Sem: Semantic Code Understanding via Git Entities

Sem shifts code analysis from line-based diffs to entity-based understanding

Sem is a new primitive for code understanding that layers semantic analysis on top of Git. Unlike traditional Language Server Protocols (LSPs), Sem identifies code as a collection of entities—such as functions, classes, and methods—rather than a series of line changes. This approach allows developers and AI agents to track what actually changed in the logic of a program, regardless of how those changes are represented in a text-based diff.

Core Capabilities and Command Set

Sem provides six primary commands to analyze a Git repository without requiring configuration files or plugins, supporting 26 languages and 5 data formats.

Entity-Level Diffing and History

  • sem diff: Replaces standard line diffs with entity-level changes. It detects renames and uses structural hashing to identify if a function was added, modified, or deleted, providing word-level inline highlights.
  • sem blame: Provides per-entity blame, identifying the last commit that modified a specific function, class, or method.
  • sem log: Tracks the evolution of a single entity across the Git history, showing every commit that touched a specific function.

Impact Analysis and Context

  • sem impact: Generates a cross-file dependency graph. It identifies every entity that depends on a given function and lists transitively affected entities and affected tests.
  • sem entities: Lists every function, class, method, and type within a given path, including their exact line ranges.
  • sem context: Creates a token-budgeted context window for LLMs. It bundles the target entity, its dependencies, and its dependents into a prompt-ready format to fit within LLM token limits.

Performance and Compatibility

Sem is designed for speed and broad compatibility, featuring a typical diff time of 8ms. It supports a wide array of languages including TypeScript, JavaScript, Python, Go, Rust, Java, C, C++, C#, Ruby, PHP, Swift, Kotlin, Elixir, Bash, HCL, Fortran, Vue, Svelte, Dart, Perl, OCaml, Scala, and Zig, as well as data formats like JSON, YAML, TOML, CSV, and Markdown.

AI Agent Integration and Benchmarks

According to Ataraxy Labs, AI agents are 2.3x more accurate when provided with Sem's entity-level output compared to raw line diffs. While the tool provides a --json flag for machine-readable output to facilitate this integration, some community members have questioned the specificity of these benchmarks, suggesting they may be tailored to Sem's own concepts of "entities" rather than general coding tasks.

Community Feedback and Implementation Concerns

While the technical utility of the tool is praised, the community has raised several concerns regarding its installation and integration:

  • Git Integration: The sem setup command configures Git to use Sem as the external diff tool (git config --global diff.external = sem) and installs a pre-commit hook. Some users have described this as "user-hostile," noting that it overrides the default git diff behavior without a clear, non-destructive way to use the CLI as a complementary tool rather than a replacement.
  • Monorepo Scaling: Questions have been raised regarding how the tool performs in massive monorepos where standard Git operations already suffer from performance degradation.
  • Application in Code Review: Some users suggest that an AST-based diff system like Sem could be integrated into code forge UIs to replace line diffs as the primary view for reviewers.

"I have been using all of ataraxy labs tools for the past few months and that have been indispensable for models to make less mistakes for me. From better git diffs to impact analysis and code reviews."

"I doubt if this actually solves a real problem for humans or agents, especially in complex projects."

Installation

Sem can be installed via Homebrew or Cargo:

brew install sem-cli
# OR
cargo install --git https://github.com/Ataraxy-Labs/sem sem-cli

To revert the Git configuration changes made by sem setup, users can run sem unsetup.

Sources