Obsidian Semantic Search

Search Obsidian vault notes by conceptual meaning, not just keywords. Builds a local embedding index over all vault notes (per-section chunking), then performs cosine similarity search at query time.

When to Use

• User asks a fuzzy/conceptual question about their own notes ("What papers discuss exploration in RL?")
• Keyword search (mcp__obsidian__search_notes) returns too many or too few results
• User wants to find notes related to a concept without knowing exact terminology
• As a pre-retrieval step before deep-reading or analyzing vault content

Architecture

Vault (.md files)  ──build_index.py──>  temp/obsidian/embeddings.json  <──search.py──  User Query
                                                │
                                                └── Shared with paper-deep-reader

• Index: Flat JSON, one embedding per .md note (full body text, frontmatter excluded), mtime-tracked incremental rebuild
• Embeddings: Reuses stable_jarvis.llm.embed() — same provider stack as paper-deep-reader (local / qwen / openai)
• Search: Brute-force cosine similarity, O(n) over all notes

Workflow

Phase 0: Check Index Freshness (Always First)

Before any search, the agent MUST check the index state:

python skills/obsidian-semantic-search/scripts/build_index.py --stats

Decision logic:

Phase 1: Build/Update Index (If Needed)

# Incremental update (recommended when index exists but stale)
python skills/obsidian-semantic-search/scripts/build_index.py

# Full rebuild (first time, or after major vault changes)
python skills/obsidian-semantic-search/scripts/build_index.py --force

The scripts locate the vault automatically via OBSIDIAN_VAULT env var or common paths (~/Documents/Obsidian, ~/Obsidian, ~/vault). If the vault is at a non-standard location, set it first:

export OBSIDIAN_VAULT="/path/to/vault"

Phase 2: Semantic Search

python skills/obsidian-semantic-search/scripts/search.py "query text" --top-k 5

Returns JSON array of results, each with:

• title — Note title (from H1 or filename)
• path — Relative path in vault (use directly with mcp__obsidian__read_note)
• score — Cosine similarity (0–1, higher = more relevant)
• text_snippet — First 200 chars of the note body
• arxiv_id — From frontmatter, if present

Phase 3: Read Full Notes (via Obsidian MCP)

After getting results, use mcp__obsidian__read_note with the path field to read full content:

mcp__obsidian__read_note(path="30 Zettelkasten/RL/MAC.md")

Phase 4: Present Results

Format search results for the user with:

• Note title as [[wikilink]]
• Relevance score
• Brief excerpt from text_snippet

Commands

Configuration

All settings via environment variables (loaded from .env):

The local provider uses sentence-transformers (paraphrase-multilingual-MiniLM-L12-v2) and requires no API key.

Index Management

• Incremental updates: Running build_index.py without --force compares file mtimes and only re-embeds changed notes. Deleted notes are automatically removed from the index.
• Shared index: The index at temp/obsidian/embeddings.json is shared with paper-deep-reader. Both skills read and benefit from the same index.
• Rebuild frequency: Run --force after major vault reorganization or when search quality degrades.

Integration Notes

• Results from search.py include the vault-relative path — this is the exact argument to pass to mcp__obsidian__read_note
• Combine with mcp__obsidian__search_notes for hybrid retrieval: semantic for recall, keyword for precision
• Index is stored at temp/obsidian/embeddings.json — shared with paper-deep-reader skill