Changelog

0.0.1

The first public release.

• The ragX method: the forage loop (RANK, READ), the slot chain and one compose call, with the paper's prompts and frozen settings.
• Seven variants: ragx, ragx-core, ragx-lexical (default), ragx-fused, ragx-rerank, ragx-fast and ragx-think, plus the classic file-level agent.
• Rankers: BM25, dense (any sentence-transformers model, cached embeddings), reciprocal-rank fusion, and a cross-encoder reranker.
• Passages with path:start-end names, and citations grounded in the passages that contain the answer.
• Providers: OpenAI (and any OpenAI-compatible endpoint), Anthropic, Google, Groq, Together, Cerebras, Fireworks, vLLM and local Hugging Face models; reasoning effort and thinking mode.
• Interfaces: Python API with streaming, the ragx command line, and an OpenAI-compatible server.
• Security: tool output and evidence fenced as untrusted data by default.