perf(index): reduce embedding batch long-tail stalls

Author: simonsysun

CHANGELOG.md

@@ -7,6 +7,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Changed
+- Full-vault indexing now embeds in smaller batches to reduce long-tail embedding stalls on real Markdown vaults.
+
 ## [0.4.0] - 2026-04-29
 
 ### Added

seeklink/ingest.py

@@ -22,7 +22,7 @@
 
 # Non-hidden top-level dirs excluded from indexing (mirrors freshness._SKIP_DIRS)
 _SKIP_DIRS = {"todo", "archive"}
-_EMBED_BATCH_SIZE = 32
+_EMBED_BATCH_SIZE = 16
 
 # Regex for YAML frontmatter block (handles empty frontmatter too).
 # Public — search.py imports this to map body-relative chunk offsets back

tests/test_ingest.py

@@ -12,7 +12,12 @@
 
 from seeklink.db import Database
 from seeklink.embedder import Embedder
-from seeklink.ingest import _parse_frontmatter, ingest_file, ingest_vault
+from seeklink.ingest import (
+    _EMBED_BATCH_SIZE,
+    _parse_frontmatter,
+    ingest_file,
+    ingest_vault,
+)
 
 
 @pytest.fixture(scope="session")
@@ -314,8 +319,8 @@ def test_batches_embeddings_across_files(self, db: Database, vault: Path):
         assert stats["ingested"] == 40
         assert stats["errors"] == 0
         call_sizes = [len(call) for call in fake.calls]
-        assert len(call_sizes) == 2
-        assert max(call_sizes) <= 32
+        assert len(call_sizes) == (40 + _EMBED_BATCH_SIZE - 1) // _EMBED_BATCH_SIZE
+        assert max(call_sizes) <= _EMBED_BATCH_SIZE
         assert sum(call_sizes) == 40
 
     def test_batch_vault_resolves_forward_refs(self, db: Database, vault: Path):