feat: extract_text, and the five UI gaps behind a working API

README.md

@@ -41,9 +41,10 @@ user row. The changes that makes possible are specified in
   in audio and video come back with a timestamp you can jump straight to.
 - **Clips and sub-videos** — mark a range non-destructively (no new file, always in step
   with its parent), or physically extract it as a standalone asset.
-- **Opt-in AI enrichment** — transcription, image description, summaries and tag
-  suggestions. Suggestions are reviewed, never applied silently, and a field you edited
-  by hand is not overwritten by a later AI run.
+- **Opt-in AI enrichment** — transcription, text extraction from PDFs and Office
+  documents, image description, summaries and tag suggestions. Suggestions are reviewed,
+  never applied silently, and a field you edited by hand is not overwritten by a later AI
+  run.
 - **AI asset creation** — generate images from images, and video from one or more
   images, with the prompt and base assets recorded so a result stays reproducible.
 - **Cost visibility** — an estimate before anything paid runs, and the real cost tracked

backend/alembic/versions/20260916_1111_add_documentpage_table.py

@@ -0,0 +1,55 @@
+"""add documentpage table
+
+Readable text pulled out of a document, so `summarize` and `autotag` have something to
+read for a PDF or a Word file. Until now they refused: `enrichment/source.py` raised
+`NoSourceMaterial` saying so out loud, because the `extract_text` job the plan lists had
+never been built.
+
+One row per page, slide, sheet or chunk rather than a column on `asset`, because
+SQLModel selects every column of every row and the library listing selects assets by the
+page — a 200-page PDF's text in an asset column would be read from disk to render a
+thumbnail grid that never looks at it. No foreign key to `asset`, matching
+`transcriptsegment`; the delete cascade is explicit in `services/assets.py`.
+
+Revision ID: 56ac14e89a0c
+Revises: c4954d1715e7
+Create Date: 2026-09-16 11:11:08.578568+00:00
+"""
+from typing import Sequence, Union
+
+from alembic import op
+import sqlalchemy as sa
+import sqlmodel
+
+
+revision: str = '56ac14e89a0c'
+down_revision: Union[str, None] = 'c4954d1715e7'
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+    op.create_table('documentpage',
+    sa.Column('id', sqlmodel.sql.sqltypes.AutoString(), nullable=False),
+    sa.Column('asset_id', sqlmodel.sql.sqltypes.AutoString(), nullable=False),
+    sa.Column('user_id', sqlmodel.sql.sqltypes.AutoString(), nullable=False),
+    sa.Column('idx', sa.Integer(), nullable=False),
+    sa.Column('page_number', sa.Integer(), nullable=True),
+    sa.Column('label', sqlmodel.sql.sqltypes.AutoString(), nullable=True),
+    sa.Column('text', sqlmodel.sql.sqltypes.AutoString(), nullable=False),
+    sa.PrimaryKeyConstraint('id')
+    )
+    with op.batch_alter_table('documentpage', schema=None) as batch_op:
+        batch_op.create_index(batch_op.f('ix_documentpage_asset_id'), ['asset_id'], unique=False)
+        batch_op.create_index('ix_documentpage_asset_idx', ['asset_id', 'idx'], unique=False)
+        batch_op.create_index(batch_op.f('ix_documentpage_user_id'), ['user_id'], unique=False)
+
+
+
+def downgrade() -> None:
+    with op.batch_alter_table('documentpage', schema=None) as batch_op:
+        batch_op.drop_index(batch_op.f('ix_documentpage_user_id'))
+        batch_op.drop_index('ix_documentpage_asset_idx')
+        batch_op.drop_index(batch_op.f('ix_documentpage_asset_id'))
+
+    op.drop_table('documentpage')

backend/app/enrichment/autotag.py

@@ -78,6 +78,11 @@ def _prompt(session: Session, asset: Asset, material: source.SourceMaterial) ->
         if material.truncated:
             header += " (the opening portion only)"
         parts.append(f"{header}:\n\n{material.text}")
+    elif material.kind == source.FROM_DOCUMENT:
+        header = "Document text"
+        if material.truncated:
+            header += " (the opening portion only)"
+        parts.append(f"{header}:\n\n{material.text}")
     elif material.kind == source.FROM_POSTER:
         parts.append("No transcript is available; a single frame from the video is attached.")
     else:

backend/app/enrichment/bulk.py

@@ -19,9 +19,15 @@
 
 from sqlmodel import Session
 
-from app.enrichment import autotag, describe, embed, summarize
+from app.enrichment import autotag, describe, embed, extract_text, summarize
 from app.models.asset import Asset
-from app.models.job import KIND_AUTOTAG, KIND_DESCRIBE, KIND_EMBED, KIND_SUMMARIZE
+from app.models.job import (
+    KIND_AUTOTAG,
+    KIND_DESCRIBE,
+    KIND_EMBED,
+    KIND_EXTRACT_TEXT,
+    KIND_SUMMARIZE,
+)
 from app.providers.base import ProviderUnavailable
 
 logger = logging.getLogger(__name__)
@@ -40,6 +46,10 @@
     KIND_SUMMARIZE: summarize.run,
     KIND_AUTOTAG: autotag.run,
     KIND_EMBED: embed.run,
+    # Safe to include where transcription is not: it calls nothing and bills nothing, so
+    # the mis-click that makes transcription too expensive to offer here costs only time.
+    # It is also the action most worth having in bulk — documents arrive by the folder.
+    KIND_EXTRACT_TEXT: extract_text.run,
 }
 
 # A selection has to be reviewable before it is run, and it is the thing standing between
@@ -140,4 +150,5 @@ def inner(_stage, _pct, _detail="", *, stage=stage, pct=pct, detail=detail):
     KIND_SUMMARIZE: "Summarising",
     KIND_AUTOTAG: "Suggesting tags",
     KIND_EMBED: "Embedding",
+    KIND_EXTRACT_TEXT: "Reading text",
 }

backend/app/enrichment/describe.py

@@ -71,6 +71,11 @@ def _prompt(asset: Asset, material: source.SourceMaterial) -> str:
                 "visual — who is on screen, the setting — which the transcript cannot "
                 "tell you."
             )
+    elif material.kind == source.FROM_DOCUMENT:
+        header = "Text of the document"
+        if material.truncated:
+            header += " (the opening portion only — it continues beyond this)"
+        parts.append(f"{header}:\n\n{material.text}")
     elif material.kind == source.FROM_POSTER:
         parts.append(
             "There is no transcript, so a single still frame from the video is "