Traceable Cross-Source RAG for Chinese Tibetan Medicine Question Answering

Fengxian Chen, Zhilong Tao, Jiaxuan Li, Yunlong Li, Qingguo Zhou

TL;DR

Researchers have developed a new Retrieval-Augmented Generation (RAG) framework specifically for Chinese Tibetan Medicine that solves the "encyclopedia bias" problem. By using a specialized routing algorithm (DAKS) and an entity-alignment graph, the system ensures that AI-generated medical answers are grounded in authoritative classical texts and clinical papers rather than just easy-to-match summary snippets.

What Problem Does This Paper Address?

In specialized fields like Traditional Chinese Medicine (TCM) or Tibetan Medicine (TM), knowledge is scattered across very different types of sources: short, information-dense encyclopedias; ancient, complex classical canons; and modern clinical research papers.

Standard RAG systems usually dump all these sources into one "flat" database. This creates a density-driven bias: because encyclopedia entries are short and keyword-rich, retrieval algorithms almost always prefer them. However, for a serious medical query, a classical principle or a clinical trial report is often a more authoritative source. Furthermore, existing systems struggle with cross-source verification—the ability to link a traditional theory from a 12th-century text to a modern clinical outcome.

Methods

The authors propose two primary innovations to create a "Traceable Cross-Source RAG":

  1. DAKS (Routing with Budgeted Retrieval): Instead of searching one giant pool, DAKS performs a "probe retrieval" across each knowledge base (KB) independently. It analyzes the distribution of scores (using metrics like entropy and "peakiness") to determine a retrieval "budget" for each source. This prevents the encyclopedia from dominating the results and ensures that authoritative sources (Classics and Papers) are represented in the final context.
  2. Alignment Graph-Guided Fusion: The team built a bipartite graph where nodes represent either text chunks or specific medical entities (e.g., Disease, Symptom, Drug, Formula). This graph acts as a bridge. If a query mentions a specific herb, the graph can find a "path" from a classical description of that herb to a modern paper discussing its clinical efficacy, even if the two texts use different terminology.
  3. Coverage-Aware Packing: Most RAG systems just concatenate the top $K$ results. This system uses a greedy algorithm to ensure the final prompt sent to the LLM contains at least one high-quality chunk from every required source category, maximizing "cross-source evidence coverage."

The generator used is openPangu-Embedded-7B, a lightweight, deployment-friendly model.

Datasets

The researchers introduced a new Chinese Tibetan-medicine QA dataset consisting of 500 queries. The dataset is balanced across four categories:

  • Definitions: Basic terminology.
  • Classics Principles: Theoretical foundations.
  • Clinical Evidence: Modern medical applications.
  • Cross-KB Synthesis: Questions that explicitly require combining info from multiple sources (e.g., "How does the classical treatment for X compare to modern clinical findings?").

Key Insights and Ideas

  • The "Authority Gap": The paper highlights that in high-stakes domains, "relevance" (how well keywords match) is not the same as "authority." DAKS treats retrieval as a resource allocation problem rather than a simple similarity search.
  • Graph-as-Bridge: By using entities as anchors, the system overcomes the "semantic gap" between ancient classical language and modern clinical terminology. The graph allows the system to perform "bridge retrieval," pulling in relevant context that a standard vector search would miss.
  • Routing + Fusion = Success: A key finding was that Graph Fusion works best when the candidate pool is already well-balanced by DAKS. If the initial retrieval is poor, the graph cannot save the final answer.

Key Results

The full system (DAKS + GraphFusion) significantly outperformed standard baselines:

  • Cross-Source Evidence Coverage (CrossEv@5): Achieved a score of 0.780, compared to just 0.620 for naive multi-KB concatenation and 0.650 for a merged KB approach.
  • Citation Correctness: Maintained a high score of 0.760, ensuring that when the AI cites a source, the evidence actually supports the claim.
  • Reduced Dominance: DAKS successfully reduced the "Encyclopedia Dominance Rate" (EDR) from 48.5% in merged retrieval to 36.2%, allowing more authoritative texts to reach the LLM.

Limitations and Future Work

  • Judge Bias: All metrics were evaluated using a single "judge" model (GLM-4.7). While this ensures consistency, it may introduce systematic biases inherent to that specific LLM.
  • Entity Extraction Dependency: The alignment graph's effectiveness relies heavily on the quality of the initial entity extraction and chunk metadata. Errors in the "knowledge graph" layer can propagate to the final answer.
  • Domain Specificity: The current work is focused strictly on Tibetan Medicine. The authors suggest that future work should validate if these "budgeted retrieval" strategies work as well in other medical or legal subdomains.

Summary updated

Abstract

Retrieval-augmented generation (RAG) promises grounded question answering, yet domain settings with multiple heterogeneous knowledge bases (KBs) remain challenging. In Chinese Tibetan medicine, encyclopedia entries are often dense and easy to match, which can dominate retrieval even when classics or clinical papers provide more authoritative evidence. We study a practical setting with three KBs (encyclopedia, classics, and clinical papers) and a 500-query benchmark (cutoff $K{=}5$) covering both single-KB and cross-KB questions. We propose two complementary methods to improve traceability, reduce hallucinations, and enable cross-KB verification. First, DAKS performs KB routing and budgeted retrieval to mitigate density-driven bias and to prioritize authoritative sources when appropriate. Second, we use an alignment graph to guide evidence fusion and coverage-aware packing, improving cross-KB evidence coverage without relying on naive concatenation. All answers are generated by a lightweight generator, \textsc{openPangu-Embedded-7B}. Experiments show consistent gains in routing quality and cross-KB evidence coverage, with the full system achieving the best CrossEv@5 while maintaining strong faithfulness and citation correctness.
BibTeX
@misc{chen2026traceable,
  author = {Fengxian Chen and Zhilong Tao and Jiaxuan Li and Yunlong Li and Qingguo Zhou},
  title = {{Traceable Cross-Source RAG for Chinese Tibetan Medicine Question Answering}},
  year = {2026},
  eprint = {2602.05195},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2602.05195},
}
Figure