forge-rag

What problem does it solve?

It prevents RAG systems from hallucinating and returning unverifiable answers by enforcing retrieval quality, chunking structure, citation discipline, and refusal behavior.

Core Features & Use Cases

• Semantic chunking with hierarchy: Chunks by meaningful units and preserves document structure so the model can answer with the right context.
• Hybrid retrieval with reranking: Combines BM25 and dense retrieval, then uses cross-encoder reranking to select the most relevant passages.
• Citation and refusal guardrails: Requires inline citations to retrieved chunks and includes a low-confidence refusal path to avoid fabrication.
• Freshness and evaluation-first iteration: Supports version-aware retrieval, scheduled indexing, and an eval set that covers both answerable and refusal cases.

Quick Start

Ask your AI to redesign your RAG pipeline to use semantic hierarchical chunking, hybrid BM25+dense retrieval with reranking, and a generation prompt that requires inline citations to verified retrieved chunks or returns a refusal when no good passages are found.