Bridging Semantic Gaps in RAG through Generated Context Knowledge Fusion
Authors: Xinkai Du, Chao Lv, Yalin Sun, Quanjie Han, Lei Yao, Maosong Sun
Organizations: Beijing Wanlian Zhilian Technology Corporation Limited, Beijing, China · Department of Computer Science and Technology, Tsinghua University, Beijing, China · Sunshine Digital Intelligence Tech Co., Ltd., Beijing, China
Retrieval-Augmented Generation has established itself as a fundamental framework in natural language processing, seamlessly integrating information retrieval with the generative capabilities of large language models. However, this process is fundamentally constrained by a critical challenge: semantic space mismatch between queries and retrieved contexts. We propose Knowledge-Aware Semantic Bridging (KASB), a novel framework that improves passage selection quality through semantic space alignment between queries and retrieved documents through intelligent knowledge fusion. Our approach leverages the complementary strengths of generative and retrieval-based knowledge through a multistage process that enhances both relevance and accuracy. We evaluate KASB on three popular open-domain Question Answering datasets to demonstrate the effectiveness of our approach.
Figures & tables
Figure 1: Overview of the KASB Framework. A finetuned generator is derived via the DPO algorithm on the training set to produce a more effective generative context generator, which then serves in the test phase. During inference, the finetuned generator first generates query-aligned generative contexts G ; meanwhile, the evidence (representing the complete retrieved knowledge base) is processed through the generated context-based selector to identify relevant retrieved contexts C , which are then fused with the generated contexts [C;G] for final answer generation.
Datasets
Train
Dev
TriviaQA
75675
8750
NQ
91334
10039
WebQ
3906
-
Table 1: Datasets statistics.
Methods
TriviaQA
NQ
WebQ
Single Knowledge
Retri-Only
62.2
48.6
48.3
Gen-Only
67.5
50.3
41.6
Two Knowledge
HyDE
72.3
53.4
54.6
COMBO
74.6
54.2
53.0
Table 2: Exact match scores on test dataset.
Figure 2: Average retrieval exact match score obtained by finetuned vs original LLM generator on generated contexts.
Model
TriviaQA
NQ
WebQ
KASB
75.4
59.6
61.2
w/o GK
69.9
54.9
56.3
w/o RK
73.2
56.8
59.5
w/o DPO
74.6
57.7
59.8
Table 4: Question answering performance (Exact Match). GK and RK denote generated knowledge and retrieved knowledge, respectively. The combined knowledge results are computed by union of single knowledge sources.
Selection
TriviaQA
NQ
WebQ
Avg.
Fixed- k , k=5
73.2
58.2
59.2
63.5
Fixed- k , k=10
73.8
58.5
59.8
64.0
Fixed- k , k=20
74.2
58.9
59.7
64.3
Fixed- k , k=50
73.5
58.1
59.4
63.7
Adaptive k⋆
75.4
59.6
61.2
65.4
Table 5: Fixed- k versus adaptive evidence selection. EM is reported on the three QA datasets; Avg. is the mean across datasets.
Figure 3: Parameter sensitivity analysis showing the effect of DPO coefficient β and initial candidate pool size on EM performance across datasets.