chace

Pronounced: /tʃeɪs/ (chase)

CHamal's AutoComplete Engine

[!WARNING] CHACE is still at early development so bugs and breaking changes should be expected.

Overview

CHACE is a Rust-based engine designed for controlled AI-assisted code completion. Traditional code completion tools like GitHub copilot or agents like cursor can hallucinate easily on large codebases when the context is misleading or too much. To mitigate that, CHACE:

• targets function declerations with empty bodies at the cursor position
• Extracts the function decleration and documentation (docstrings)
• Sends only the minimal context to the LLM
• Retrive only the function implementations from the LLM for optimal token efficiency

This approach keeps the AI focused on the specific task, reduces token usage, maintains precision and efficiency and produces more predictable results.

Inspiration

The idea is heavily inspired by ThePrimeagen's new approach to AI-assisted coding. While his implementation is built with Lua and requires Opencode to work (not yet open-sourced), I took a different architectural approach: built as a standalone Rust binary that operates independently and can be integrated into any editor through plugins. This design ensures the core application to be editor-agnostic, lightweight, and easy to extend to other development environments.

Architecture

CHACE runs as a Unix socket server (/tmp/chace.sock) that accepts JSON requests containing source code and cursor position. The engine:

1. Parses the source code using Tree-sitter
2. Locates empty functions at the cursor
3. Sends function declerations to the configured LLM backend
4. Returns the generated function body with precise byte offsets

Supported LLM Backends

• Google Gemini (gemini-2.5-flash)
• Groq (gpt-oss-20b)

Language Support

Currently supports:

• Rust
• TypeScript/ TypeScript XML
• JavaScript/ JavaScript XML

Installation

Install via Cargo

cargo install chace

Configuration

Set the required environment variables:

export GEMINI_API_KEY="your-gemini-api-key"
export GROQ_API_KEY="your-groq-api-key"

Usage

Running the Server

chace

The server listens on /tmp/chace.sock and handles concurrent connections.

Request Format

Send JSON-encoded requests via the Unix socket:

{
  "source_code": "fn add(a: i32, b: i32) -> i32 {\n\n}",
  "cursor_byte": 35,
  "backend": "Gemini",
  "context_snippets": ["struct User {\n  id: u32,\n  name: String\n}"]
}

Optional Fields:

• context_snippets (array of strings): Additional code snippets to provide context for better code generation

Response Format

{
  "start_byte": 35,
  "end_byte": 36,
  "body": "    a + b",
  "usage": {
     "prompt_tokens": 300,
     "completion_tokens": 1200,
     "total_tokens": 1500
    }
  "error": null
}

Optional Fields:

• error (string or null): Error message if the request failed, null on success

IDE Integration

CHACE is designed to be integrated with IDEs via plugins. See chace.nvim for reference.

Protocol

CHACE uses a line-delimited JSON protocol via a Unix socket:

• Each request is a single JSON object terminated by a newline
• Each response is a single JSON object terminated by a newline
• Multiple requests can be sent over the same connection
• Connections are handled asynchronously

Development

Build from Source

git clone https://github.com/chamal1120/chace.git
cd chace
cargo build --release

Adding Language Support

To add support for a new language:

1. Add the Tree-sitter grammar to Cargo.toml
2. Create a new backend in src/languages/
3. Implement the LanguageStandard trait with the following methods:
4. Return FunctionInfo objects containing:
5. Update the request handler in main.rs

Adding LLM Backends

To add a new LLM provider:

1. Create a new module in src/ai/
2. Implement the LLMBackend trait
3. Add initialization in main.rs
4. Update the backend selection logic

Testing

refer tests.

License

MIT License - see LICENSE for details.