feat(core): support frontmatter field to skip embeddings for individual notes

[phernandez]

Summary

Add a frontmatter field that allows individual notes to opt out of semantic embeddings. Notes with this field will still be indexed and searchable via full-text search (FTS), but will skip the embedding generation step during sync/indexing.

Motivation

Some notes don't benefit from semantic search — large reference documents, logs, templates, or auto-generated content. Skipping embeddings for these notes:

• Reduces sync time and resource usage (embedding API calls or local model inference)
• Keeps the vector index focused on semantically meaningful content
• Gives users fine-grained control over what gets embedded

Proposed Frontmatter Field

---
title: My Reference Note
embed: false
---

embed: false skips embedding generation. Default behavior (field absent or embed: true) is unchanged — embeddings are generated as usual.

Field name rationale: embed is concise and directly describes what it controls. Alternatives considered: embeddings: false (verbose), skip_embeddings: true (negative framing), searchable: false (ambiguous — FTS still works).

Implementation Notes

• During indexing, check for embed: false in parsed frontmatter
• Skip the embedding provider call for that entity
• Entity still gets full-text indexed normally
• Existing entities that gain this field on re-index should have their embeddings cleared
• Search behavior: these notes appear in FTS results but not in semantic/vector search results

Test Plan

• Note with embed: false is indexed without embeddings
• Note without the field behaves as before (embeddings generated)
• FTS still finds notes with embed: false
• Semantic search does not return notes with embed: false
• Re-indexing a note that adds embed: false clears existing embeddings
• Re-indexing a note that removes embed: false generates embeddings

[claude]

Claude finished @phernandez's task in 1m 37s —— View job

---

Triage Analysis

• Gather context and read issue
• Explore embedding-related code to assess complexity
• Check for duplicate issues
• Apply labels
• Provide triage summary

---

Classification

Field	Value
Type	Feature Request
Priority	Low
Component	Sync / Embedding Pipeline
Complexity	Medium

Duplicate Check

No duplicate found. Issue #578 (closed) was a bug about embeddings being silently skipped on dependency failure — unrelated.

Analysis

This is a well-specified feature request to allow per-note opt-out of embedding generation via embed: false in frontmatter. The issue includes clear motivation, implementation notes, and a test plan.

Code impact assessment:

The embedding pipeline lives in sync/sync_service.py. After syncing files, it collects all synced entity IDs and calls search_service.sync_entity_vectors_batch(synced_entity_ids) in a batch. Frontmatter metadata is already parsed and available via EntityFrontmatter.metadata (a flexible dict), so reading embed: false from it is straightforward.

Key implementation touchpoints:

• sync/sync_service.py (~line 313-374): Filter out entity IDs where embed: false before the batch embedding call. Also handle clearing existing embeddings for entities that newly set embed: false.
• services/search_service.py: May need a method to clear embeddings for specific entities.
• repository/search_repository_base.py + concrete implementations: Possibly expose a delete-by-entity-id for embeddings.
• Tests: Unit + integration tests for all 6 scenarios listed in the test plan.

Design note: The embed key would live in EntityFrontmatter.metadata as arbitrary metadata — no schema changes needed. The check is purely at sync time.

Complexity is Medium because:

• Parsing the flag is trivial (metadata dict access)
• The clearing-on-reindex path requires careful handling (sync determines delta, must detect the embed field changed)
• Both SQLite and Postgres search repositories need to support embedding deletion
• Full test coverage required per project standards

Labels Applied

• enhancement ✓

Note: Other labels from the triage guide (priority, component, complexity) do not appear to exist in this repository yet.