Retrievers

The ranker answers every RANK, every hop of the slot chain and the backfill. Each variant picks one; retriever= overrides it.

BM25

RagX("./docs", retriever="bm25")

Okapi BM25 over each passage's title and text, from an in-memory inverted index built when the first question is asked. It needs no extra dependencies and no GPU, and handles corpora of tens of thousands of passages comfortably.

Dense

RagX("./docs", retriever="dense", embed_model="Qwen/Qwen3-Embedding-0.6B", device="cuda")

Exact inner-product search over normalised passage embeddings from any sentence-transformers model. The default is the paper's embedder, Qwen/Qwen3-Embedding-0.6B. Queries use the model's query prompt when the model defines one.

Embeddings are cached in <corpus>/.ragx/embeddings/ keyed by passage content, so the corpus is embedded once and, after a change, only new or edited passages are embedded again.

Requires pip install "pyragx[dense]". Set device="cpu" or device="cuda" to choose where the embedder runs; by default sentence-transformers picks a GPU when one is available.

Fused

RagX("./docs", retriever="fused")

Reciprocal-rank fusion of the BM25 and dense rankings: each passage scores 1 / (60 + rank) in each of the two top-100 lists, and the sums are ranked. It combines exact-term matching with semantic matching.

Rerank

RagX("./docs", retriever="rerank", rerank_model="BAAI/bge-reranker-v2-m3", rerank_depth=100)

The dense top rerank_depth passages (100 by default) are scored by a cross-encoder, default BAAI/bge-reranker-v2-m3, and reordered. The cross-encoder's input is capped at the model's position limit, so models with 512 positions work as well as long-context ones.

Retrieval without a model

RagX.search runs the ranker alone and makes no LLM call, which is useful for checking retrieval:

rag = RagX("./docs", variant="ragx-fused")
rag.search("Ashford University founder", top=5)
# ['ashford_university.md:1-5', 'thomas_wren.md:1-5', ...]

ragx search "Ashford University founder" -c ./docs -v ragx-fused -k 5