pdf2beamer 0.1.0

Local-first conversion of native scientific PDFs into editable Beamer decks.

pdf2beamer

pdf2beamer is a local-first Python package for converting native scientific PDF papers into editable, compilable Beamer presentations.

The project is intentionally structured around inspectable intermediate representations:

PDF -> PaperIR -> ArgumentGraph -> DeckPlan -> SlideIR -> Beamer

The package defines strict Pydantic v2 data models and keeps real Docling, PyMuPDF, Nemotron generation, validation, rendering, and compilation behind local integration points.

Constraints

• No external API calls.
• No OCR or scanned-PDF fallback.
• Local Nemotron generation plus Qwen embedding/reranking adapters are dependency-injected.
• LLM components generate structured JSON only.
• Beamer is rendered deterministically from SlideIR.

Public API Sketch

from pdf2beamer import PdfToBeamerPipeline, PipelineConfig

config = PipelineConfig(
    model_path="./models/nemotron-3-nano-4b-gguf/NVIDIA-Nemotron-3-Nano-4B-Q4_K_M.gguf",
    embedding_model_path="./models/Qwen3-Embedding-0.6B",
    reranker_model_path="./models/Qwen3-Reranker-0.6B",
    duration_minutes=10,
    audience="technical",
    theme="clean",
)

pipeline = PdfToBeamerPipeline(config)
result = pipeline.generate("paper.pdf", "out/")

Local Models

Real model and PDF backends are optional. The base package imports without installing heavy extraction or model dependencies, and the library never downloads model files at runtime.

Install the extras you need:

pip install -e ".[pdf,docling,latex,models]"

Expected local files, auto-detected by --real-models when present:

• Generation: models/nemotron-3-nano-4b-gguf/NVIDIA-Nemotron-3-Nano-4B-Q4_K_M.gguf
• Embedding: models/Qwen3-Embedding-0.6B
• Reranking: models/Qwen3-Reranker-0.6B

You can override paths with --model, --embedding, or --reranker.

Model files are local assets and should not be committed. Store them under models/; .gitignore excludes models/ and common model-weight formats.

Download Models From Hugging Face

Install the models extra, which includes Hugging Face download tooling:

pip install -e ".[pdf,docling,latex,models]"

If your Hugging Face account needs access to a model, authenticate once:

hf auth login

Download the expected local model layout:

mkdir -p models/nemotron-3-nano-4b-gguf \
  models/Qwen3-Embedding-0.6B \
  models/Qwen3-Reranker-0.6B

hf download nvidia/NVIDIA-Nemotron-3-Nano-4B-GGUF \
  NVIDIA-Nemotron-3-Nano-4B-Q4_K_M.gguf \
  --local-dir models/nemotron-3-nano-4b-gguf

hf download Qwen/Qwen3-Embedding-0.6B \
  --local-dir models/Qwen3-Embedding-0.6B

hf download Qwen/Qwen3-Reranker-0.6B \
  --local-dir models/Qwen3-Reranker-0.6B

Quick local check:

test -f models/nemotron-3-nano-4b-gguf/NVIDIA-Nemotron-3-Nano-4B-Q4_K_M.gguf
test -d models/Qwen3-Embedding-0.6B
test -d models/Qwen3-Reranker-0.6B
git check-ignore -v models/nemotron-3-nano-4b-gguf/NVIDIA-Nemotron-3-Nano-4B-Q4_K_M.gguf

Then run with real local models. Use --no-compile if you only want the editable out/main.tex file:

uv run --extra pdf --extra docling --extra models pdf2beamer generate paper.pdf --real-models --no-compile

LaTeX Compilation

pdf2beamer always writes out/main.tex. To also produce out/main.pdf, install a TeX distribution that provides the latexmk command, then run without --no-compile.

Debian/Ubuntu:

sudo apt update
sudo apt install latexmk texlive-latex-recommended texlive-latex-extra texlive-fonts-recommended

Windows:

winget install --id MiKTeX.MiKTeX --exact
winget install --id StrawberryPerl.StrawberryPerl --exact

MiKTeX provides the TeX toolchain, and latexmk needs Perl on Windows. Restart the terminal after installation so the updated PATH is visible.

macOS:

brew install --cask mactex-no-gui

Check that latexmk is available:

latexmk --version

Compile during generation:

uv run --extra pdf --extra docling --extra models pdf2beamer generate paper.pdf --real-models

Fake-model command for lightweight local development:

pdf2beamer generate paper.pdf \
  --use-fake-models \
  --duration 10 \
  --output out/

Real local-model command:

pdf2beamer generate paper.pdf \
  --real-models \
  --duration 10 \
  --audience technical \
  --output out/

Structured GGUF Output

The GGUF generator is exposed as LocalNemotronGenerator and loads a local Nemotron instruct GGUF through llama-cpp-python. For ArgumentGraph and SlideIR, it first tries Instructor's local llama-cpp-python integration:

instructor.patch(
    create=llama.create_chat_completion_openai_v1,
    mode=instructor.Mode.JSON_SCHEMA,
)

This path returns Pydantic response models directly and retries validation failures. If Instructor or the OpenAI-compatible llama.cpp method is unavailable, the generator falls back to llama.cpp response_format JSON schema/object mode, then to strict JSON parsing.

CLI controls:

pdf2beamer generate paper.pdf \
  --real-models \
  --instructor \
  --instructor-max-retries 2 \
  --no-compile \
  --output out/

Disable Instructor and use llama.cpp response format fallback:

pdf2beamer generate paper.pdf --real-models --no-instructor --output out/

13. 12_quality_report.json - synthese qualite finale.
14. assets/ - images extraites du PDF et reutilisables dans les slides.