Research Local Corpus

Query a local research corpus via wiki or qmd.

Skill metadata

Source	Bundled (installed by default)
Path	skills/research/research-local-corpus
Version	1.0.0
Author	VibeOS
License	MIT
Platforms	linux, macos, windows
Tags	research, rag, wiki, qmd, corpus, local
Related skills	llm-wiki, professional-research, qmd, obsidian

Reference: full SKILL.md

信息

The following is the complete skill definition that VibeOS loads when this skill is triggered. This is what the agent sees as instructions when the skill is active.

Research Local Corpus Skill

Reuse prior research without a new memory provider. Prefer a curated wiki or local hybrid search over re-crawling the web every turn.

When to Use

• User says “from my notes / wiki / research folder / previous briefs”.
• Repeating analysis over a saved corpus under ~/.vibeos/research/ or WIKI_PATH.
• Building a compounding KB after several vibeos research runs.

Do not use for:

• First-time web research → professional-research.
• Cloud docs in Notion/Drive → those MCPs first.

Prerequisites

Pick one path (install only what you need):

Path	Setup
LLM Wiki	bundled skill llm-wiki; set WIKI_PATH or use ~/wiki
qmd	vibeos skills install official/research/qmd (macOS/Linux)
Research workspaces	already created via vibeos research run → ~/.vibeos/research/

How to Run

1. Orient — list the corpus root and recent artifacts.
2. Retrieve — wiki index / qmd query / search_files over research dirs.
3. Synthesize — write a cited brief; cite local paths as sources.
4. Ingest (optional) — file durable conclusions back into the wiki (llm-wiki).

# Workspaces from CLI
ls "${VIBEOS_HOME:-$HOME/.vibeos}/research/" | head

# Wiki (if configured)
WIKI="${WIKI_PATH:-$HOME/wiki}"
test -f "$WIKI/index.md" && head -40 "$WIKI/index.md"

Decision tree

Situation	Prefer
Small curated KB, cross-links matter	llm-wiki
Large markdown dump, semantic recall	qmd
Single past brief / compare runs	open ~/.vibeos/research/<id>/
Notes in Obsidian vault	Obsidian MCP or vault path + search_files

Procedure

1. Never invent file paths — verify with search_files / read_file.
2. Treat local extracts like web sources: grade confidence, note staleness.
3. When promoting web findings into the wiki, keep raw/ immutable (llm-wiki).
4. For spreadsheet tables in the corpus, use execute_code or Google Sheets MCP (spreadsheet-research routing below) — not pixel OCR first.

Spreadsheet routing (no Excel COM core tool)

Need	Prefer
Google Sheets	vibeos mcp install google-workspace → read_sheet_values
Local CSV/TSV/XLSX	execute_code + pandas/openpyxl; or quantitative-analysis
Visible Excel UI only	computer_use after choose-interaction-surface (last resort)

Pitfalls

• Do not add an in-tree memory provider for this — policy forbids it.
• Do not dump the whole corpus into the prompt; retrieve then read.
• Keep prompt-cache safe: do not swap toolsets mid-session.

Verification

• Corpus root exists and was listed.
• Every claim cites a real local path or prior brief id.
• Confidence states whether the corpus may be stale vs live web.