Local RAG for Lab Notes and Slides on AMD and NVIDIA DGX Spark
I sometimes remember a figure from a lab meeting but forget which slide deck it came from. I built a local retrieval-augmented generation (RAG) system so we can ask questions about our Markdown notes and PowerPoint slides in Japanese or English and find the source of an answer.
The AMD Ryzen AI Max+ 395 workstation indexes the documents and retrieves relevant passages. DeepSeek runs across two NVIDIA DGX Spark systems and uses those passages to answer. Open WebUI shows the answer and its sources together.
Making slide figures searchable
A text extractor may capture a chart’s title but miss the figure itself. Two models running through Ollama on the AMD workstation make that content searchable:
- Gemma 4 31B describes figures from rendered slides during indexing. The index labels each description as AI-generated so readers can distinguish it from the slide author’s text.
- Qwen3-Embedding-4B embeds both the document text and Gemma’s descriptions. It embeds questions in the same vector space, so a question and a relevant passage can use different words.
Qdrant stores each vector with the passage it represents, the file path, and the starting line or slide number.

Answering questions with retrieved passages
For each question in Open WebUI, the AMD workstation retrieves up to four passages from Qdrant and sends them with the question to the local API for DeepSeek-V4.1-Flash-1017. TensorFold runs that single model across both DGX Spark systems.
The service checks the citation IDs it recognizes in the answer against the IDs it sent to DeepSeek. The source list includes every passage supplied to the model, even if the answer doesn’t cite it. Researchers can open a note at the listed line or return to the original slide to check a chart’s legend or values.
Related article
Building a Local RAG System for Lab Notes and Slides with AMD and DGX Spark
Related projects
Tools & Technologies
Python · Qdrant · Qwen3-Embedding-4B · Gemma 4 31B · Ollama · AMD Ryzen AI Max+ 395 · DeepSeek-V4.1-Flash-1017 · NVIDIA DGX Spark · Open WebUI
