AGENT SKILL

rag-parse

Process/ingest the documents in /work/data/input into the knowledge stores (Obsidian vault + vector RAG index). Incremental: only new or changed files are processed, and files removed from input are pruned from the stores. Use when the operator wants to load, process, parse, ingest, index, or update their data. Triggers: parse, parse the files, process the input, ingest, index the documents, update the store, re-parse, load my data, refresh the index.

License 许可
MIT
所属仓库
castellotti/glove
最近更新
2026-11-11
Stars
★ 1
Forks
0

01 前言 Front Matter

name
rag-parse
description
Process/ingest the documents in /work/data/input into the knowledge stores (Obsidian vault + vector RAG index). Incremental: only new or changed files are processed, and files removed from input are pruned from the stores. Use when the operator wants to load, process, parse, ingest, index, or update their data. Triggers: parse, parse the files, process the input, ingest, index the documents, update the store, re-parse, load my data, refresh the index.

03 SKILL.md 正文

Parse input into the knowledge stores

Turn everything in /work/data/input into both knowledge stores with one incremental, deterministic command. It extracts text + source locators, (re)builds the vector index, and writes Obsidian source-pages — skipping unchanged files and pruning ones the operator removed. It reads PDFs (OCR'ing scanned ones), text and Markdown files, and images (OCR'd with tesseract).

kstore sync

Then:

  • Report the summary line it prints (new / changed / unchanged / removed, OCR'd, vault pages).
  • If it lists files as needs OCR (scanned PDFs, and ocrmypdf was not available), OCR each into the cache and re-run, e.g.: ocrmypdf --skip-text "/work/data/input/FILE.pdf" /work/data/store/cache/FILE-ocr.pdf then kstore sync. (In this sandbox ocrmypdf is installed, so sync normally OCRs scanned PDFs automatically.)
  • Never write extracted text, OCR output, or page images to /work/data/output. That directory is for final deliverables only. All derived and transient artifacts belong under /work/data/store (the tools handle this).

Optional: richer Obsidian synthesis

sync populates the vault with one source-page per document (fast, mechanical). For a cross-linked knowledge graph with claim ledgers and Maps of Content, the operator can additionally ask for synthesis — run /skill:wiki-ingest on specific documents. That is a heavier, LLM-driven layer on top of the source pages; only do it when asked, and a few documents at a time (it consumes context).

04 同仓库更多 Skills