// the find
gomate-community/TrustRAG
TrustRAG：The RAG Framework within Reliable input,Trusted output
TrustRAG is a modular Python RAG framework from an academic team at the Chinese Academy of Sciences. It covers the full pipeline — document parsing, hybrid retrieval, reranking, citation matching, and a recursive DeepResearch mode — and ships with adapters for most major vector DBs and LLM providers. The name implies adversarial robustness, but that's mostly aspirational branding.
The reranker selection is genuinely broad: pointwise (relevance generation + query generation via flan-t5), pairwise (allpair + bubblesort), setwise, and BGE cross-encoder — each backed by the actual papers, not hand-waving. The hybrid retriever with configurable BM25/dense weight mixing is straightforward to tune. The citation matching module (both rule-based and LLM-based) for grounding answers back to source passages is a real differentiator that most RAG frameworks skip. The DeepResearch pipeline's depth-first recursive loop with explicit action types (answer, reflect, search, read, coding) and a token budget guard is a reasonable architecture for multi-step research tasks.
Despite the English README, this is heavily Chinese-first: the default tokenizer (huqie) is Chinese-specific, most example documents are in Chinese, and the primary LLM examples target GLM4 and ZhipuAI — an English-only user will spend real time working around these defaults. The 'Trust' framing is mostly empty: the attack_eval module exists but is undocumented and disconnected from the main pipeline, so there's no actual robustness guarantee at inference time. The DeepResearch code is duplicated — there's one copy in api/agent/ and another in trustrag/modules/deepresearch/ with no clear ownership between them. Test coverage is nearly nonexistent: three files covering a small slice of reranking and query rewriting, nothing for document parsing or the retrieval-to-generation path.