Exploring the design space of AI based code completion engines

Artificial Intelligence (AI) based code completion tools such as Github Copilot have recently gained tremendous popularity due to their ability to suggest arbitrary length snippets, improving developer productivity dramatically. However, there is little public understanding of what it takes to build such a tool. In this thesis, we explore the design space of building such a tool. We study the importance of the two key components of such a tool: the Large Language Model (LLM) that predicts the suggestions, and the system around it that feeds it the right context and filters out the bad suggestions. We start by exploring the design of Github Copilot to understand the state of the art, and describe the three key components of Copilot: Prompt Engineering, Model Invocation and Feedback loop. We then study the various factors that affect the quality of the suggestions generated by the LLM. We study both (a) the impact of the context fed to the LLM, and (b) the impact of the LLM itself. For the former, we study the impact including context from other files and code after the cursor along with different methods of context collection and amount of collected context. For the latter, we study the impact of the size of the LLM and the training procedure. Apart from factors affecting the quality of suggestions, we also study the factors affecting the latency of such code completion engines, as low latency is critical for building good code completion engines. We find that the context fed to the model makes a significant difference in the quality of generated suggestions, and good context collection can improve the quality of suggestions by 11-26% points (20-113% relative improvement) on the exact match metric for one line suggestions. Models that can exploit the context after the cursor can further improve the quality of suggestions by 6-14% points (12-16% relative improvement). Our experiments show that increasing the prompt length beyond a point does not improve suggestion quality significantly, and that 2048-4096 tokens are sufficient. We also find that the size of the LLM has a much smaller impact on the quality of suggestions than other factors such as the context fed to the model and the training procedure used. For example, we found that the SantaCoder model (1.1B parameters) provided better suggestions than the 16B CodeGen-Multi model despite being 16x smaller. This is because the SantaCoder model was trained on a larger dataset and can also exploit the suffix context, unlike the CodeGen model. Overall, we believe these findings will be useful for the community to understand the design trade-offs involved in building such code completion tools.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2023

Author dc:creator

• Thakkar, Parth

Contributors dc:contributor

• Xu, Tianyin

• Code Completion
• Artificial Intelligence
• Machine Learning
• Software Engineering

Statement dc:rights

• Copyright 2023 Parth Thakkar

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/120156

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F120156