ARKS: Active Retrieval in Knowledge Soup for Code Generation

Hongjin Su1, Shuyang Jiang2, Yuhang Lai2,
Haoyuan Wu1, Boao Shi1, Che Liu1, Qian Liu3, Tao Yu1
1The University of Hong Kong, 2Fudan University, 3Sea AI Lab
Paper Code Twitter Data

We introduce ARKS, a general pipeline for retrieval-augmented code generation (RACG). We construct a knowledge soup integrating web search, documentation, execution feedback, and evolved code snippets. Through active retrieval in knowledge soup, we demonstrate significant increase in benchmarks about updated libraries and long-tail programming languages (8.6% to 34.6% in ChatGPT)

Abstract

Recently the retrieval-augmented generation (RAG) paradigm has raised much attention for its potential in incorporating external knowledge into large language models (LLMs) without further training. While widely explored in natural language applications, its utilization in code generation remains under-explored. In this paper, we introduce Active Retrieval in Knowledge Soup (ARKS), an advanced strategy for generalizing large language models for code. In contrast to relying on a single source, we construct a knowledge soup integrating web search, documentation, execution feedback, and evolved code snippets. We employ an active retrieval strategy that iteratively refines the query and updates the knowledge soup. To assess the performance of ARKS, we compile a new benchmark comprising realistic coding problems associated with frequently updated libraries and long-tail programming languages. Experimental results on ChatGPT and CodeLlama demonstrate a substantial improvement in the average execution accuracy of ARKS on LLMs. The analysis confirms the effectiveness of our proposed knowledge soup and active retrieval strategies, offering rich insights into the construction of effective RACG pipelines.

Arks is effective because:
  • Knowledge soup:   Diverse knowledge soup consists of not only documentation, but also more resources including web search, code snippets and execution feedback
  • Query Formulation:   Apart from natural language question, Arks employ the explained code as more informative query, which helps retrieval models match desired knowledge source
  • Active Retrieval:   Compared to RAG, we actively refine query and knowledge soup, which helps to retrieve more relevant information
  • Performance:   Across all datasets, ChatGPT and CodeLlama achieve performance gain of 24.0% and 23.8% respectively, compared to the vanilla setting without knowledge augmentation

Main Results

Diverse resources in general help LLM generalization.

ChatGPT and CodeLlama execution accuracy with different knowledge sources. Tensor-M refers to Tensorflow-M and Avg. refers to the average score across four benchmarks. Web denotes the web search content; Exec denotes the execution feedback from compiler/interpreter; Code denotes the code snippets generated by LLMs in previous rounds that are verified to be free of syntax error; Doc refers to the documentation. Adding more knowledge sources consistently enhances the performance, which demonstrates the advantage of a diverse knowledge soup in the RACG pipeline.

Analysis

We demonstrate the significant performance increase provided by active retrieval on top of one-time retrieval

Next, we show the critical role of both query formulation and retrieval model choice

Furthermore, we evaluate the retrieval accuracy, revealing that both retrieval and generator models have large rooms to improve for better generalization

With long-context models and more knowledge included in the prompt, model generalization performance is not guaranteed to improve, calling for more delicate mechanism in RACG pipeline

Last but not least, we perform a case study o provide a more intuitive understanding of the improvement achieved by ARKS on top of LLMs,

Dataset Statistics

To evaluate Arks, we curate 4 datasets covering two realistic scenarios of RACG: updated libraries and long-tail programming languages. Here are the data statistics: