Configuration

Every setting is a keyword argument of RagX (or a field of ragx.Settings):

python

RagX("./docs", variant="ragx", steps=8, backfill=24, embed_model="BAAI/bge-m3", device="cuda")

Models

Setting	Default	Meaning
`model`	from the environment	`provider:model`; see Models and providers
`scout_model`	`model`	model for the forage loop and the slot chain
`answer_model`	`model`	model for the compose call
`think`	the variant's (`ragx-think`: on)	thinking mode in every call
`think_budget`	8192	extra tokens a thinking call may spend
`max_context_tokens`	32768 for `vllm:`/`hf:`, none otherwise	the reader's context; the evidence is trimmed to fit
`temperature`	0.0	sampling temperature
`cache`	`False`	memoise identical model calls and tool results within the process

Setting	Default	Meaning
`variant`	`ragx-lexical`	one of `ragx`, `ragx-core`, `ragx-lexical`, `ragx-fused`, `ragx-rerank`, `ragx-fast`, `ragx-think`, `classic`
`retriever`	the variant's	override the ranker: `bm25`, `dense`, `fused`, `rerank`
`backfill`	the variant's (48 or 0)	the question's own top passages appended at compose
`steps`	10	forager turns
`rank_k`	8	passages listed per `RANK`
`hop_k`	4	passages retrieved per hop
`max_hops`	4	sub-questions at most
`read_cap`	8	read passages composed
`extra_cap`	12	extra passages composed (hop passages first)
`tools`	`["RANK", "READ"]`	forager tools; add `"SEARCH"` for line search
`fence`	`True`	fence tool output and evidence as untrusted data
`compose_reasoning`	`False`	a reasoning pass before the compose: `True`, or `"auto"` for number, date and comparison questions
`typed_hints`	`True`	the answer-type hint in the compose prompt

Setting	Default	Meaning
`embed_model`	`Qwen/Qwen3-Embedding-0.6B`	sentence-transformers embedder for dense ranking
`rerank_model`	`BAAI/bge-reranker-v2-m3`	cross-encoder for `rerank`
`rerank_depth`	100	dense candidates the reranker (and fusion) consider
`device`	automatic	`"cpu"`, `"cuda"`, ... for the embedder and reranker
`passage_words`	120	passage size when files are cut

Setting	Default	Meaning
`verify_citations`	`True`	cite only passages that contain the answer (grounding)
`abstain_on_weak_evidence`	`False`	allow "I don't know" and replace ungrounded answers with a refusal

Setting	Default	Meaning
`cache_dir`	`<corpus>/.ragx`	where the mirror, index and embeddings live
`include`	`[]` (everything supported)	glob allowlist
`exclude`	`[".git/", "node_modules/", ".ragx/**"]`	glob blocklist

Used only with variant="classic"; see The classic agent.

Setting	Default	Meaning
`route`	`agentic`	`agentic`, `adaptive`, `deep` or `both`
`max_steps`	5	tool-loop turns
`auto_budget`	`True`	fewer turns and extras for scouts of 2.5B parameters or less
`max_result_lines`	40	search lines shown per call
`max_read_chars`	4000	characters per read
`prerank`	`True`	seed the loop with a BM25 ranking of likely files
`single_shot_reads`	4	files the `adaptive` route reads first; files per hop in `deep`
`deep_hops`	4	sub-questions in `deep`
`deep_surf_cap`	12	surfaced files composed in `deep`
`deep_chain_notes`	`False`	include the resolved chain in the compose prompt
`surf_cap`	16	surfaced files composed in `agentic`
`compose_surfaced`	`True`	compose over surfaced files, not only reads
`semantic_fallback`	`False`	offer the `SEMANTIC` tool (`pyragx[semantic]`)
`fuzzy`	`True`	offer `FUZZY` when ugrep is installed
`context_lines`	1	lines of context around each search hit
`sandbox_observations`	`False`	fence tool output as quoted data