diff --git a/backend/alembic/versions/d6e7f8091a2b_search_the_article_body.py b/backend/alembic/versions/d6e7f8091a2b_search_the_article_body.py
new file mode 100644
index 0000000..527c079
--- /dev/null
+++ b/backend/alembic/versions/d6e7f8091a2b_search_the_article_body.py
@@ -0,0 +1,63 @@
+"""Full-text search over what an article actually says.
+
+`search_vector` weighted title, summary and `content` — but 323 of the 331
+articles have `content` NULL, because everything the generator writes goes into
+the `sections` JSON and only the eight hand-seeded samples ever used the column.
+So for 98% of the library the body contributed nothing, and a learner searching
+for a drug name, a diagnostic criterion or an eponym that appears only in a
+section got no result and no indication that the search had not looked there.
+
+A generated column cannot contain a subquery, so the extraction lives in an
+IMMUTABLE function it can call. `content` stays in the expression for the eight
+articles that use it.
+
+Revision ID: d6e7f8091a2b
+Revises: c5d6e7f8091a
+"""
+from alembic import op
+
+revision = "d6e7f8091a2b"
+down_revision = "c5d6e7f8091a"
+branch_labels = None
+depends_on = None
+
+EXTRACT = """
+CREATE OR REPLACE FUNCTION article_sections_text(sections json)
+RETURNS text LANGUAGE sql IMMUTABLE PARALLEL SAFE AS $$
+  SELECT COALESCE(string_agg(CONCAT_WS(' ', s->>'title', s->>'content'), ' '), '')
+  FROM json_array_elements(COALESCE(sections, '[]'::json)) AS s
+$$;
+"""
+
+OLD = """
+  setweight(to_tsvector('english', COALESCE(title, '')), 'A') ||
+  setweight(to_tsvector('english', COALESCE(summary, '')), 'B') ||
+  setweight(to_tsvector('english', COALESCE(content, '')), 'C')
+"""
+
+NEW = """
+  setweight(to_tsvector('english', COALESCE(title, '')), 'A') ||
+  setweight(to_tsvector('english', COALESCE(summary, '')), 'B') ||
+  setweight(to_tsvector('english',
+    COALESCE(content, '') || ' ' || article_sections_text(sections)), 'C')
+"""
+
+
+def _rebuild(expression: str) -> None:
+    # The expression of a generated column cannot be altered in place, and
+    # dropping the column takes its index with it.
+    op.execute("DROP INDEX IF EXISTS ix_articles_search_vector")
+    op.execute("ALTER TABLE articles DROP COLUMN IF EXISTS search_vector")
+    op.execute(f"ALTER TABLE articles ADD COLUMN search_vector tsvector "
+               f"GENERATED ALWAYS AS ({expression}) STORED")
+    op.execute("CREATE INDEX ix_articles_search_vector ON articles USING gin (search_vector)")
+
+
+def upgrade():
+    op.execute(EXTRACT)
+    _rebuild(NEW)
+
+
+def downgrade():
+    _rebuild(OLD)
+    op.execute("DROP FUNCTION IF EXISTS article_sections_text(json)")
diff --git a/backend/app/routers/questions.py b/backend/app/routers/questions.py
index 48f3d4c..ff624f6 100644
--- a/backend/app/routers/questions.py
+++ b/backend/app/routers/questions.py
@@ -41,7 +41,11 @@ router = APIRouter()
 MAX_SEARCH_RESULTS = 500
 # A matched test stays a study aid, not a dump of the bank.
 MAX_MATCHED_QUESTIONS = 30
-MAX_UPLOAD_BYTES = 10 * 1024 * 1024
+# A document here is a query, never content: it is read once to find matching
+# questions in the bank and then discarded, so the cap is about how much text
+# is worth reading rather than about storage. Two megabytes is a long syllabus
+# or a score report; past that somebody is uploading a textbook.
+MAX_UPLOAD_BYTES = 2 * 1024 * 1024
 MAX_QUERY_CHARS = 6000