vault-research

name: vault-research description: > Deep evidence research across the Obsidian vault using Keppi MCP tools and the Obsidian CLI. Use when wiki-search has been tried and is insufficient — raw meeting transcripts, comprehensive multi-note analysis, or when the wiki has no coverage. NOT for quick entity/concept lookups — use wiki-search for those. Triggers: "find everything about X across my notes", "what did we discuss in meetings about Y", "give me the full picture on Z", any question requiring 4+ raw note reads.

Vault Research (Deep)

Comprehensive vault research using Keppi MCP tools for graph navigation and note reading via the best available method. Use this when wiki-search is insufficient.

Keppi calls use MCP tools directly — not bash, not CLI. Note reading uses the best available tool in priority order — see Tool Availability below.

Tool Availability — Check First

Before starting, determine what tools are available. This controls the read strategy.

To check vault path (needed for read_file fallback):

Desktop Commander:start_process
  command: obsidian vault info=path
  shell: powershell.exe
  timeout_ms: 8000

Returns something like C:\Users\username\Documents\Obsidian Vault. Store this — you'll need it to build full paths for read_file.

When to Use This Skill

• Deep evidence retrieval from raw meeting transcripts
• Comprehensive analysis requiring 4+ note reads
• Questions that need specific quotes or step-by-step reconstruction
• The wiki has no coverage or is stale for this topic
• "Summarize everything across all my notes about X"

Try wiki-search First

If the question is about a known entity, person, project, or concept, try wiki-search first. It answers 80% of queries in 2-3 reads at 1/10th the token cost. Come here when wiki-search sends you — or when you know you need raw source depth.

The Karpathy Loop (Deep Research)

Even deep research participates in the promotion cycle. Every fact you assemble from raw notes that would answer a future query on its own belongs in the wiki.

1. Check wiki for map of the territory (sources, related pages)
2. Use Keppi to find entry points and trace connections
3. Read raw notes to answer the question
4. Promote key facts back to the wiki

Without step 4, the next researcher re-reads the same 8-16K tokens of raw notes. With promotion, the next query resolves at wiki-lane cost (~400 tokens).

Steps

Step 0 — Quick Wiki Check

Even for deep research, check the wiki first. It may have a synthesis that reframes what you need from raw notes. Wiki reads are ~400 tokens vs 3,000–8,000 for transcripts.

Desktop Commander:start_process
  command: obsidian read file=index
  shell: powershell.exe
  timeout_ms: 8000

If a relevant wiki page exists, read it. Use it as a map for the raw note reads ahead — the sources: frontmatter field tells you exactly which raw notes to target.

Step 1 — Find Entry Points

Try semantic search first — it finds conceptually related notes even when phrased differently, replacing 2-3 keyword iterations with one call.

Check embedding status (skip if already checked in Step 0 of wiki-search):

keppi:get_embed_status()

If ready_for_semantic_search is true, run semantic search across full vault:

keppi:semantic_search(query="<natural language description>", limit=5)

Use results with distance < 0.5 as blast_radius entry points (Step 2). If all distances > 0.5 or no results, fall through to keyword search.

Keyword fallback (use only if semantic unavailable or all weak):

keppi:keyword_search(query="<2-4 word topic>", limit=5)

Run at most 2 differently-worded queries. Semantic search eliminates the need to rephrase the same concept multiple times.

Step 2 — Trace Connections (Blast Radius)

keppi:blast_radius(note="<TopResultNote>", depth=2)

This gives you the structural map — which notes are connected and how strongly. related_to edges (weight 2.0) are more meaningful than tag overlaps (weight 0–0.5). Use depth=2; depth 3+ gets noisy and slow on large vaults.

Performance note: Avoid keppi:query_node unless you specifically need full inbound + outbound edge metadata. On a 2,600+ node vault it returns 150+ edges (large payload, slow). blast_radius is faster and gives you the structure you need.

Step 3 — Build the Context Pack

keppi:context_pack(query="<topic>", budget=8000)

Keppi selects the minimal set of notes fitting within the token budget, ranked by relevance to the query. This is your reading list.

Token budget guide:

No Desktop Commander / obsidian-cli? Stop here. The context pack itself contains the note content — synthesize directly from it without separate reads.

Step 4 — Read the Notes

Read each note from the context pack via the best available method, highest relevance first.

Option A: obsidian-cli via Desktop Commander (preferred)

CRITICAL: Always use shell: powershell.exe — not shell: cmd. CMD truncates arguments at the first space, breaking all filenames with spaces. PowerShell handles spaces correctly when using single-quoted arguments.

Single-word names:

Desktop Commander:start_process
  command: obsidian read file=NoteName
  shell: powershell.exe
  timeout_ms: 10000

Names with spaces — wrap the entire argument in single quotes:

Desktop Commander:start_process
  command: obsidian read 'file=Note Name With Spaces'
  shell: powershell.exe
  timeout_ms: 10000

Read by exact vault-relative path (also handles spaces correctly in PowerShell):

Desktop Commander:start_process
  command: obsidian read path=1-Projects/Acme/Some Note With Spaces.md
  shell: powershell.exe
  timeout_ms: 10000

Option B: Desktop Commander read_file (fallback — no obsidian-cli needed)

Use Desktop Commander:read_file with the full Windows path. Handles spaces natively. Build the path as: <vault_path>\<relative_path_from_keppi_result>

Example:

Desktop Commander:read_file
  path: C:\Users\username\Documents\Obsidian Vault\1-Projects\Acme\2025-12-16 Meeting.md

The exact relative path comes from the path field in keppi search/blast_radius results.

Option C: No Desktop Commander — keppi:context_pack only

Skip Step 4 entirely. The context pack from Step 3 contains the note content. Synthesize directly from that output.

Step 5 — Synthesize

Combine graph structure (blast radius relevance scores) with note content to answer the question. Cite notes by title. If the research reveals new cross-connections not in the wiki, note them — they're candidates for a wiki update.

Step 6 — Promote Facts to Wiki (MANDATORY)

Deep research almost always surfaces facts that should live in the wiki. If you assembled an answer from multiple raw notes, those facts need a home.

The rule: If you assembled facts from raw notes that would answer the question on their own, promote them. The wiki is a queryable facts layer, not a summary layer.

How to promote:

1. After synthesizing your answer, identify the 1-3 key facts that required raw note reads
2. Find (or create) the relevant wiki page in your wiki subfolder (e.g. wiki/entities/ or wiki/concepts/)
3. Add a concise section: ## Pricing, ## Key Dates, ## Status, etc.
4. Add a wikilink to each source note: Source: [[Source Note Name]]
5. Update the sources: frontmatter list to include the source note
6. Set updated: frontmatter to today's date

Result: Next query on this topic resolves in a single wiki read (400 tokens) instead of a full research pass (8,000–16,000 tokens).

Example: A "summarize everything about the Cetera project" query required reading 5 raw meeting notes. The key facts — contract status, deliverables, deadlines — get promoted to the Cetera wiki page. Next time, the wiki fast lane answers the factual questions; only deep evidence retrieval needs the research path.

Obsidian CLI Quick Reference

Desktop Commander:start_process
  command: obsidian <command> <args>
  shell: powershell.exe   ← ALWAYS PowerShell. CMD truncates on spaces.
  timeout_ms: 10000

Keppi MCP Tools Used in This Skill

Performance Rules

• Prefer keyword_search + blast_radius over query_node — query_node on a 2,600+ node vault returns 150+ edges (large payload, slow round-trip over MCP).
• Keep keyword_search limit ≤ 5 — more results = more MCP payload for little gain.
• Keep blast_radius depth ≤ 2 — depth 3+ fans out to thousands of nodes.
• Wiki reads are cheap — ~400 tokens vs. 3,000–8,000 for a raw transcript. Always try the wiki fast lane before committing to raw note reads.
• context_pack with a tight budget is better than reading 10 raw notes individually.
• Promote key facts after deep research — if you assembled an answer from 4+ raw notes, add the key facts to the relevant wiki page. Next query resolves at wiki-lane cost (400 tokens) instead of a full research pass (8,000–16,000 tokens).