How AI Can Turn Codebases into Knowledge Graphs: A Practical Guide

Origin

Li Ming, a new R&D engineer, quickly encounters undocumented code, obscure business rules, and missing requirement documents, leading to production incidents and confusion among teammates.

Breakthrough

Realizing the need for a systematic solution, Li Ming decides to leverage large language models (LLMs) to connect scattered knowledge across code, requirements, and operations.

Stage 1: Basic Application

He creates a simple script that indexes company documentation and code commit records, allowing keyword search to retrieve related documents. The prototype demonstrates that an LLM can answer historical feature questions more efficiently than manual searching.

Stage 2: Knowledge Integration

Li Ming defines three key capabilities:

Basic query – provide standard answers for common business issues.

Knowledge association – link code changes with requirement documents and incident records.

Intelligent suggestions – automatically surface relevant historical experience when new requirements are developed.

He builds a knowledge base containing essential artifacts such as TRD, ERD, system design docs, and common Q/A, and integrates them with the LLM.

Stage 3: Deep Application

Advanced features include:

Code change traceability – retrieve historical modifications for any code segment.

Requirement analysis – help newcomers understand system evolution.

Development assistance – generate basic code snippets from requirement descriptions.

Experience transfer – suggest implementation ideas based on similar past cases.

Technical Implementation

Key steps involve binding requirements to code, cleaning and annotating data, and uploading it to a dataset platform (e.g., DIFY). Example API calls:

curl -H "Authorization: token YOUR_TOKEN" "https://api.github.com/repos/{owner}/{repo}/issues/{issue_number}"

curl -H "Authorization: token YOUR_TOKEN" "https://api.github.com/repos/{owner}/{repo}/commits/{commit_sha}"

curl -H "Authorization: token YOUR_TOKEN" "https://api.github.com/search/commits?q=repo:{owner}/{repo}+[REQ-123]+in:message"

Dataset creation and segment upload are performed with similar curl commands, attaching code‑requirement pairs and optional metadata such as programming language.

Future Optimization

Identified improvement areas include enhancing code generation quality for stable modules, improving the accuracy of knowledge association, and refining retrieval‑augmented generation (RAG) techniques to better match queries with relevant documents.

Conclusion

The multi‑stage approach transforms fragmented knowledge into a systematic, AI‑enhanced knowledge base, reducing onboarding difficulty, preserving institutional memory, and enabling sustainable optimization of software development processes.