Features¶

PPT-Werkstatt covers the path from uploaded material to finished slide deck and supports step-by-step revision through chat afterwards. The features fall into the areas of material ingestion, briefing, slide generation, detail editing, and export.

Use cases¶

Generate a slide deck from a set of documents — Several documents are uploaded; the application segments them into typed content entries and derives a content-grounded presentation. The material may be thematically broad or narrow.
Adapt the slide deck to audience and presentation duration — A briefing dialogue clarifies purpose, audience, language, tone, and presentation duration; slide count and content depth follow accordingly.
Provide structural preferences and topic priorities — A rough outline or topic priority list can be entered as free text before generation; the application reflects this in the slide sequence.
Add icons and diagram graphics — Optionally, icons matching the slide content are selected; for suitable content, graphical layouts (processes, cycles, pyramids, matrices, hierarchies) are proposed instead of plain text slides.
Revise slides through chat — After generation, individual slides or the whole deck can be adjusted through chat instructions: shorten, rewrite, change layout, insert new slides, reorder, group.
Use existing slide decks as material — Uploaded PPTX files are also indexed as content sources; their contents feed into slide generation like any other document.

At a glance¶

Connection to four model services: primary LLM, fast LLM, embedder, reranker.
Import: PDF, DOCX, PPTX, Markdown, plain text.
Export: PPTX (with custom template), DOCX, Markdown.
Over thirty slide layouts in six content categories.
Three working phases: briefing dialogue, automatic slide generation, detail editing in chat.
Multi-stage quality assurance with an automatic correction loop.
Versioning with undo; visible placeholders for material gaps.

Material ingestion¶

Uploaded documents are read, segmented into content units, and assigned semantic vectors. Larger documents are chunked along recognised structural markers (headings, page breaks, slide breaks) so that the analysis works on meaningful units.

Supported document formats:

PDF — text extraction with page breaks as structural markers.
DOCX — paragraph text, headings, and tables are processed.
PPTX — slide texts and speaker notes are extracted, so existing decks can also be used as source material.
Markdown — headings and lists are recognised as structural markers.
Plain text (TXT) — taken in unchanged as a content block.

Indexed entries are typed (text, table, list, decision, quote, metric, heading) and labelled with title, summary, keywords, and source position. The inventory is held in memory for the duration of the session and not persisted.

Briefing dialogue¶

The chat clarifies the framing parameters of the presentation iteratively: purpose, audience, presentation duration, language, tone, as well as topic priorities and exclusions. The briefing agent draws on the existing inventory as context and works through the open points until the briefing can be approved.

Slide generation¶

From the approved briefing and the inventory, a multi-stage pipeline generates a complete slide deck:

A coarse structure defines slide titles, sequence, and slide roles (intro, body, summary, divider).
Each slide is assigned a suitable layout from the catalogue.
The slot inventory of each layout is filled in parallel with content from the material.
Visible placeholders are inserted when a required slot cannot be covered from the material; this surfaces gaps rather than closing them through model speculation.

Layout catalogue¶

The catalogue covers six content categories — text, comparison, process, relation, metric, and visual. Concretely, the layouts include bullet lists, structured lists with sub-points, pro/con boxes, two-column comparisons, section dividers, quote highlights, box and card grids, comparison tables, large metrics, simple tables, process flows (horizontal, vertical, chevron), funnels, timelines, cycles (basic, radial, segmented), org charts, hierarchies, pyramids (upright, inverted, segmented), matrices (basic, labelled, SWOT), and radial diagrams.

Icons and diagram graphics¶

Optionally, two visual elements are added:

Icons — A curated subset of an open-source icon set (Bootstrap Icons, MIT licence; a pool of around 150 icons suitable for presentations) is used to select content-appropriate icons per slide.
Diagram graphics — Diagram-style layouts are produced through an integrated external template project that renders content as SVG. The SVG is then embedded into the slide as a raster image.

Quality assurance¶

Three mechanisms safeguard the material fidelity of generated slides:

Rule-based slot check — Each slot of a layout has constraints (required, character limits, minimum and maximum number of list items). Violations are reported directly.
Embedding coverage — Slide texts are matched semantically against the material; slides with low material coverage are flagged.
LLM hallucination check — A validator call inspects the slides for claims not supported by the material and assigns an overall verdict (ok / concerns / problematic).

Flagged slides enter an automatic correction loop (up to two iterations). Remaining mandatory gaps are exposed as visible placeholders. A final homogenization stage checks the finished deck for inconsistencies in tone and terminology and proposes corrections.

Detail editing in chat¶

After generation, the application switches to a chat-based editing mode. Free-text instructions are classified into one of 28 action types grouped into five categories:

Structure — insert, delete, move, duplicate, merge, split, group, or address slides by group.
Content — shorten, expand, sharpen language, rewrite, modify selectively, translate, adjust tone.
Form — change layout, adjust bullets, rename slides, insert graphics.
Meta — speaker notes, metadata, undo, export, status, overall overview.
Navigation — move the editing focus to a slide or group, ask questions about the material or about a specific slide.

A classification check intervenes when confidence is low, when multiple actions are detected, or when a context conflict is found, and asks back instead of executing an unclear instruction.

Source attribution and versioning¶

Every slide content carries a reference to the underlying inventory entries. Material can also be addressed directly through deterministic references (entry IDs, document/page references), which avoids an LLM call for source resolution during detail editing. Before any modifying operation, a version snapshot is stored; the last 15 states can be reverted through undo.

Export¶

The finished presentation can be produced in three formats:

PPTX — A custom PPTX or POTX template is preserved; master layouts, fonts, and colours remain intact. Existing slides in the template are replaced.
DOCX — Readable print variant with slide titles, content, and speaker notes.
Markdown — Structured intermediate format for further processing or preview.