Generate Slides — Native PPTX engine for Open WebUI

Part of NEURA Office: native Word, PowerPoint and Excel for Open WebUI.

A Open WebUI Tool that generates native PowerPoint (.pptx) presentations from a JSON spec produced by the model. It doesn't export HTML or images: it builds the slides directly with python-pptx, with a coherent visual system, layered decorative shapes, native charts, icons inside circles and rich layouts.

The resulting file is saved through Open WebUI's Files API (with a /cache/files fallback) and a clickable download link appears in the chat.

License: MIT · Author: IANUSTEC

Slides generated from the example examples/deck.json → examples/demo_tech_deck.pptx.

Features

• Native, editable .pptx (text, charts and shapes are editable in PowerPoint/Keynote/LibreOffice).
• Native Office charts: bar, line, area, pie, doughnut, radar, stacked_bar.
• ~25 ready-made layouts: cover, sections, bullets, columns/comparison, KPIs, timeline, process flow, icon lists, icon grids, pillars, quote, alert, tables, diagrams (funnel, pyramid, cycle, quadrant, bullseye) and image layouts.
• Curated themes + custom accent (theme:"auto" infers the theme from the content).
• Lucide-style icons bundled in the file (no network dependency for icons).
• Optional images from Unsplash (with a key) or generated via Open WebUI.
• Template Mode: build a deck by cloning shapes from a user .pptx template and overlaying content in the template safe zone.
• inspect_slides: JSON inventory of the template for the model: text, table and picture ids per slide, and every image inside the .pptx (slides, backgrounds, layout/master logos) with a ref the model can place on new slides.
• Single-file: one self-contained .py, ready to paste into the Tools registry.

See CHANGELOG.md for release notes.

Requirements

• Open WebUI >= 0.4.0
• Python: python-pptx, pillow (declared in the frontmatter → Open WebUI installs them automatically)
• Optional: httpx (fetch images from URL/Unsplash)

Installation

Option A — from the Open WebUI community

1. Open the tool page on the Open WebUI community site.
2. Click Get / Import to your instance.

Option B — manual

1. In your Open WebUI instance go to Workspace → Tools → +.
2. Paste the contents of generate_slides.py.
3. Save. The declared dependencies are installed on first use.
4. Enable the tool for the model (or chat) that should use it.

Usage

The model calls generate_slides(content), where content is a single JSON string. Minimal structure:

{
  "title": "Presentation title",
  "subtitle": "Optional subtitle",
  "author": "Author / company",
  "theme": "auto",
  "footer": "Footer label",
  "slides": [
    { "layout": "cover", "title": "...", "subtitle": "...", "icon": "cpu" },
    { "layout": "kpi_row", "title": "...", "stats": [ { "value": "-40%", "label": "..." } ] }
  ]
}

See the full example in examples/deck.json.

Classic customization (no template)

Works on the standard theme engine (no reference_file_id required):

Field	Effect
palette.primary, palette.dark, palette.accent, palette.light_accent	Override named theme colors
accent, primary (top-level)	Shortcuts for accent / primary
heading_font, body_font	PowerPoint font names (e.g. "Calibri", "Arial") for titles and body text

{
  "title": "Branded deck",
  "theme": "ocean",
  "palette": { "accent": "C99A3B", "primary": "1E3A5F" },
  "heading_font": "Calibri Light",
  "body_font": "Calibri",
  "slides": [{ "layout": "title_bullets", "title": "Hello", "bullets": ["One"] }]
}

Template Mode (3-step workflow)

Use this when the user attaches a template .pptx and wants the output to keep its look (backgrounds, logos, decorative shapes) while replacing slide content.

1. Step 1 — inspect_slides — Pass the Files API file_id (or rely on chat attachment auto-detection only in this tool). Returns a compact JSON for the model: per slide the text, table and picture ids with their verbatim text, and images[] (ref, size, slides, on). Decorations, bbox and safe zones stay server-side: decorations are cloned and safe zones applied by the tool.
2. Step 2 — Model — Builds the content JSON plus the same reference_file_id. Prefer per-slide reuse (slide, keep_ids, drop_ids, text from inspect shape "id") when mirroring each template slide; otherwise template_mapping + template_edits.
3. Step 3 — generate_slides(content) — Re-downloads the reference, clones allowed shapes, overlays layouts in the safe zone, saves the .pptx. Requires admin valve template_mode_enabled=true.

Important: generate_slides does not infer the template from chat attachments. Only an explicit reference_file_id in the JSON enables Template Mode.

Example spec: examples/template-deck-with-reference.json. Deep dive: doc/template-mode-implementation-plan.md. Model instructions (Neura): doc/neura-template-mode-model-hints.md.

Template Mode JSON fields

Field	Description
reference_file_id	Files API id of the template .pptx
slides[].reuse	Per output slide: clone template slide index + shape ids + in-place text (see examples/template-reuse-slide.json). Template tables (kind: "table") are refilled in place with text["<id>"] = {"headers": [...], "rows": [[...]]} keeping cell formatting
template_mapping	Fallback: maps layout role → reference slide index (cover, section, content, closing, default, …)
template_edits.defaults	Clone policy: drop_text, drop_placeholders, drop_offslide (conservative defaults; set drop_text: true only when the user wants template wording removed)
template_edits.slides[]	Per reference slide: index (0-based in the template file), drop_ids, optional keep_ids

With template_mode_enabled=false (default), reference_file_id is ignored and a classic deck is generated (same behavior as v1.0.3).

Post-generate edit (optional)

Separate from Template Mode’s 3-step workflow: after the user opens a deck from generate_slides, they may ask for small fixes (title too large, text overlap, split title vs body). Neura should call edit_presentation with a JSON content string — not rerun generate_slides for those tweaks.

Flow: user feedback → Neura → optional inspect_slides on the output file UUID → edit_presentation with presentation_file_id + operations[] (set_text, shrink_font, …).

Requires admin valve presentation_edit_enabled=true (default off; independent of template_mode_enabled). Input must be a Files API UUID from the last generate/edit response, not the template reference_file_id.

Docs: doc/presentation-edit-implementation-plan.md · Model hints: doc/neura-presentation-edit-hints.md · Chat examples: doc/neura-presentation-edit-chat-examples.md.

Minimal payload shape:

{ "presentation_file_id": "<uuid>", "operations": [{ "op": "set_text", "slide": 1, "shape_id": 1, "text": "..." }] }

v2 ops (same JSON): enable_autofit, replace_text_and_fit, split_text (from_shape_id, to_shape_id, mode), resize_shape (delta_height_in), set_table_cell (row, col, text), set_font_pt (font_pt or font_pt_delta). shrink_font optional "mode": fit (default), to_min, target (+ target_pt). JSON "slide" uses PowerPoint numbering when EDIT_SLIDE_INDEX_ORIGIN=1 in generate_slides.py.

Themes

auto (default, inferred from content) · midnight · forest · ocean · coral · terracotta · teal · berry · sage · cherry · charcoal · slate. You can force the accent with "accent": "#C99A3B".

Available layouts (main fields)

Layout	Main fields
cover	title, subtitle, author, eyebrow, icon, date, chips[]
section	number ("01"), eyebrow, title, lead
title_bullets	title, eyebrow, bullets[]
title_body	title, eyebrow, body (paragraphs separated by \n)
two_column_text / comparison_two	left{}, right{} or columns[] (heading, icon, points[], highlight, badge)
kpi_row	stats[] with {value, label, change}
timeline_horizontal / process_flow	steps[] with {when, title, description}
icon_list_vertical	items[] with {icon, title, description}
icon_grid_2x2 / icon_grid_3 / pillars	items[] with {icon, title, description}
chart	chart_type, labels[], values[] or datasets[]{label,data[]}, insight[]
funnel / pyramid / cycle / quadrant / bullseye	nodes[] with {label, description}
quote	quote, author, role
alert	title, level (info|tip|warning|danger), body or bullets[]
table	headers[], rows[]
text_image_right / image_left_text_right	title, bullets[]/body, image_hint or image_url or base64
image_full_caption	title, subtitle, image_hint/image_url
closing	title, eyebrow, takeaways[], contact

Screenshots

Cover	KPI row
Native chart	Funnel

Valves (configuration)

Valve	Default	Description
default_theme	auto	Default theme when the spec doesn't set one
footer_label	""	Default footer (overridden by spec.footer)
unsplash_access_key	""	Unsplash key for stock images (optional)
image_generation	false	Enable AI image generation via Open WebUI
max_image_px	1600	Maximum image width
emit_status	true	Emit status events in chat
pptx_export_dir	/app/backend/data/cache/files	Fallback directory for saving
template_mode_enabled	false	Enable reference download + template clone in generate_slides
inspect_slides_enabled	true	Allow the inspect_slides tool
inspect_extract_images	true	Upload the template images to the Files API cache on inspect (images[] is returned either way)
template_strict_mode	false	Fail generate when reference safe_zone is not computed or is too small
presentation_edit_enabled	false	Enable edit_presentation (post-generate fixes on an existing output UUID)
presentation_edit_min_font_pt	14	Floor for shrink_font when op omits min_pt
presentation_edit_shrink_step_pt	2	Step for shrink_font when op omits step_pt
presentation_edit_progress_every	5	Emit progress every N ops (0 = off)
presentation_edit_warn_noop	false	Status hint when all ops complete but none mutated the deck

Roadmap (Fase 8+)

Deferred post-v1: auto-inject inspect on .pptx upload (Open WebUI/Neura integration), optional inspect thumbnails, PNG/JPG background-only templates.

Admin notes (Neura pilot)

• Keep template_mode_enabled=false on customer instances until Template Mode is explicitly rolled out.
• Enable template_mode_enabled=true only on internal pilot workspaces.
• Presentation Edit: keep presentation_edit_enabled=false by default; enable presentation_edit_enabled=true only on internal pilot workspaces where post-generate fixes are tested. Independent of template mode. Rollout checklist and log monitoring: doc/neura-presentation-edit-hints.md — Rollout pilot Neura.
• After deploy: Workspace → Tools → paste the updated generate_slides.py and save (same Files API workflow as v1.0.3).
• Rollback: turn the valve off instantly (template_mode_enabled and/or presentation_edit_enabled), or pin the previous tool version.

How it works

1. For a custom template, the model may call inspect_slides first, then generate_slides with the same reference_file_id.
2. The model produces the JSON spec and calls generate_slides (classic or template path).
3. The engine resolves theme/accent, prefetches images (if any image layouts are present), and for each slide invokes the corresponding layout renderer (template path clones reference shapes first).
4. Shapes, text and charts are written as native OOXML objects.
5. The .pptx is saved via the Files API (fallback /cache/files) and the link is returned in chat.

Local development / testing

Requires python-pptx and (optionally) pillow/httpx:

pip install -r examples/requirements-dev.txt httpx
python examples/build.py              # → examples/demo_tech_deck.pptx
python examples/check_baseline.py     # NF5 golden regression smoke
python examples/fixtures/build_fixtures.py   # template-mode dev fixtures
python examples/parse_template_fixtures.py   # Fase 1 template parse smoke
python examples/inspect_template_fixtures.py # Fase 2 inspect JSON (offline)
python examples/inspect_extract_media_fixtures.py  # Fase 3.1 media extract (offline)
python examples/inspect_shape_kinds.py           # Fase 3.2 kind + cloneable (offline)
python examples/inspect_online_templates.py  # real Slidesgo decks in doc/online_templates/
python examples/clone_template_fixture.py    # Fase 3 clone smoke → examples/output/
python examples/generate_slides_async_smoke.py   # NF9 / NF5 / F9 / F10 async smoke
python examples/template_mode_nf6_15_slides.py # NF6: 15-slide template deck
python examples/run_production_checks.py       # full local battery

Optional: open examples/output/clone_corners.pptx in LibreOffice for a visual check.

inspect_slides in Open WebUI needs the Files API runtime; the scripts above test parse/JSON locally.

See examples/fixtures/README.md for fixture types. Local dev was verified with Python 3.11 (Docker python:3.11-slim) and pinned deps in examples/requirements-dev.txt. Record the Open WebUI Neura container Python version on deploy (≥3.10).

The file is designed to run inside Open WebUI: the open_webui.* imports are optional and the tool degrades gracefully when they're missing (handy for isolated render tests).

Contributing

Issues and PRs welcome. Please keep the file single-file and free of mandatory network dependencies for the core features.

License

MIT © IANUSTEC