Give your team a search box that answers in your voice, with citations.

Permissions are resolved at query time, not at index time. The retriever filters candidates by the asking user's identity through your source-system ACLs, so a user with no access to a document never sees it cited — and we prove this on an adversarial test set every release.

Three layers. The prompt requires citation; uncited generations are flagged and routed back through retrieval; the eval harness measures citation faithfulness on every release. We publish the score in your repo so regressions are loud.

Sometimes. We start with the strongest hosted embedding for the domain, then evaluate whether a small domain-tuned model measurably shifts retrieval scores against the question set. We write the cost-quality trade in the decision log.

In your cloud, in the vector store you choose — Qdrant, pgvector, or the managed option from your hyperscaler. We do not host indexes for clients, and source documents never leave your tenant.

Yes, through typed tools, with the same approval and audit boundaries our agentic systems practice uses. Read-only by default; writes require an explicit tool and a human in the loop where the risk warrants it.