Docs Search

Lightweight search utility and usage guide for the Toss / Apps-in-Toss llms-full.txt documents. It retrieves snippets from three sources with simple keyword + similarity scoring.

Sources

• Apps-in-Toss Developer Center: https://developers-apps-in-toss.toss.im/llms-full.txt
• Toss Mini TDS Mobile: https://tossmini-docs.toss.im/tds-mobile/llms-full.txt
• Toss Mini TDS React Native: https://tossmini-docs.toss.im/tds-react-native/llms-full.txt

Files

• requirements.txt: Python dependency (requests).
• skills/docs-search/core/docs_search.py: CLI searcher that downloads and caches llms-full.txt files, then ranks paragraphs.
• skills/docs-search/core/korean_similarity.py: Korean-aware similarity utilities.

Installation

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Usage

python -m core.docs_search -q "Unity WebGL optimization" -k 5

• -q / --query: search query (required)
• -k / --topk: number of results to show (default 5)
• --refresh-cache: bypass cache and re-download all sources
• Output: [index] <source URL> - <heading> / score: <score> / snippet

How it works

• Downloads each llms-full.txt and caches under <temp>/apps-in-toss/skills/.cache/ (cross-platform: /tmp on Linux/macOS, %TEMP% on Windows).
• Splits by blank lines into paragraphs and attaches the most recent heading (#, ##, etc.) as context.
• Ranks by keyword overlap plus SequenceMatcher similarity; returns top results.

Examples

• Quickly find where an API or guide is described.
• Search optimization/porting/auth/payment references with queries like “Unity WebGL performance”.
• Open the returned source URL in a browser for full context.

Notes

• Searches only the text contained in llms-full.txt; it does not crawl linked pages.
• Requires network access; deleting the cache triggers re-downloads.

Core workflow

1. Fetch & cache: download each llms-full.txt and store under <temp>/apps-in-toss/skills/.cache/ (reuse cache when present).
2. Chunk: split by blank lines, tagging each paragraph with the most recent heading.
3. Score: keyword overlap (weight 2) + similarity; Korean text uses korean_similarity, otherwise SequenceMatcher.
4. Rank & output: sort by score, return top‑k snippets (source URL, heading, score, snippet).

from core.docs_search import search

results = search("Unity WebGL optimization", topk=5, refresh=False)
for item in results:
    print(item["source"], item["heading"], item["score"])
    print(item["snippet"])

Requirements

• Python 3.9+
• requests>=2.31.0 — HTTP 요청 및 llms-full.txt 다운로드