Varagity
Drag or use arrow keys to turn the model. Press Home to return to its original orientation.
retrieval, with context.
A self-hosted RAG system that turns a document collection into grounded, cited answers. It combines semantic and keyword search on local GPUs, with an evidence panel that lets you follow each answer back to the passages it used.
Documents are parsed, split into chunks, and given a short context blurb before being indexed in pgvector and Elasticsearch. An optional reranker refines the matches before generation. The web app exposes those steps through source citations and per-chunk scores, while Prefect and Grafana make the pipeline easier to inspect. A separate message-graph workflow explores relationships across an imported conversation history.
- Hybrid semantic and BM25 retrieval
- Source citations and per-chunk evidence
- Local GPU inference and pipeline observability
Inside Varagity.
A conversation with the document collection, with retrieval evidence alongside the answer.
A view of the modules and connections that make up the application.
Prefect tracks the tasks and runs behind the document pipeline.
Grafana brings query, ingestion, and infrastructure metrics into view.