Embedding pipeline improvements building on the concurrent batching
foundation:
- Track docs_embedded vs chunks_embedded separately. A document counts
as embedded only when ALL its chunks succeed, giving accurate
progress reporting. The sync command reads docs_embedded for its
document count.
- Reuse a single Vec<u8> buffer (embed_buf) across all store_embedding
calls instead of allocating per chunk. Eliminates ~3KB allocation per
768-dim embedding.
- Detect and record errors when Ollama silently returns fewer
embeddings than inputs (batch mismatch). Previously these dropped
chunks were invisible.
- Improve retry error messages: distinguish "retry returned unexpected
result" (wrong dims/count) from "retry request failed" (network
error) instead of generic "chunk too large" message.
- Convert all hot-path SQL from conn.execute() to prepare_cached() for
statement cache reuse (clear_document_embeddings, store_embedding,
record_embedding_error).
- Record embedding_metadata errors for empty documents so they don't
appear as perpetually pending on subsequent runs.
- Accept concurrency parameter (configurable via config.embedding.concurrency)
instead of hardcoded EMBED_CONCURRENCY=2.
- Add schema version pre-flight check in embed command to fail fast
with actionable error instead of cryptic SQL errors.
- Fix --retry-failed to use DELETE instead of UPDATE. UPDATE clears
last_error but the row still matches config params in the LEFT JOIN,
making the doc permanently invisible to find_pending_documents.
DELETE removes the row entirely so the LEFT JOIN returns NULL.
Regression test added (old_update_approach_leaves_doc_invisible).
- Add chunking forward-progress guard: after floor_char_boundary()
rounds backward, ensure start advances by at least one full
character to prevent infinite loops on multi-byte sequences
(box-drawing chars, smart quotes). Test cases cover the exact
patterns that caused production hangs on document 18526.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
c2036c64e9