FAQ

Do I need a vector database?

No. The default variant, ragx-lexical, ranks with BM25 and needs no embeddings. The dense variants keep their embeddings in a NumPy file next to the text mirror and search them exactly in memory.

Do I need a GPU?

Not for ragx-lexical with a hosted model. The dense variants embed the corpus once, which is much faster on a GPU but works on a CPU; ragx-rerank runs a cross-encoder on every RANK, which a GPU speeds up considerably. Local models (hf:, vLLM) need whatever their size requires.

Which variant should I use?

Start with ragx-lexical. Move to ragx-rerank for the best accuracy or ragx for the most complete pipeline when you can run an embedder. See Variants.

Does it work on single-hop questions?

Yes. The forager often answers in two or three turns, and the slot chain then has a single hop. For questions that one lookup answers, a single-shot pipeline is cheaper; ragX pays off when the evidence is spread across documents the question does not all name.

Is ragX trained or fine-tuned?

No. It works with any instruction-following model through prompts alone.

Which models work best?

Models that reliably follow a one-action-per-turn format: most models from about 4B parameters up, and every hosted frontier model tested. Very small models (below about 4B) often break the action format.

How large a corpus can it handle?

The paper's benchmark ran it on all of English Wikipedia (21M passages) with dedicated retrieval services. The library searches in memory, which comfortably covers corpora of tens of thousands to a few hundred thousand passages on one machine.

Can I see what the agent did?

Yes: answer.trace holds every step, ask_stream yields them live, and ragx ask shows them as it runs.

How do I cite ragX?

See Citing ragX.