Applies to VAST Cluster 5.5.
Complete applications built on VAST DataEngine, published by VAST on GitHub. Each one runs a DataEngine pipeline next to its data and writes the results to VAST DataBase, with an application on Kubernetes on top. Use them to see how the pieces in this section fit together in a real system.
They aren't verified against these docs. These pages haven't run either application, and their READMEs don't state which VAST Cluster version they target. Check each repository's own deployment guide before you use it.
| Application | What it does | DataEngine in it | Other parts |
|---|---|---|---|
| VSS blueprint | Video search and analysis: search uploaded video by what happens in it | An ingest pipeline that segments, analyses and embeds each uploaded video, and a scheduled enrichment function | VAST DataBase for segment rows and vectors, S3 for the video, NVIDIA NIM and Cosmos models on a GPU host, a web UI and REST API on Kubernetes |
| Research assistant agent | A RAG chat assistant over your documents | An event-based ingestion pipeline that chunks and embeds each uploaded document | VAST DataBase for the vectors and the conversation history, S3 for the documents, a chat UI and agent API on Kubernetes, a choice of LLM provider |
VSS blueprint
github.com/vast-data/vss-blueprint brings NVIDIA's video search and summarization (VSS) reference designs to VAST. Videos uploaded to a bucket go through an ingest pipeline of DataEngine functions, in order:
- Segment: split each video into short clips.
- Detect: run object detection on each clip.
- Reason: describe what happens in each clip as searchable text, with NVIDIA Cosmos-Reason2.
- Embed: compute vectors from the description and from the clip itself.
- Write: store the segment rows and their vectors in VAST DataBase.
An optional enrichment function runs on a schedule. It samples recent descriptions, suggests search terms and key events, and writes them to a VAST DataBase table.
The models run as containers on a separate GPU host. A web UI and REST API run on Kubernetes and search the vectors in VAST DataBase.
What the README says you need: VAST DataEngine, a Kubernetes cluster, a container registry for the images, a GPU host for the models, and VAST credentials. The repository's deployment guides cover the models, the Kubernetes application, and the pipeline, which you can create from the Web UI or with vastde.
The pages in this section that match its parts:
- Write a function, triggers and pipelines as code: the ingest pipeline and the scheduled function.
- VAST DataBase and Vector Search from a function: the writer step.
Research assistant agent
github.com/vast-data/vast-research-assistant-agent is a retrieval-augmented generation (RAG) assistant. Documents uploaded to an S3 bucket start an event-based DataEngine ingestion pipeline. It chunks each document, embeds the chunks and stores the vectors in VAST DataBase.
A chat UI and an agent API run on Kubernetes. The agent retrieves chunks from VAST DataBase, keeps each conversation's history in VAST DataBase too, streams answers with citations and can add web search results. The language model is NVIDIA NIM by default, or Google Gemini or Vertex AI, chosen with one setting.
What the README says you need: VAST DataEngine, a Kubernetes cluster for the chat application, and credentials for your chosen model provider. GPUs are needed only if you run NIM yourself. The repository has deployment guides for the Kubernetes application and for the ingestion pipeline.
The pages that match its parts:
- Triggers and pipelines as code: the ingestion pipeline.
- VAST DataBase and Vector Search from a function: writing embeddings from a function.
- Build RAG retrieval with permissions, in VAST Vector Search: the retrieval side, including returning only chunks a user may read.
See also
- Overview: what DataEngine is for.
- End-to-end patterns: the same patterns as outlines, step by step.