Annotate reports with an AI-written story - #1923
Conversation
Add a Section.include_in_export flag and route to_notebook/to_html through _export_views/_export_view so a report can be exported with a subset of its sections. Defaults keep every section included, so existing exports are unchanged.
Render each report accordion card header as a checkbox bound to the section include_in_export flag next to its title, so users can pick which sections to keep before exporting. The checkbox appears once the report has outputs and leaves the rendered report unchanged.
Apply the HoloViz cleanup review: replace the list-building loop in Report._export_views with a comprehension. No behavior change.
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## main #1923 +/- ##
==========================================
+ Coverage 72.48% 73.47% +0.98%
==========================================
Files 203 208 +5
Lines 35960 37534 +1574
==========================================
+ Hits 26066 27578 +1512
- Misses 9894 9956 +62 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
|
Can the story interweave plots and tables like a blog post? Also, do you have an updated video? Also can user write a prompt to ask the LLM to focus on a specific angle? |
|
Yup.. sharing the video only compressing that |
Untitled.-.10.July.2026.at.02.51.11.3.mp4 |
Only Report is ever exported, so the base TaskGroup _export_views / _export_view wrappers (which just returned self.views / self._view) were dead code. Drop them and move to_notebook / to_html onto Report next to its _export_views / _export_view overrides.
|
Okay thanks for sharing the video. Ideally, the story interweaves pictures and texts. Also, on clicking the annotate, it should show a popup to enter text to guide the story. |
|
Hmm Okay Andrew, will refactor this and add this |
3f567c6 to
6c01e04
Compare
Screen.Recording.2026-07-10.at.6.25.28.PM.mp4 |
|
I like it (poor low performer 😭)! Let's think about UI/UX. The current implementation is a bit weird ongoing back and forth on base report vs story report. I wonder if we can use Tabs or some other layout to differentiate. Also, what if user wants to tweak a specific section, or make manual edits (should this be supported)? Lastly, what if user simply wants to regenerate the full story? Or wants to create a new version? |
|
Haha 😅 Yupp, need suggestions over the UI/UX, planning it, I too feel back and forth weird.
For this, I was thinking to add some button edit with AI.
Hmm, Better UX needed for this.... |
If it can work like Google docs, that'd be cool (ReactiveHTML is outdated now, use JSComponent) |
Nested sections rendered without a checkbox because only the report drew section headers, and the export read each top level section's aggregated views, so a nested section's selection was ignored. Render a heading and checkbox for nested sections, and walk the tasks when collecting the views so a discarded section is left out of the notebook and the HTML at any depth.
Selection only reached whole sections, so two charts made in one session could not be kept or dropped separately. Move include_in_export onto Task and give every task its own checkbox, and walk the tasks when collecting the views so discarding one chart leaves the rest of its section intact. The report explains the checkboxes at the top rather than on a per-checkbox info icon.
Add a StoryAgent that turns the report's charts and tables into ordered prose and view blocks, plus the catalog builder that describes each view for the prompt.
Add an Annotate action that opens a guidance popup (tone presets plus free text) and writes an AI story that interleaves prose with the charts and tables it discusses. The story is a separate view reachable from the report, an Arrange dialog reorders sections via a JSONEditor, and the notebook and HTML exports follow whichever view is shown.
Replace the toggle that swapped the report view for the story view with a Tabs layout, so the two are told apart at a glance instead of by flipping back and forth. The Story tab appears once a story is generated and is removed when it is discarded, and the exports follow whichever tab is selected.
Render each prose block as a contenteditable JSComponent so the story can be tweaked by hand like a document, and keep the prose as text on the story blocks so an edit flows into the notebook, HTML and Word exports. The text is only pushed back into the element when it changes elsewhere, otherwise the caret would jump to the start while typing.
Add a Regenerate action on the story that reopens the guidance popup so the whole story can be written again with a different angle, and give each paragraph an edit action that asks the LLM to rewrite just that paragraph from an instruction, leaving the charts and the rest of the story untouched.
Regenerating now adds a version instead of replacing the story, and a picker on the story switches between them, so a version the user has edited by hand is never thrown away by asking for a fresh take. Each version owns its blocks, so edits stay with the version they were made on.
The editable paragraphs showed the raw Markdown the model writes, so the story read as source rather than prose. Render each paragraph and swap to the source only while it is being edited, rendering again on blur, using the same Markdown engine the Markdown pane uses.
6c01e04 to
6bd11e3
Compare
The report was vertically centred in the main area, so a tall report or story never received a bounded height and could not scroll, while a short one left a gap below it. Stretch the report to fill the area and scroll its own content instead. Story charts and tables now render as fixed-height figures so they flow like blog-post images rather than stretching to fill and swallowing the scroll, and the export hint is shown only on the report tab where the checkboxes it describes live.
The story catalog labelled each figure with the view's title, which is the pipeline step name (e.g. "Prepare data", "Create pie chart"), so those names leaked into the prose. Drop the title from the catalog and let the data summary describe each figure, and instruct the model to refer to figures by what they show.
|
@ahuang11, How about this? Screen.Recording.2026-07-19.at.1.44.19.PM.mp4 |
fix: conflicts report - keep the story feature (tabs, editable prose, versions) and the fill/scroll fix layered on main's export-selection base, which is identical to ours fix: conflicts ui - take main's hsplit navigation; the report now fills and scrolls through _nav_content (which stretches, unlike the old centered _main), so our ui.py stretch toggle is no longer needed; kept the story's llm and jsoneditor wiring
Add a Report mode section covering the Annotate feature: generating an interleaved prose-and-chart story, editing paragraphs by hand or with AI, keeping each generation as a version, and exporting whichever tab is shown, with a short worked example.
| ) | ||
|
|
||
|
|
||
| class EditableProse(JSComponent): |
There was a problem hiding this comment.
Should this be in report.py? Or story.py? Or even a panel-extension package https://github.com/panel-extensions/copier-template-panel-extension?
There was a problem hiding this comment.
moved to editor only
| _STORY_TAB = 1 | ||
|
|
||
| _STORY_PRESETS = [ |
There was a problem hiding this comment.
Should these be in CAPS? Maybe use https://holoviz-dev.github.io/holoviz-skills/contributing-to-holoviz/cleanup/ (and if it doesn't fix this, please update that skill to reflect)
| self._story_blocks = [] | ||
| self._story_versions = [] | ||
| self._story_version = 0 | ||
| self._story_outline = [] |
There was a problem hiding this comment.
If there's so many _story related variables, should this be a dict or a Story param class?
| instruction = (self._edit_instruction.value or "").strip() | ||
| if self.llm is None or self._edit_index is None or not instruction: | ||
| return | ||
| from .agents.story import StoryAgent, build_catalog |
- Move EditableProse out of report.py into editors.py, alongside LumenEditor, so the JSComponent lives with the other editor components. - Import StoryAgent and build_catalog at the module top rather than inline, now that there is no circular import to avoid. - Group the report's story state (title, blocks, versions, version, outline) into a StoryState param class instead of five loose instance attributes.
| self._story_generate = Button( | ||
| label="Generate", variant="contained", icon="auto_stories", | ||
| on_click=self._generate_story, | ||
| ) | ||
| self._story_dialog = Dialog( | ||
| Typography("Pick a tone or write your own guidance, then generate the story.", variant="body2"), | ||
| self._story_presets, | ||
| self._story_guidance, | ||
| Row(self._story_generate, align="end", margin=(10, 0, 0, 0), sizing_mode="stretch_width"), | ||
| show_close_button=True, | ||
| title="Annotate Report", | ||
| width_option="sm", |
There was a problem hiding this comment.
I'm beginning to think we should do inheritance to have a StoryReport(Report) and migrate this to story.py
These are private class attributes rather than public constants, so drop the all-caps naming; give the preset list its own name so it no longer collides with the presets Row.
StoryState is story domain state, so it belongs in agents/story.py with the other story models rather than in report.py.
Move the AI story-annotation tab off the base Report onto a StoryReport subclass so the base report stays free of story state and UI. Reports in the app now instantiate StoryReport, base Report keeps only the core report/export behaviour.
|
Exciting feature! However a few bugs:
|
Hide the row of a task that renders nothing, so an unviewable output no longer leaves a lone export checkbox in empty space, and skip the header of an untitled nested section for the same reason. Walk the whole exploration tree when building the report. Followups nest arbitrarily deep, and descending a single level dropped the newest exploration from the report. Reuse the report between visits instead of rebuilding it, so the Story tab and its prose survive a trip back to the chat. Leave the Explore menu item inactive and outlined while report mode is on, rather than flipping it to an active timeline icon. Leave report mode when an exploration is picked in the navigation.
Hide the row of a task that renders nothing, so an unviewable output no longer leaves a lone export checkbox in empty space, and skip the header of an untitled nested section for the same reason. Walk the whole exploration tree when building the report. Followups nest arbitrarily deep, and descending a single level dropped the newest exploration from the report. Reuse the report between visits instead of rebuilding it, so the Story tab and its prose survive a trip back to the chat. Leave the Explore menu item inactive and outlined while report mode is on, rather than flipping it to an active timeline icon. Leave report mode when an exploration is picked in the navigation.
776ed00 to
8774862
Compare
|
@ahuang11, I think this can be merged now!! tested this :) |
2 participants
Adds an Annotate button to the report toolbar that uses an LLM to write a story about the selected tables and charts: an overall summary at the top plus a short note under each section. The narrative renders in the report and flows into both the Notebook and HTML export. Closes #1912.
Note: this stacks on #1911 (#1922); the first commits belong to that PR, so please review after it merges.
Untitled.-.09.July.2026.at.20.50.01.1.mp4