Agent-Skills.md

Agent Skills: /graph_find — Parallel Multi-Algorithm Vault Search

Parallel multi-algorithm Obsidian vault search. Without arguments — auto-summarizes current session and maps it to vault nodes. With query — runs 3 search methods in parallel (LLM semantic, MCP text+dataview, filesystem), deduplicates & ranks results. Use when user says: 'find in vault', 'what do we have about [topic]', 'найди в обсидиане', 'search obsidian', 'vault search', 'найди контекст', 'graph find', 'map session'.

UncategorizedID: tekliner/improvado-agentic-frameworks-and-skills/graph-find
41

Install this agent skill to your local

pnpm dlx add-skill https://github.com/tekliner/improvado-agentic-frameworks-and-skills/tree/HEAD/skills/graph-find

Skill Files

Browse the full folder contents for graph-find.

Download Skill

Loading file tree…

skills/graph-find/SKILL.md

Skill Metadata

Name
graph-find
Description
"Parallel multi-algorithm Obsidian vault search. Without arguments — auto-summarizes current session and maps it to vault nodes. With query — runs 3 search methods in parallel (LLM semantic, MCP text+dataview, filesystem), deduplicates & ranks results. Use when user says: 'find in vault', 'what do we have about [topic]', 'найди в обсидиане', 'search obsidian', 'vault search', 'найди контекст', 'graph find', 'map session'."

/graph_find — Parallel Multi-Algorithm Vault Search

Thesis: Without a query — auto-summarize the current session into a blended query and map it to vault nodes. With a query — run 3 independent search methods in parallel, merge & deduplicate results, compute composite relevance scores, and present a ranked list with confidence classification.

Execution Protocol

When the user invokes /graph_find <query> or /graph_find (no arguments):

Step 0: Query Resolution (if no query provided)

If no query argument was passed, auto-generate a session summary to use as the search query.

0a. Analyze Current Session

Review the full conversation history and identify:

  • All distinct tasks/topics worked on (there may be multiple!)
  • Key entities mentioned: customer names, product names, technologies, people
  • Artifacts created: files, skills, configs, documents
  • Domain areas: which parts of the vault this touches (Improvado, Miras, Daniel Personal, etc.)

0b. Generate Blended Summary Query

Write a single blended query (1-3 sentences) that covers ALL session aspects. This is used as ONE search — not split per topic.

Format:

[Main topic/action]. Related: [entity1], [entity2], [entity3]. Areas: [domain1], [domain2].

Example:

Obsidian vault skills for graph search and content addition. Related: graph-find, graph-add, suggest_placement, CLAUDE.md. Areas: Claude Code Infrastructure, Obsidian Knowledge Base.

Why blended, not per-topic: Running N separate searches is slow. A single well-composed blended query captures enough keywords for each method to find relevant hits — LLM catches the semantic meaning, MCP catches exact terms, Grep catches keywords.

0c. Present Session Mapping

After searching (Steps 1-4), present the results as a session-to-vault mapping:

SESSION MAPPING

Summary: [blended query used]

BEST MATCHES:
1. [0.89] Data sources and vendors/Claude code/ — Sources: LLM, MCP
2. [0.72] Miras Knowledge Platform/ — Sources: LLM

RECOMMENDED related_to LINKS:
  - "[[Claude code Skills base.base]]"
  - "[[Miras Knowledge Platform]]"

Step 1: Launch All 3 Search Methods in Parallel

Use the Task tool or direct tool calls to run ALL methods simultaneously. Do NOT run them sequentially.

Method 1: LLM Semantic (suggest_placement)

$PROJECT_ROOT/claude_venv/bin/python -c "
import sys; sys.path.insert(0, '$PROJECT_ROOT')
from data_sources.obsidian.ops import suggest_placement
results = suggest_placement('QUERY_HERE')
for r in results:
    print(f'{r.score:.2f}|{r.path}|{r.reason}')
"
  • Score range: 0.6–1.0 (pre-filtered by suggest_placement)
  • Strength: Best for fuzzy topic matching, understands semantic context, analyzes vault hierarchy
  • Label results as: source=LLM

Method 2: MCP Search (text + dataview)

Run both MCP tools in parallel:

Text search:

mcp__obsidian__search_vault_simple(query="QUERY_HERE")
  • Score: 0.7 for all matches (exact text match)
  • Label as: source=MCP-text

Dataview/structured search:

mcp__obsidian__search_vault(query="QUERY_HERE")

  • Score: 0.7 for matches (frontmatter, tags, properties)

  • Label as: source=MCP-dv

  • Fallback: If MCP unavailable (Obsidian not running), log warning and skip both. This method degrades gracefully.

Method 3: Filesystem (Grep + Glob)

Run both in parallel:

Content search (Grep):

Grep(pattern="KEYWORD", path="data_sources/obsidian/vault/", glob="*.md", output_mode="files_with_matches", head_limit=20)
  • Score: 0.6 for content matches
  • Label as: source=Grep

Filename search (Glob):

Glob(pattern="**/*KEYWORD*", path="data_sources/obsidian/vault/")

  • Score: 0.5 for filename matches

  • Label as: source=Glob

  • Strength: Fastest, zero LLM cost, always available, catches exact keywords

  • Query preprocessing: Extract 1-3 most distinctive keywords from query (skip articles, prepositions)

Step 2: Merge & Deduplicate

  1. Collect all results into flat list: [{path, source, score, snippet}]
  2. Normalize paths: Strip vault root prefix, lowercase for comparison, treat folder/ and folder/index.md as same
  3. Group by normalized path (deduplicate same file from multiple sources)
  4. For each unique path, track:
    • sources_count — how many methods found it (max 3: LLM, MCP, Filesystem)
    • max_score — highest relevance score from any method
    • sources_list — which methods found it (e.g., "LLM, MCP-text, Grep")

Step 3: Compute Composite Score

For each unique result:

composite = (sources_count / 3) * 0.4 + max_score * 0.6
  • Multi-source bonus: Files found by all 3 methods score highest
  • Quality anchor: High relevance from even 1 method still ranks well (0.6 * max_score)
  • Sort by composite score descending

Step 4: Classify & Present

Confidence levels:

| Level | Criteria | |-------|----------| | HIGH | composite >= 0.8 AND sources_count >= 2 | | MEDIUM | composite >= 0.6 OR sources_count >= 2 | | LOW | composite < 0.6 AND sources_count == 1 | | NO_MATCH | No results from any method |

Output format:

VAULT SEARCH: "query"

TOP MATCH: path/to/file.md [0.91]
  Found by: LLM(0.92), MCP-text, Grep

RANKED RESULTS:
1. [0.91] path/to/file1.md — Sources: LLM, MCP-text, Grep
   Snippet: first relevant line or reason
2. [0.78] path/to/folder/ — Sources: LLM, MCP-dv
   Snippet: ...
3. [0.65] path/to/file2.md — Sources: Grep, Glob
   Snippet: ...

RECOMMENDATION:
  Best file: path/to/file1.md
  Best folder: path/to/folder/
  Confidence: HIGH
  • Show top 10-15 results max
  • Include snippet or reason for each
  • State confidence level

Graceful Degradation

If any method fails, log warning and continue with remaining methods.

| Working methods | Quality | |----------------|---------| | 3/3 | Full confidence | | 2/3 | Good results, note failure | | 1/3 (filesystem always works) | Usable, warn about reduced confidence | | 0/3 | Report failure, suggest manual browsing |

Warning: [!] Method X unavailable: [reason]. Continuing with N remaining methods.

Query Preprocessing

Before searching:

  1. Extract keywords: For Grep/Glob, use 1-3 distinctive words
  2. Keep full query: For LLM and MCP methods, use full natural language
  3. Russian support: If query in Russian, also search English equivalents
  4. Folder hints: If query mentions known entity (customer, product, person), add vault path hints

Integration with /graph_add

When called from /graph_add, return structured data (not just formatted output):

  • Full results list with scores and sources
  • Confidence classification
  • Best file and best folder recommendations

Related Skills

| Skill | Relationship | |-------|-------------| | /graph_add | Uses graph_find results to decide create/update/refactor | | /obsidian | Primary vault operations (MCP tools, Python ops) | | /find-right-folder-for-context | Simpler folder recommendation (repo-level, not vault search) |

Created: 2026-02-05 | Version: 2.0.0 Changes: v2 — honest 3 methods (removed invented search_vault_smart), blended query for session mode