type fast, without switching input method:

smart_recall("readiness probe failure threshold") → ✅ found
smart_recall("depuroi") → ❌ nothing  (before v0.5.37)
smart_recall("デプロイ") → ✅ found

The lesson is there. But if you search in romaji — the Latin-alphabet transliteration of Japanese used when typing without a Japanese input method — it was invisible.

What We Built

Katakana → romaji tokenization at index time.

When cachly indexes a document containing katakana, it now also generates the Hepburn romaji equivalent as an additional search token.

"コンテナのヘルスチェックで問題発生"
  ↓ bigrams: コン, ンテ, テナ, ...
  ↓ romaji tokens: "kontena", "herusuche" (from ヘルス+チェック)

Now a search for kontena matches the stored katakana content.

The Converter

We implemented a full Hepburn romanization converter — 70+ mapping rules — handling:

Real Examples

# Stored lesson (Japanese)
"コンテナのヘルスチェックで127.0.0.1を使用する必要がある"

# These all find it now:
smart_recall("kontena")          → ✅ (romaji for コンテナ)
smart_recall("container")        → ✅ (cross-lingual synonym)
smart_recall("コンテナ")          → ✅ (direct katakana match)
smart_recall("healthcheck 127")  → ✅ (mixed language, Latin terms match directly)

# Stored lesson (katakana deployment note)
"デプロイメントが失敗、ポート3000を開放することで解決"

smart_recall("depuroimento")     → ✅ (full romaji)
smart_recall("depuroi")          → ✅ (partial romaji, fuzzy match)
smart_recall("port 3000")        → ✅ (Latin terms match directly)
smart_recall("deploy")           → ✅ (cross-lingual: deploy↔デプロイ)

Why Romaji Matters

Japanese developers switch input modes constantly. When typing fast in a terminal, in a commit message, or in a search query — romaji often comes out first. Making romaji search work means zero friction between "I need to find that lesson" and "found it."

This is especially true for katakana loanwords — the vast majority of technical terms in Japanese are written in katakana and have direct romaji equivalents.

Upgrade

npx @cachly-dev/mcp-server@latest autopilot