Code-Survey: Uncovering Insights in Large-Scale Codebases with LLM

• Do we truly understand how complex systems, like the Linux kernel, work?
• How can we grasp the high-level design choices and evolution of such intricate systems?

Code-Survey is the first step in using LLMs to gain meaningful insights into large-scale software systems. Unlike Retrieval-Augmented Generation (RAG) or fine-tuning, Code-Survey offers a unique approach.

Imagine if every entry-level kernel developer or a graduate student studying the kernel， could participate in a survey about every commit. What valuable information could we uncover from the results?

Code-Survey helps you explore and analyze some of the world's largest and most complex codebases, such as the Linux kernel. By carefully designing a survey and transforming unstructured data like commits and emails into organized, structured, and easy-to-analyze formats, you can perform quantitative analysis with ease. Code-Survey simplifies the process of uncovering valuable insights in modern complex software systems.

See our arxiv for details: https://arxiv.org/abs/2410.01837

Here is an example of analysis for the eBPF subsystem: docs/report_ebpf.md. (Not yet complete...more is adding)

Code-survey is a prototype and a on going poject now. We are working on the next step: Can we automatically mine meaningful pattern from everying we produced in the software developement? Just tell AI what you need, and ask AI to give you a compresenive report.

Why Code-Survey?

As far as we know, Code-Survey is the first method and tool that leverages LLMs to systematically explore and analyze large-scale codebases, providing quantitative results. By transforming unstructured data like git messages and emails into structured data, you can perform detailed analyses on software evolution, uncovering insights related to design, implementation, maintenance, reliability, and security.

Software development is also a social activity, allowing the application of established social science techniques. By treating LLMs as human participants in a survey, Code-Survey bridges the gap between high-level design decisions and the intricate details of implementation. This enables the analysis of large-scale codebases like the Linux kernel in ways that were previously impossible.

Why Choose Code-Survey?

• Unprecedented Analysis: No human can analyze large-scale codebases as thoroughly as AI can.
• No Chatbots or RAG: Forget about chatbots, RAG document searches, or code generation. Stop the ineffective AI!
• Simple and Flexible: Use only data like git messages and email data. Design a survey and run it with just a few lines of Python code. Easily apply Code-Survey to other projects or subsystems by designing your own code-survey.

Let's do Code-Survey!

Join us in revolutionizing the way we understand and analyze complex software systems.

What Questions Can Code-Survey Help Answer?

• How do new feature introductions impact the stability and performance of existing components over time?
• What are the identifiable phases in a component's lifecycle (e.g., development, stabilization, optimization)?
• How have specific features evolved over successive commits?
• Which components or files have the highest bug frequency?
• What dependencies have emerged between features and components?
• How can lessons from the development history of one subsystem improve others?

Workflow / Methodolog

The core idea of Code-survey is to treat LLMs like human participants in a survey:

• they can process data faster and cheaper, but are also prone to errors and limitations.
• By applying traditional human survey methods, we can efficiently conduct LLM-based surveys, while human experts provide oversight and validation to ensure accuracy.
• You can let LLM help you with survey design and data

Our approach follows a well-defined workflow:

1. Human Experts or LLM Agents design surveys: Tailored questions for each unstructured data type (e.g., commits, emails) to extract key information.
2. LLM Agents complete the surveys: Answering yes/no, tagging relevant data, and summarizing key information. This is the key steps to turn unstructured data into structured data.
3. Human Experts or LLM Agents analysis results: Ensuring accuracy and uncover new insights easily from that. If the results are not statisfied, go back to step 1 to enhance the survey.
4. Generate Reports: Summarize the results and provide insights into the data.

Survey Design with LLM Assistance

A key aspect of Code-Survey is designing effective surveys to generate accurate data. Surveys can be designed by humans or LLM agents. We identify three key steps to guide LLM agents in designing surveys. The following prompts serve as a framework or LLM input for survey creation:

• Design High-Level Insightful Questions: If you could ask every kernel developer to complete a survey or questionnaire about a commit or an email, what are the most insightful questions related to design, implementation, maintenance, reliability, and security? Describe the possible questions in detail.
• Identify Required Data Types and Sources: What data types and sources are required to answer the insightful questions described previously? Describe the data types and sources for each question in detail.
• Design Survey Questions to Retrieve Data: What survey questions can you design to obtain the required data types for the insightful questions from the data sources described previously? Describe the survey questions for kernel developers in detail.

This workflow ensures that the LLM-driven survey design process leads to structured data that offers deeper insights into complex software systems, such as the Linux kernel. By guiding LLM agents through these steps, we can systematically extract valuable information from unstructured data sources.

Linux-bpf Dataset

The Linux-bpf dataset focuses on the eBPF subsystem and is continuously updated via CI. The dataset includes:

• 680+ expert-selected commits: Features, commit details, types (Map, Helper, Kfunc, Prog, etc.). Human experts tagged these commits and can be analyzed by LLM Agents. dataset here
• 12,000+ BPF-related commits: LLM Agent surveys and summaries. You can download dataset here.
• 150,000+ BPF subsystem-related emails: LLM Agent surveys and summaries(TODO).

To see more details abot what we find, check the analysis in report_ebpf.md.

A simplest approach to see how these data works is just Upload the CSV to ChatGPT(**Or other platforms) and Ask questions to let it Analysis for you!

Note this is just a very simple demo now --- there are hundreds of ways to improve the survey accuracy:

• It's using gpt4o API, o1 model can be much better;
• You can simply run it multiple times to get multiple survey results and then average them. This is typically a real survey d and the result would be much better, but we need time and more money for API.
• More Advance Agent design with multi-steps and reasonging, or multi-agent;
• Better prompt engineering;

Survey Example

You can find this example in survey/commit_survey.yml, which analysis all the 10000+ bpf commits in the Linux kernel eBPF subsystem.

title: "Commit Classification Survey"
description: "A survey about the commit in Linux eBPF, to help better understand the design and evolution of bpf subsystem. For choice, try to be as specific as possible based on the commit message and code changes. If the commit message is not clear or does not provide enough information, you can choose the 'I'm not sure' option."
hint: "For example, when seems not related to eBPF, confirm it's a rare cases really has nothing to do with eBPF in all it's contents, such as btrfs or misspelled commit message. Do not tag subsystem changes related to eBPF as not."
questions:
- id: summary
  type: fill_in
  question: "Please provide a summary of It in one short sentence not longer than 30 words. Only output one sentence."
  required: true

- id: keywords
  type: fill_in
  question: "Please extract no more than 3 keywords from the commit. Only output 3 keywords without any special characters."
  required: true

- id: commit_classification
  type: single_choice
  question: "What may be the main type of the commit?"
  choices:
    - value: A bug fix. It primarily resolves a bug or issue in the code.
    - value: A new feature. It adds a new capability or feature that was not previously present.
    - value: A performance optimization. It improves the performance of existing code such as reducing latency or improving throughput.
    - value: A cleanup or refactoring in the code. It involves changes to improve code readability maintainability or structure without changing its functionality.
    - value: A documentation change or typo fix. It only involves changes to documentation files or fixes a typographical error.
    - value: A test case or test infrastructure change. It adds or modifies test cases test scripts or testing infrastructure.
    - value: A build system or CI/CD change. It affects the build process continuous integration or deployment pipelines.
    - value: A security fix. It resolves a security vulnerability or strengthens security measures.
    - value: It's like a merge commit. It merges changes from another branch or repository.
    - value: It's other type of commit. It does not fit into any of the categories listed above.
    - value: I'm not sure about the type of the commit. The nature of It is unclear or uncertain.

- id: major_related_implementation_component
  type: single_choice
  question: "What major implementation component is modified by the commit? It's typically where the code changes happened."
  choices:
    - value: The eBPF verifier. This component ensures that eBPF programs are safe to run within the kernel.
    - value: The eBPF JIT compiler for different architectures. It changes how eBPF bytecode is translated into machine code for different hardware architectures.
    - value: The helper and kfuncs. It modifies or adds helpers and kernel functions that eBPF programs can call.
    - value: The syscall interface. It changes the system calls through which user-space programs interact with eBPF.
    - value: The eBPF maps. It changes how data structures shared between user-space and kernel-space (maps) are created or managed.
    - value: The libbpf library. It affects the library that simplifies interaction with eBPF from user-space applications.
    - value: The bpftool utility. It modifies the bpftool utility used for introspecting and interacting with eBPF programs and maps.
    - value: The test cases and makefiles. It adds or modifies test cases or makefile scripts used for testing or building eBPF programs.
    - value: The implementation happens in other subsystem and is related to eBPF events. e.g. probes perf events tracepoints network scheduler HID LSM etc. Note it's still related to how eBPF programs interact with these events. 
    - value: It's like a merge commit. It includes significant changes across multiple components of the system.
    - value: It's not related to any above. It affects an implementation component not listed but does related to the BPF subsystem.
    - value: It's not related to any above. It affects an implementation component is totally unrelated to the BPF subsystem.  It's not related to any above because it totally not related to the BPF subsystem. It's a rare case wrong data and need removed.
    - value: I'm not sure about the implementation component of the commit. The component affected by It is unclear.
......

run the survey example

There are two examples to run the survey:

• survey/survey_agent.py is a simple script to run the survey/feature_survey.yml survey, which works on the feature_commit_details.csv dataset. It try to answer questions one at a time.
• survey/survey_struct.py is a simple script to run the survey/commit_survey.yml survey, which works on the commit_survey.csv dataset. It's using struct output and try to answer all questions at once. It will also revise the survey result to make it more accurate.

They are just simple scripts, you can set the OPENAI_API_KEY in the .env file to run it.

Why LLM?

LLMs have been proven effective in survey, summarization, and analysis tasks in fields like market research and chemistry. With LLMs, we can analyze unstructured data, which traditional methods struggle to handle efficiently.

Why Not Other Methods? Why human experts can't?

• Domain Knowledge: Required for Linux kernel analysis Domain Knowledge. Just use keyword search or simple NLP tools can't handle that.
• Unstructured Data: Commit messages and emails are difficult to process with traditional tools.
• Expert Cost: Manually analyzing this data is time-consuming and expensive. It's impossible to do that for the whole Linux kernel commits. Here’s a focused version emphasizing Designing Questions and Using Statistical Methods for ensuring correctness and consistency:

How to Make Sure the Result is Correct and Meaningful?

• Design Questions for Correctness: Create well-structured, specific questions that minimize ambiguity. Design questions to compare similar features, commits, or components across different contexts, which helps identify inconsistencies and errors in the results.
• Cross-Check with Known Data: Ensure that the questions target areas where the answers can be validated with reliable sources like commit logs, documentation, or expert k