An AI agent is only as trustworthy as what it answers from. Retrieval-augmented generation (RAG) grounds every answer in your organisation's own content rather than the model's memory of the public internet. ATX Agent Hub makes RAG a governed, project-scoped capability — and shows its work on every answer.
Knowledge base collections ingest PDF, DOCX, Markdown, HTML and CSV files, Confluence and SharePoint pages and web crawls. Content is chunked, embedded and indexed on a schedule you choose, so the knowledge base stays current as source documents change. A ready-made workflow pattern takes a Confluence space straight into a knowledge base.
Chunking, embedding model and retriever are configurable per collection. Hybrid search combines semantic similarity with keyword matching, and re-ranking orders the candidates before they reach the model — the difference between an answer that quotes the right clause and one that quotes a similar-sounding one.
Every answer carries its sources. Chat with Data offers four citation styles — minimal, inline, full (document, section, timestamp or paragraph) and direct source snippets — plus a confidence score and document lineage down to the chunk used. Answer mode and answer language are configurable per project.
Knowledge bases belong to a project, and access can be set per source. An agent can only retrieve from collections its project owns, so a policy library indexed for one department is not quietly available to another department's chatbot.
When the platform is self-hosted, the documents, the embeddings and the vector collections stay in your own cloud tenancy or data centre, and embedding and answer models are called through your own endpoints. For many regulated teams that is the condition for putting internal documents into RAG at all.
Retrieval-augmented generation with the controls an organisation needs: private ingestion of internal content, access control per source, configurable retrieval, citations on every answer, and deployment within the organisation's own environment.
PDF, DOCX, Markdown, HTML and CSV files, Confluence and SharePoint pages, and web crawls.
Each answer shows its citations in the chosen style, a confidence score and document lineage to the exact chunk used, so it can be checked against the source.
This site needs JavaScript for the full experience. Every page above is reachable without it.