Enhancing Knowledge Density via Hierarchical Context Pruning for Retrieval-Augmented Repository-Level Code Completion

Xiaohan Chen; Xiang Chen; Yu Tian; Piji Li; Sheng-Jun Huang; Songcan Chen

cs.SEAug 13, 2025

Enhancing Knowledge Density via Hierarchical Context Pruning for Retrieval-Augmented Repository-Level Code Completion

Authors: Xiaohan Chen, Xiang Chen, Yu Tian, Piji Li, Sheng-Jun Huang, Songcan Chen

Organizations: MIIT Key Laboratory of Pattern Analysis and Machine Intelligence, College of Computer Science and Technology, Nanjing University of Aeronautics and Astronautics · Tsinghua University

Abstract

Retrieval-Augmented Generation (RAG) has advanced repository-level code completion by retrieving external code snippets, API documentation, or type definitions to expand the model's contextual awareness. However, under the finite token budget of the context window, existing retrieval-augmented methods often saturate the budget with highly similar or even redundant code, which not only wastes tokens but also suppresses critical cross-file dependencies, preventing synergistic interaction between semantically similar reference snippets and cross-file dependency definitions. Therefore, further gains are limited by low knowledge density in the retrieved context rather than by model capacity. We propose SaraCoder, a retrieval-augmented framework that enhances knowledge density via hierarchical context pruning under a fixed context budget. A Hierarchical Feature Optimization module prunes an over-complete candidate pool in four stages. The freed tokens are then allocated to an External-Aware Identifier Disambiguator, which selectively injects only the most necessary cross-file dependency definitions. SaraCoder transforms context construction into a budget-aware optimization problem, balancing reference snippets and cross-file dependencies to build a context that is both knowledge-dense and semantically complete. Experiments on CrossCodeEval and RepoEval-Updated, across multiple backbone code LLMs, show that SaraCoder consistently outperforms strong RAG baselines such as GraphCoder and Repocoder in both code-level and identifier-level metrics, without increasing token usage, and yields synergistic gains when combined with other cross-file methods.

Figures & tables

Fig. 1: The pitfalls of pure similarity retrieval vs. the efficiency of Saracoder . Pink boxes highlight how traditional retrieval leads to “Crowding out” the context with redundant semantics or ambiguous noise. In contrast, green boxes showcase the “Compact & Inject” strategy of Saracoder , which distills essential knowledge and injects precise, high-value insights.

Fig. 2: An illustration of Saracoder framework. (1) Database Construction. This phase constructs a key-value codebase. This involves using a slicing algorithm to create induced graph slices, which are then precisely mapped to source code snippets. (2) Code Retrieval. This phase takes code context as input and retrieves similar code, then refines suggestions via Hierarchical Feature Optimization. Concurrently, an External-Aware Identifier Disambiguator clarifies external symbols via dependency analysis, delivering highly accurate candidates. (3) Code Generation. This phase generates prompts by integrating outputs from code retrieval with the code completion context. These prompts are then fed into an LLM to predict completion statements.

Fig. 3: Prompt template. (

\mathcal{S}_{ref}

corresponds to similar code snippet,

\mathcal{D}_{ext}

corresponds to external symbols enhancement, and

C_{\text{context}}

corresponds to unfinished code context.)

	CrossCodeEval		RepoEval-Updated
	Python	Java	Python	Java
Total Repositories	471	239	10	8
Total Files	14348	5868	3258	8260
Total Task cases	2665	2139	2000	1600

TABLE I: CrossCodeEval vs. RepoEval-Updated.

Relationship	Definition and Type-Sensitive Characteristics	Syntax Examples (Python / Java)
Assignment	A variable acquires its type identity through assignment, which enables type inference and propagation.	Python: count: int = 10 ; Java: String s = new String() .
Contextual Binding	Temporary type bindings are created within specific syntactic structures, and their lifecycle is managed depending on the context.	Python: with open(file) as f: ; Java: try (BufferedReader br = ...) .
Reference	Existing variables or properties are accessed, which requires late-bound type resolution.	Python: obj.calculate() ; Java: this.value .
Type Declaration	Variable or return types are explicitly annotated, which provides the basis for static type checking.	Python: def func() -> list[str]: ; Java: List<Integer> list = new ArrayList<>() .
Parameter Constraint	Function parameters are constrained by types, which ensures input type validation.	Python: def sort(items: Sequence[T]) ; Java: void sort(List<? extends Comparable> l) .
Return Constraint	Function return values are constrained by types, which guarantees output type consistency.	Python: @return_type(float) ; Java: public int getValue() { ... } .

TABLE II: Key Associations for Analysis, Refactoring, and Debugging Modeled in Context Graph Construction

Dataset	Language	Methods	Models and Metrics
			Codegen2-7b				Codegen25-7b				deepseek-coder-6.7b-instruct				CodeLlama-7b-Instruct
			Code Match		Identifier Match		Code Match		Identifier Match		Code Match		Identifier Match		Code Match		Identifier Match
			EM	ES	EM	F1	EM	ES	EM	F1	EM	ES	EM	F1	EM	ES	EM	F1
RepoEval-Updated	Python	No Rag	17.40	32.54	23.75	30.21	19.55	34.48	25.75	32.16	11.50	30.33	15.30	22.39	17.35	33.05	23.55	30.53
		Shifted Rag	32.70	59.22	40.10	55.66	36.45	61.96	43.20	58.31	20.90	42.95	26.50	38.88	33.90	60.28	41.50	56.39
		Vanilla Rag	38.70	63.58	46.45	60.43	42.25	66.26	48.75	62.79	22.20	41.48	27.85	37.58	40.30	65.03	47.55	61.06

TABLE III: Performance comparison on RepoEval-Updated and CrossCodeEval datasets. Numbers are shown in percentage (%). The top results are bolded , and the second best are underlined .

Fig. 4: Impact of top_k on RepoEval-Updated and CrossCodeEval.

Language	Methods	Codegen2-7b				Codegen25-7b				CodeLlama
		Code Match		Identifier Match		Code Match		Identifier Match		Code Match		Identifier Match
		EM	ES	EM	F1	EM	ES	EM	F1	EM	ES	EM	F1
Python	No Rag	5.44	57.85	11.71	42.22	7.77	60.52	14.45	45.40	9.49	61.97	16.44	47.36
	Shift Rag	4.87	58.36	11.64	42.91	7.44	60.20	14.17	44.78	6.95	60.35	13.75	45.36
	Vanilla Rag	9.52	61.87	17.42	48.01	12.43	63.81	20.74	51.00	11.48	63.66	19.42	50.72
	Saracoder	12.16	54.16	18.37	45.47	11.50	44.90	16.22	38.95	13.32	48.04	19.54	42.34

TABLE IV: Performance benefits of Saracoder when integrated orthogonally with other cross-file approaches (

\%

). The CrossCodeEval dataset is used in this part.

Method	Codegen2-7b		deepseek-coder-6.7b-instruct		CodeLlama-7b-Instruct
Method	#In	#Out	#In	#Out	#In	#Out
No RAG	1397.38	50.44	1383.12	95.22	1388.38	50.146
Shifted RAG	821.627	95.15	778.52	93.48	763.93	96.86
HF_OP (top_k=10)	1665.50	88.15	1621.92	90.86	1611.37	96.99
HF_OP (top_k=4)	1644.26	89.20	1596.83	88.57	1590.55	98.06

TABLE V: Quantitative analysis of top_k and token when max_num_tokens is 2048.

Dataset	Language	Method	hit@1	hit@5	hit@10
RepoEval_Updated	python	No Comp.	25.40%	32.15%	34.35%
	python	HF_OP	25.00%	34.50%	36.45%
	java	No Comp.	12.69%	19.81%	22.25%
	java	HF_OP	15.12%	21.94%	23.19%
CrossCodeEval	python	No Comp.	0.11%	0.26%	0.38%
	python	HF_OP	2.61%	3.97%	4.16%

TABLE VI: Hit@k Performance on RepoEval_Updated and CrossCodeEval Datasets. “No Comp.” denotes the original retriever without any compression processing.

Fig. 5: Ablation study. (Each three-data-point group represents CodeGen2-7B, CodeGen2.5-7B, and CodeLlama-7B-Instruct models. Bar lengths show their average performance, with I-shaped error bars indicating standard deviation)

LLM Code Generation

Large language model code generation studies how models translate instructions and context into executable programs and how to make them reliable.

See all 309 papers

Effective and Efficient Context Retrieval via Partial Dependency Graph for Repository-Level Code Generation
Aug 3, 2026Zhongxin Liu, Zhonghao Jiang, Zhifan Ye +3Retrieval-Augmented Generation Graph-Based Retrieval
College of Computer Science and Technology and the State Key Laboratory of Blockchain and Data Security, Zhejiang University · School of Computer and Computing Science, Hangzhou City University · Faculty of Computing, Harbin Institute of Technology +1
When Retrieval Hurts Code Completion: A Diagnostic Study of Stale Repository Context
May 14, 2026Haojun Weng, Qianqian Yang, Hao Fu +2Retrieval-Augmented Generation Retrieval Robustness
Independent Researcher, California, USA · Independent Researcher, Beijing, China
Code Comprehensor: Hierarchical Documentation and Cross-Modal Retrieval for Large Repositories
Oct 21, 2026Sujoy Roychowdhury, Joseph H. R. Isaac, Jiangning Gao +2Cross-Modal Retrieval Semantic Search
Ericsson R&D

Aug 3, 2026J/K move · Enter open · S save

Effective and Efficient Context Retrieval via Partial Dependency Graph for Repository-Level Code Generation

Zhongxin Liu, Zhonghao Jiang, Zhifan Ye, Haoye Wang+2

College of Computer Science and Technology and the State Key Laboratory of Blockchain and Data Security, Zhejiang University · School of Computer and Computing Science, Hangzhou City University · Faculty of Computing, Harbin Institute of Technology+1

Retrieval-Augmented Generation Graph-Based Retrieval Repository-Level Code Generation

Open paper AskPDF arXiv

LLM-based repository-level code generation aims to generate code using the context available in a software repository, requiring LLMs to reason over complex code dependencies. Due to limited context windows and insufficient repository-specific understanding, LLMs typically rely on retrieval-augmented generation (RAG) to incorporate relevant code. Early RAG approaches primarily employ similarity-based retrieval, which often fails to retrieve code snippets that the target function depends on. Recent work introduces graph-based retrieval to model such dependencies, but typically relies on manually designed rules and static global graphs, leading to limited flexibility and high construction and maintenance costs. In contrast, human developers collect helpful context by implicitly constructing a partial dependency graph and iteratively inspecting along it. Inspired by this behavior, we propose DyRetriever, an efficient context retrieval method via partial dependency graphs. DyRetriever uses an LLM to first select a set of entry-point functions and then perform multi-hop reasoning along the code dependency graph. During multi-hop reasoning, it uses the LLM's semantic understanding to validate whether a function can help generate the target function, eliminating manually designed rules and enabling flexibility across scenarios. Instead of statically constructing a global dependency graph, DyRetriever builds a partial graph on demand and discards it after use, reducing construction and maintenance costs. We integrate DyRetriever with a similarity-based code retriever to build DyCoder and evaluate it on CoderEval and DevEval. Experimental results show that DyCoder achieves relative Pass@1 improvements of 25.63% and 59.73% on CoderEval and DevEval, respectively, compared with existing RAG-based methods, while being 7.4x faster than baselines based on static dependency graph construction.

Abstract

LLM Code Generation

Explore similar work

Effective and Efficient Context Retrieval via Partial Dependency Graph for Repository-Level Code Generation

When Retrieval Hurts Code Completion: A Diagnostic Study of Stale Repository Context

Code Comprehensor: Hierarchical Documentation and Cross-Modal Retrieval for Large Repositories