Eight bundled connectors, scheduled re-ingest, source ACLs preserved. A single Knowledge Query node sits before any LLM step. Whatever the user can't see, the agent can't surface — without you wiring up a thing.
The bundled set covers the corpus most enterprises actually want grounded — wikis, drives, the warehouse, and the public web. Anything else, you wire through an MCP server.
Each source flows through the same five-stage pipeline. You configure the source; the pipeline does the rest. Rebuild full or incremental, on a schedule or on demand.
Vendor-specific connector pulls only what changed since the last sync.
PDF, DOCX, HTML, MD, structured rows. OCR fallback for scanned PDFs.
Headings, sections, table rows. Chunk size tuned per source type, not globally.
Bedrock, OpenAI, or your own gateway. Re-embeds on model upgrade — never silently.
Stored alongside the source's permission tuple. Query-time filter happens before retrieval.
Most RAG systems retrieve first and filter after. That leaks: timing channels, ranking signal, "we know it exists, you just can't read it." Nexus stores the source's ACL alongside every chunk. The user's identity flows through the agent into the retrieval call, and the index filters before similarity search runs. If your user can't read the Confluence page, the chunk is never a candidate. Not redacted. Not sorted away. Never seen.
We'll connect one of your sources during the demo, run an ingest, and show you a Knowledge Query node grounding a Claude response with citations.