diff --git a/backend/migrations/0005_chunk_indexes.sql b/backend/migrations/0005_chunk_indexes.sql
new file mode 100644
index 0000000..670ad1f
--- /dev/null
+++ b/backend/migrations/0005_chunk_indexes.sql
@@ -0,0 +1,6 @@
+-- enrich.embed looks up claims and documents without a chunk on every batch. chunk_uniq is an expression index
+-- (COALESCE(...)), which those lookups can't use, so each batch scanned the whole table: after the history
+-- backfill (over a million claims) a run took hours and blocked the other enrichers. Built CONCURRENTLY by hand on
+-- 2026-11-14; IF NOT EXISTS makes this a no-op there.
+CREATE INDEX IF NOT EXISTS chunk_claim_idx ON chunk (claim_id) WHERE claim_id IS NOT NULL;
+CREATE INDEX IF NOT EXISTS chunk_document_idx ON chunk (document_id) WHERE document_id IS NOT NULL;