diff --git a/backend/alembic/versions/y7e8f9a0b1c2_media_assets.py b/backend/alembic/versions/y7e8f9a0b1c2_media_assets.py
new file mode 100644
index 0000000..54b13bf
--- /dev/null
+++ b/backend/alembic/versions/y7e8f9a0b1c2_media_assets.py
@@ -0,0 +1,57 @@
+"""Media as a searchable corpus, ready for a vision embedder.
+
+Images are currently found only by the filename someone typed. This gives them a
+caption, alt text and tags to search on now, and an embedding column that a
+vision-capable model can fill later without another migration — `embedding_model`
+records which model produced each vector, so a text-embedded caption and a
+vision-embedded image are distinguishable rather than silently mixed.
+
+Revision ID: y7e8f9a0b1c2
+Revises: x6d7e8f9a0b1
+"""
+from alembic import op
+
+revision = "y7e8f9a0b1c2"
+down_revision = "x6d7e8f9a0b1"
+branch_labels = None
+depends_on = None
+
+
+def upgrade():
+    op.execute("""
+    CREATE TABLE IF NOT EXISTS media_assets (
+        id SERIAL PRIMARY KEY,
+        path VARCHAR(500) UNIQUE NOT NULL,
+        title VARCHAR(300),
+        caption TEXT,
+        alt_text TEXT,
+        kind VARCHAR(20) NOT NULL DEFAULT 'image',
+        category_id INTEGER REFERENCES question_categories(id) ON DELETE SET NULL,
+        user_id INTEGER REFERENCES users(id) ON DELETE SET NULL,
+        embedding vector(1024),
+        embedding_model VARCHAR(120),
+        embedded_at TIMESTAMP,
+        created_at TIMESTAMP NOT NULL DEFAULT CURRENT_TIMESTAMP,
+        search_vector tsvector GENERATED ALWAYS AS (
+            setweight(to_tsvector('english', coalesce(title, '')), 'A') ||
+            setweight(to_tsvector('english', coalesce(caption, '')), 'B') ||
+            setweight(to_tsvector('english', coalesce(alt_text, '')), 'C')
+        ) STORED
+    )
+    """)
+    op.execute("CREATE INDEX IF NOT EXISTS ix_media_search ON media_assets USING GIN (search_vector)")
+    op.execute("CREATE INDEX IF NOT EXISTS ix_media_model ON media_assets(embedding_model)")
+    op.execute("""
+    CREATE TABLE IF NOT EXISTS media_tag_links (
+        id SERIAL PRIMARY KEY,
+        media_id INTEGER NOT NULL REFERENCES media_assets(id) ON DELETE CASCADE,
+        tag_id INTEGER NOT NULL REFERENCES question_tags(id) ON DELETE CASCADE,
+        CONSTRAINT uq_media_tag UNIQUE (media_id, tag_id)
+    )
+    """)
+    op.execute("CREATE INDEX IF NOT EXISTS ix_media_tag_media ON media_tag_links(media_id)")
+
+
+def downgrade():
+    op.execute("DROP TABLE IF EXISTS media_tag_links")
+    op.execute("DROP TABLE IF EXISTS media_assets")
diff --git a/backend/app/main.py b/backend/app/main.py
index 76f0f72..97f490d 100644
--- a/backend/app/main.py
+++ b/backend/app/main.py
@@ -168,7 +168,7 @@ def setup_pgvector():
     # Import new models so create_all picks them up
     from app.models import quiz_category, quiz_question_link, question_category, favorite  # noqa
     from app.models import flashcard, course  # noqa
-    from app.models import category_grant  # noqa
+    from app.models import category_grant, exam, media  # noqa
 
     # Kill stale idle-in-transaction connections from previous killed startups.
     # They hold DDL locks and cause ALTER TABLE below to hang indefinitely.
diff --git a/backend/app/models/media.py b/backend/app/models/media.py
new file mode 100644
index 0000000..ff665ec
--- /dev/null
+++ b/backend/app/models/media.py
@@ -0,0 +1,38 @@
+from datetime import datetime
+
+from sqlalchemy import Column, DateTime, ForeignKey, Integer, String, Text, UniqueConstraint
+
+from app.database import Base
+from app.models.embeddable import Embeddable
+
+
+class MediaAsset(Base, Embeddable):
+    """An image in the bank, searchable by what it shows rather than its filename.
+
+    The embedding column is filled from the caption today. A vision-capable model
+    can fill it from the image itself later; `embedding_model` records which one
+    produced each vector, so the two are never mixed silently.
+    """
+
+    __tablename__ = "media_assets"
+
+    id = Column(Integer, primary_key=True, index=True)
+    path = Column(String(500), unique=True, nullable=False)
+    title = Column(String(300), nullable=True)
+    caption = Column(Text, nullable=True)
+    alt_text = Column(Text, nullable=True)
+    kind = Column(String(20), default="image")
+    category_id = Column(Integer, ForeignKey("question_categories.id", ondelete="SET NULL"), nullable=True)
+    user_id = Column(Integer, ForeignKey("users.id", ondelete="SET NULL"), nullable=True)
+    created_at = Column(DateTime, default=datetime.utcnow)
+
+
+class MediaTagLink(Base):
+    __tablename__ = "media_tag_links"
+    __table_args__ = (UniqueConstraint("media_id", "tag_id", name="uq_media_tag"),)
+
+    id = Column(Integer, primary_key=True, index=True)
+    media_id = Column(Integer, ForeignKey("media_assets.id", ondelete="CASCADE"), nullable=False, index=True)
+    # `question_tags` is created by raw DDL rather than an ORM model, so the
+    # constraint is declared in the migration; the mapper only stores the id.
+    tag_id = Column(Integer, nullable=False, index=True)
diff --git a/backend/app/services/embedding_service.py b/backend/app/services/embedding_service.py
index adbc613..fb9950d 100644
--- a/backend/app/services/embedding_service.py
+++ b/backend/app/services/embedding_service.py
@@ -112,6 +112,9 @@ EMBEDDABLE = {
     "article": lambda row: _join(row.title, row.summary, row.content),
     "flashcard": lambda row: _join(row.front, row.back),
     "article_section": lambda row: _join(row.title, row.content),
+    # Text today; a vision model can embed the image itself later without
+    # changing anything here but this line.
+    "media": lambda row: _join(row.title, row.caption, row.alt_text),
 }
 
 
@@ -150,12 +153,15 @@ def embed_question(question) -> bool:
 
 def embeddable_models() -> dict:
     """The ORM class behind each embeddable kind."""
+    from app.models import media  # noqa — registers media_assets on the mapper.
     from app.models.article import Article, ArticleSectionIndex
     from app.models.flashcard import Flashcard
     from app.models.question import Question
 
+    from app.models.media import MediaAsset
+
     return {"question": Question, "article": Article, "flashcard": Flashcard,
-            "article_section": ArticleSectionIndex}
+            "article_section": ArticleSectionIndex, "media": MediaAsset}
 
 
 def stale_embedding_counts(db) -> dict:
diff --git a/backend/app/services/search_service.py b/backend/app/services/search_service.py
index 81708f0..b4c5766 100644
--- a/backend/app/services/search_service.py
+++ b/backend/app/services/search_service.py
@@ -78,6 +78,7 @@ CORPORA = {
     "article": ("articles", ("title", "summary", "content")),
     "flashcard": ("flashcards", ("front", "back")),
     "article_section": ("article_section_index", ("title", "content")),
+    "media": ("media_assets", ("title", "caption", "alt_text")),
 }
 
 
diff --git a/backend/tests/test_hybrid_search.py b/backend/tests/test_hybrid_search.py
index 751acae..2116faf 100644
--- a/backend/tests/test_hybrid_search.py
+++ b/backend/tests/test_hybrid_search.py
@@ -14,6 +14,7 @@ from sqlalchemy.pool import StaticPool
 
 from app.database import Base
 from app.models.course import Course  # noqa — Quiz.course_id FK needs the table in metadata.
+from app.models.media import MediaAsset  # noqa — the health report walks every embeddable table.
 from app.models.question import Question
 from app.models.user import User
 from app.services import search_service