Qwen PDF VDB POC

This repository contains a local proof of concept that:

• reads all PDFs from sources/
• renders every page to an image
• extracts native PDF text with PyMuPDF
• stores page renders as real image vectors
• stores embedded PDF images as separate real image vectors when the PDF exposes them
• embeds native text chunks with the same multimodal model
• writes everything into a local Qdrant database inside this repo
• supports retrieval-only search and ask commands against the joint text/image vector space

Stack

• Qwen3-VL-Embedding-2B for true multimodal embeddings
• PyMuPDF for PDF text extraction and rendering
• Qdrant local mode for persistent vector storage
• local repo-scoped .venv/, .hf/, and models/

Repo-local runtime

• Virtual environment: .venv/
• Hugging Face cache: .hf/
• Embedding model: models/Qwen3-VL-Embedding-2B/
• Vector DB: data/qdrant/
• Page renders and extracted assets: artifacts/page_renders/
• Run reports: artifacts/reports/
• Future-agent docs: .codex/

Install or refresh dependencies

python -m pip --python .\.venv\Scripts\python.exe install torch torchvision --index-url https://download.pytorch.org/whl/cu128 --extra-index-url https://pypi.org/simple transformers safetensors qwen-vl-utils huggingface-hub accelerate pymupdf qdrant-client pillow

Download the embedding model into this repo

New-Item -ItemType Directory -Force .hf, .hf\hub, models\Qwen3-VL-Embedding-2B | Out-Null
.\.venv\Scripts\hf.exe download Qwen/Qwen3-VL-Embedding-2B --cache-dir .\.hf\hub --local-dir .\models\Qwen3-VL-Embedding-2B --max-workers 4

Ingest all PDFs

.\.venv\Scripts\python.exe .\main.py ingest --recreate --embed-batch-size 1

Optional flags:

• --limit-pages 3 limits ingest for quick tests
• --render-dpi 96 changes page render resolution
• --embed-batch-size 2 can improve throughput if VRAM allows it
• --collection my_collection changes the local Qdrant collection name

Multimodal search

.\.venv\Scripts\python.exe .\main.py search "Welche Designregeln fuer Perception werden vorgestellt?" --limit 6

The search operates in one shared embedding space across:

• text chunk records
• full page-image records
• embedded PDF image records

Ask questions / retrieve contexts

.\.venv\Scripts\python.exe .\main.py ask "Welche Designregeln fuer Perception werden vorgestellt?" --limit 8

This retrieves the most relevant multimodal contexts for a natural-language question and returns the matching chunks/pages directly.

Stats

.\.venv\Scripts\python.exe .\main.py stats

Metrics

The ingest report tracks:

• pages per second
• vector records per second
• embedding items per second
• embedding input tokens per second
• text-vs-image embedding item counts

Important note:

• page and extracted-image records are embedded as images, not by converting them to text first