把 Gemma 4 e2b 接成 Obsidian 的本地 Helper：摘要、翻譯、Tag 建議Wire gemma4:e2b Into Obsidian as a Local Daily Helper

Why not an existing Obsidian plugin?

The community has several Ollama integration plugins. I tried two. The issue isn't features, it's boundary control:

• Most ship "send the whole note as prompt" — no precise selection like "summarize just these three paragraphs."
• Prompts live in plugin settings; tweaking one means digging through a UI.
• Many add a model picker and temperature slider in the name of generality. For daily use that's noise.

What I want: select text → run one fixed task → paste result back. Three tasks, three shortcuts. No menus.

Architecture: 60 lines of shell + macOS Shortcuts

Obsidian: select text
   ↓ ⌘⌥1 / ⌘⌥2 / ⌘⌥3
macOS Shortcuts (captures the selection)
   ↓ Run Shell Script
~/bin/llm-helper.sh {task} {selection}
   ↓ curl
Ollama HTTP API (localhost:11434)
   ↓ JSON response
Paste back into Obsidian (replaces the selection)

Three tasks (summary / translate / tag) share one script, branched on the first argument. macOS Shortcuts handles "capture selection → pipe to script → replace selection with stdout."

Shell script

~/bin/llm-helper.sh:

#!/bin/bash
set -euo pipefail

TASK="${1:-}"
INPUT="${2:-}"
MODEL="gemma4:e2b"
ENDPOINT="http://localhost:11434/api/generate"

if [[ -z "$TASK" || -z "$INPUT" ]]; then
  echo "usage: llm-helper.sh {summary|translate|tag} <text>" >&2
  exit 1
fi

case "$TASK" in
  summary)
    PROMPT="請務必使用繁體中文台灣用語回答。請用三句話以內摘要下面這段文字，
直接回摘要本體，不要前綴、不要解釋：

---
$INPUT
---"
    ;;
  translate)
    PROMPT="請務必使用繁體中文台灣用語翻譯下面這段英文，
保留原文的技術名詞（用括號註原文），直接回譯文，不要解釋：

---
$INPUT
---"
    ;;
  tag)
    PROMPT="請務必使用繁體中文台灣用語回答。請為下面這段筆記建議 3-5 個
Obsidian tag（kebab-case，不含 # 字元，用逗號分隔），
不要解釋、不要其他文字、只回 tag 清單：

---
$INPUT
---"
    ;;
  *)
    echo "unknown task: $TASK" >&2
    exit 2
    ;;
esac

curl -s "$ENDPOINT" \
  -H "Content-Type: application/json" \
  -d "$(jq -nc \
        --arg model "$MODEL" \
        --arg prompt "$PROMPT" \
        '{model: $model, prompt: $prompt, stream: false, think: false, options: {temperature: 0.2}}')" \
  | jq -r '.response' \
  | sed -e 's/^[[:space:]]*//' -e 's/[[:space:]]*$//'

Five things worth pointing out:

1. All prompts have the Chinese prefix. gemma4:e2b defaults to English (prior article). Don't drop the line.
2. stream: false — the Shortcut needs a complete response in one shot.
3. temperature: 0.2 — summary, translation, tags all want stability, not creativity.
4. think: false — gemma4:e2b is a thinking model. Without this flag, the num_predict budget gets eaten by the internal reasoning pass first, and response comes back empty. A lesson you only need to learn once.
5. jq -nc to build the JSON — avoids shell-variable quoting hell.

macOS Shortcuts setup

Three Shortcuts, same shape, different argument:

1. New Shortcut → Quick Action
2. Receives: Text from Obsidian
3. Action 1: Get Text from Shortcut Input
4. Action 2: Run Shell Script

- Shell: /bin/zsh - Pass input: as arguments - Script: $HOME/bin/llm-helper.sh summary "$1"

1. Action 3: Replace Selection (with stdout)

Name them Obsidian Summary, Obsidian Translate, Obsidian Tag, then bind ⌘⌥1 / ⌘⌥2 / ⌘⌥3 under System Settings → Keyboard → Keyboard Shortcuts → Services.

Three tasks, measured

I picked three different input types per task from my own Obsidian vault (this repo's content/ folder doubles as the vault), ran 10 trials per type — 90 real Ollama calls across the three tasks, zero empty responses. Raw records: tests/outputs/functional_test_results.jsonl; aggregates: tests/outputs/functional_test_summary.json.

Much faster than the 8–20s I'd originally guessed — gemma4:e2b is only 5.1B and none of these three tasks need a long output. Tag format compliance was the shakier one: several of the 90 runs came back with a leading # or a stray English word, matching the observation in "Limits and what's next" below.

Summary sample

Input (excerpt from my own ARCHITECTURE.md):

此監看倉並非強迫所有站點使用同一個 SSG、theme 或 deploy 目標。
每個站點保留最適合自身內容的技術棧。共享的是 knowledge/ skills/
pipeline/ metrics/ n8n-workflows/；不共享 SSG 選擇、視覺識別、
部署目標、是否雙語。

Output (gemma4:e2b):

各站點可根據自身內容選擇最適合的技術棧，不強制使用統一的SSG、theme或部署目標。
僅共享知識、技能、管道、指標等資訊。SSG選擇、視覺識別、部署目標及雙語選項不共享。

Translate sample

Input: The unified memory architecture on Apple Silicon allows large models to run without dedicated VRAM, but inference throughput remains bound by memory bandwidth, not compute.

Output:

Apple Silicon 上的統一記憶體架構 (unified memory architecture) 允許大型模型在沒有專用
VRAM 的情況下運行，但推論吞吐量 (inference throughput) 仍然受限於記憶體頻寬
(memory bandwidth)，而非運算能力 (compute)。

Tag sample

Input (a note on RAG chatbot embedding retrieval):

今天研究了怎麼幫 RAG chatbot 做 embedding retrieval。用 nomic-embed-text 把每篇文件
切成 300-500 字的 chunk，算 cosine similarity 找出最相關的前 5 段，再把這些段落塞進
prompt context 讓 LLM 生成回答。測試發現 chunk 太短會失去上下文，chunk 太長又容易稀
釋掉關鍵句子的權重，抓 top-k 的 k 值也要跟 chunk 大小一起調。晚點想試試看 hybrid
search，把關鍵字比對（BM25）跟向量相似度加權混合，看看能不能提升召回率。

Output:

obsidian tag:
RAG
embedding retrieval
nomic-embed-text
chunking
context window
hybrid search
bm25

This run is a good demonstration of the tag task's problem — the prompt explicitly says "no explanation, no other text, just the tag list," and the model still prepended obsidian tag:, one-per-line instead of comma-separated, and never converted to kebab-case. This format drift wasn't a one-off across the 90 functional-test runs.

Compared to my previous Claude flow

The point isn't "Gemma e2b replaces Claude." It's offloading these three repeated chores so frontier-model budget goes to work that actually needs it (like writing the English version of this article).

Limits and what's next

• Long input: past ~3K tokens, summary starts dropping the back half. Manual chunking is the current workaround.
• Mixed-language tags: occasional English tag slips in. The "Chinese nouns only" instruction helps, doesn't fully fix.
• Shortcut selection misses: Obsidian's preview-mode selection behaves differently from source mode. Still working around this.

Next: extending this to a fourth task — "given this note, find related notes in the vault." That needs embeddings and gets meaningfully more complex; it'll be its own article.

FAQ

Q: Why not use Obsidian's Smart Connections plugin with e2b? A: Smart Connections is embedding-based. This article is generation. Complementary, not competing.

Q: Can I swap in a bigger model? A: Architecturally, yes — just change the MODEL line in llm-helper.sh. This machine only has gemma4:e2b (5.1B) installed; I haven't run a larger size, so I can't hand you real numbers on whether tag compliance improves with more parameters. That's an honest gap, not something I'm skipping out of laziness.

Q: Windows / Linux? A: The shell script ports directly. Replace macOS Shortcuts with AutoHotkey (Windows) or xdotool + a keybinding (Linux). Same shape.

Verdict

Waveform: STEADY_PULSE

Good fit if you:

• Use Obsidian and have enough notes for the chores to recur daily
• Have ollama already running, or are willing to spend 30 minutes setting it up
• Want to free frontier-model budget for work that actually needs it
• Care about privacy: notes never leave the machine

Skip if you:

• Notes routinely exceed 5K tokens (wait for the streaming/chunking version)
• Don't want to touch shell scripts or macOS Shortcuts

See gemma4:e2b on Ollama Library →