A fully local, open-source agentic pipeline that takes a PDF and a natural language query, reasons about the right presentation structure, and generates a styled
.pptxfile — no cloud APIs, no data leaving your machine.
# 1 — Ingest a PDF (run once per document)
python main.py offline path/to/paper.pdf
# 2 — Generate a presentation
python main.py online "make a 10 slide technical presentation on transformer architecture" Attention_is_all_you_NeedOutput: a styled .pptx in ./outputs/ with titles, bullets, speaker notes, and a takeaway bar per slide.
The system follows a two-path design:
OFFLINE (once per PDF)
PDF → ingest → chunker → embedder → Milvus Lite
ONLINE (per query) — LangGraph StateGraph
Query → S0 QueryCompiler → S3c Retrieval → S4 Strategy → S4c grade
→ S5 Planner → S5c grade → S6 ContentGen → S7 Compress
→ S9 StyleResolver → S11 Renderer → .pptx
The online path runs as a LangGraph StateGraph with conditional retry loops at both the strategy and planner stages. Critics are deterministic Python — no extra LLM calls at validation time.
| Component | Tool |
|---|---|
| PDF parsing | PyMuPDF (fitz) |
| Chunking | LangChain RecursiveCharacterTextSplitter |
| Embeddings | all-MiniLM-L6-v2 (sentence-transformers) |
| Vector store | Milvus Lite |
| LLM | qwen2.5:14b via Ollama |
| Agent orchestration | LangGraph |
| PPTX generation | python-pptx |
| Config | YAML + Pydantic |
# Create environment
conda create -n pdf-agent python=3.11 -y
conda activate pdf-agent
# Install dependencies
pip install -r requirements.txt
# Fix OMP conflict on macOS Apple Silicon
conda env config vars set KMP_DUPLICATE_LIB_OK=TRUE -n pdf-agent
conda deactivate && conda activate pdf-agent
# Pull the LLM
ollama pull qwen2.5:14b# Technical overview
python main.py online "make a 10 slide technical presentation on the transformer architecture" Attention_is_all_you_Need
# Executive summary
python main.py online "create a 5 slide executive summary of the attention mechanism" Attention_is_all_you_Need
# Beginner tutorial with dark theme
python main.py online "build a beginner tutorial on self-attention dark theme" Attention_is_all_you_Need| Keywords | Theme |
|---|---|
dark, minimal |
Dark background, blue accent |
corporate, formal |
White, navy accent |
colorful, creative |
Light grey, pink accent |
| (default) | White, blue accent |
pdf-agent/
├── main.py # CLI entry point
├── config.yaml # model names, chunk params, retrieval settings
├── requirements.txt
├── README.md
├── README.html # Styled full-page README
└── src/
├── models.py # Pydantic artifact definitions
├── utils.py # shared LLM caller, JSON parser, logger
├── ingest.py # S1 — PDF → DocTree
├── chunker.py # S3a — DocTree → list[Chunk]
├── embedder.py # S3b — list[Chunk] → Milvus
├── retrieval.py # S3c — query → EvidencePack
├── query_compiler.py # S0 — raw string → PresentationRequest
├── strategy.py # S4 — PresentationStrategy (LLM)
├── planner.py # S5 — SlidePlan (LLM)
├── content.py # S6 — SlideContent (LLM)
├── compression.py # S7 — trim bullets
├── visual.py # S8 — VisualSpec per slide
├── style.py # S9 — StyleConfig from description
├── layout.py # S10 — LayoutSpec (deterministic)
├── renderer.py # S11 — python-pptx → .pptx
└── agent.py # LangGraph StateGraph wiring
- Phase 1 — MVP (offline ingestion + online generation + PPTX render)
- Phase 2 — LangGraph agent loop with plan critic
- Phase 3 — Semantic DocTree + BGE-M3 embeddings + hybrid search
- Phase 3 — S4 strategy critic + S7 compression agent
- Phase 4 — Gradio UI + FastAPI endpoint
- Phase 5 — Evaluation harness + open-source release
Typed artifacts at every boundary — every stage produces a Pydantic model. No raw strings between stages.
Constraints propagate as data — slide count, audience, style extracted at S0 and enforced downstream mechanically.
Offline / online separation — ingestion happens once per document. Query-time work starts at retrieval. These paths never mix.
Critics are cheap — the plan critic is deterministic Python. LLM calls only when schema validation fails.
Built during an internship at Unicloud. Architecture informed by synthesis of multiple agentic RAG patterns. Built on top of Milvus, LangGraph, and Ollama.