Xenith RAG
Hybrid retrieval over private documents, with citations you can check.
Xenith RAG connects private document collections to an LLM through hybrid retrieval (BM25 plus dense vectors), reranking, and span-level citations, with retrieval quality measured on questions your experts wrote. We ingest the corpus, index access rules as metadata, tune recall against the golden set, and hand over a running system with re-indexing jobs and a quality dashboard. Runs in your VPC or on-premise when the corpus cannot leave.
Deployed as a scoped engagement for your stack - not a self-serve trial or live sandbox.
How it works
- Hybrid search combining vector similarity and BM25 keyword for best of both worlds
- Cross-encoder re-ranking for precision that basic RAG can't match
- Multi-format ingestion: PDF, Word, HTML, Notion, Confluence, SharePoint
- Real-time index updates: add a document and it's searchable in seconds
- Accuracy evaluation dashboard with retrieval and generation scores
Who this is for
What's included
Typical engagement deliverables
Scoped to your stack and compliance needs. We deploy the working system into your environment and hand over runbooks - not a sandbox login.
Your document archive ingested and indexed
Hybrid search live and returning accurate results
Accuracy benchmarked against 50 real queries from your team
Admin panel for adding, removing, and updating documents
API endpoint your applications can query immediately
Questions buyers ask
Frequently asked questions
- How do you stop fluent but wrong answers?
- Every claim in an answer is grounded in a retrieved span and carries a citation to the page or clause. Claims without a supporting span are suppressed. Retrieval recall is measured on a golden question set your experts wrote before we tune anything.
- Which vector database do you use?
- pgvector when you already run Postgres and the corpus is under a few million chunks; Qdrant or Weaviate for dedicated scaling or heavy filtering. Hybrid retrieval with BM25 and a reranker matters more than the database choice.
- Can it run on-premise?
- Yes. Embedding models (BGE-M3, Nomic), rerankers, and generation models (Llama 3, Qwen) run on GPUs you own. The legal research case study ran on client infrastructure with privileged documents.
- How are document permissions handled?
- Access control lists are indexed as metadata and enforced at retrieval time, so users only see chunks they could open in the source system. Permission sync runs on the re-indexing schedule.
When this is the wrong fit
- Small, static FAQ sets; a search index or a well-written page is cheaper.
- Questions answered by computing over structured data; that is an analytics problem, see Xenith Analytics.