FAQ#

Does my data leave my network?#

Not unless you choose a hosted model. Self-hosted deployments run embeddings, generation, and storage entirely on your infrastructure. Hosted LLM providers only ever see the retrieved chunks in the prompt — never your raw corpus.

Which vector database should I use?#

Start with the embedded store in development (zero ops), move to Qdrant when you self-host at scale, or a managed store when you'd rather not run one. The storage interface is identical — switching is a config change and a re-index.

How do citations stay accurate?#

Chunks carry exact provenance (file, page, position) from ingestion, and the model cites by index into the retrieved set — so a [2] can only ever point at something that was actually retrieved. Confidence scores flag answers whose citation coverage is thin.

What happens when the answer isn't in my documents?#

The system prompt requires the model to say so rather than guess, and the low retrieval scores drive the confidence score down. You can route these to a fallback UX or a human queue.

How big can a corpus get?#

HNSW-indexed collections serve top-k in single-digit milliseconds into the tens of millions of vectors per node; beyond that, collections shard. BM25 sidecars scale with the same partitioning.

Can I bring my own models?#

Yes — embeddings, rerankers, and LLMs all sit behind provider registries. Implement the interface, register it, select it with an env var. See the model explorer for what's supported out of the box.