RAG with receipts: citations, rerankers, and recursive deep search

What it does TrustRAG is a Python framework that assembles retrieval-augmented generation pipelines from pluggable parts: document parsing, hybrid retrieval (BM25 plus dense), multiple reranking strategies, citation generation, and a recursive “DeepResearch” mode that breaks queries into sub-questions and decides whether to search, reflect, read, or answer. It targets the “reliable input, trusted output” niche by letting you configure each stage rather than hiding them behind a single API call.

The interesting bit The project treats reranking as a first-class research problem, shipping pointwise, pairwise, setwise, and BGE rerankers side-by-side—though listwise and TourRank are still on the waiting list. Its DeepResearch module runs a depth-first search through sub-queries with token-budget and action-depth guardrails, essentially turning the LLM into a methodical research assistant that knows when to stop.

Key highlights

• Modular pipeline with swappable retrievers (BM25, dense, hybrid), rerankers (BGE, FLAN-T5, Qwen2), and judges (BgeJudge)
• DeepResearch mode with recursive sub-query expansion, intent parsing, and action-type reasoning (search, reflect, read, code, answer)
• Document parsing for text, docx, ppt, excel, html, pdf, and md, plus integration with MinerU for extraction
• Multimodal RAG support via GLM-4V-Flash and vector database backends including Milvus and Qdrant
• Citation generation and HyDE query rewriting built in

Caveats

• ListWise-Rerank and TourRank are explicitly marked “waiting to implement”
• The changelog shows major features like Deep Search and web search landed in early 2025, so the surface is still expanding

Verdict Good fit for RAG practitioners who need fine-grained control over retrieval and ranking components and don’t mind assembling the pipeline themselves. Skip it if you want a fully managed, batteries-included platform with stable abstractions.