EvidenceRAG is a document question-answering system built for technical, industrial, and engineering documents - the kind where a sentence taken out of context is worse than no answer at all. It's for teams who need to trust the answer, not just get one.

Markdown, PDF (including scanned/image-only pages via Gemini vision), Word, Excel, and CSV. Every parser produces the same hierarchical KnowledgeNode tree, so retrieval logic doesn't depend on the source format.

Hybrid BM25 + dense retrieval with reciprocal rank fusion, cross-encoder reranking, relevance grading, full parent-section reconstruction, and one bounded follow-up retrieval hop when confidence is low - all before generation even starts.

It says so. There are no hardcoded fallback answers - when retrieval confidence is too low or the evidence doesn't support a claim, the pipeline returns an explicit failure state and shows you the retrieved references instead of guessing.

Every answer ships with source file, page number, and section heading path. An explainability matrix also exposes BM25 score, dense score, fusion rank, and rerank score for every candidate chunk that was considered.

A second, cheap LLM call acts as a judge: given the generated answer and the exact chunks it was allowed to use, it returns a 0-1 support score and flags any unsupported claims. This catches hallucination that retrieval metrics alone can't.