Sia Code

Local-first codebase search with semantic understanding and multi-hop code discovery.

Benchmark Results

89.9% Recall@5 on RepoEval benchmark (1,600 queries, 8 repositories)

• +12.9 percentage points better than cAST (77.0%)
• Lexical-only search outperforms hybrid (BM25 > BM25+embeddings)
• Publication-quality results with ±1.5% confidence interval

See docs/BENCHMARK_RESULTS.md for full analysis.

Features

• 89.9% Recall@5 - State-of-the-art code search performance on RepoEval benchmark
• Lexical-First Search - BM25 + FTS5 optimized for code queries (outperforms semantic-only)
• Multi-Hop Research - Automatically discover code relationships and call graphs
• AST-Aware Chunking - Tree-sitter preserves function/class boundaries
• Project Auto-Detection - Automatic language detection and indexing strategy
• Tiered Search - Filter by project code, dependencies, or both
• 12 Languages - Python, JS/TS, Go, Rust, Java, C/C++, C#, Ruby, PHP (full AST support)
• Watch Mode - Auto-reindex on file changes with incremental updates
• Portable Index - Usearch HNSW + SQLite FTS5 in .sia-code/ directory

Installation

# From PyPI (recommended)
pip install sia-code

# Or with uv
uv tool install sia-code

# Or from source
uv tool install git+https://github.com/DxTa/sia-code.git

# Try without installing (ephemeral run)
uvx sia-code --version
uvx sia-code search "authentication logic"

# Verify installation
sia-code --version

Quick Start

# Initialize and index
sia-code init
sia-code index .

# Search
sia-code search "authentication logic"           # Hybrid search (default: BM25 + semantic)
sia-code search --regex "def.*login"             # Lexical-only search (BM25)
sia-code search --semantic-only "handle errors"  # Semantic-only search

# Multi-hop research (discover relationships)
sia-code research "how does the API handle errors?"

# Check index health
sia-code status

Commands

Command	Description
sia-code init	Initialize index in current directory
sia-code index .	Index codebase
sia-code index --update	Re-index changed files only
sia-code index --watch	Auto-reindex on file changes
sia-code search "query"	Hybrid search (BM25 + semantic)
sia-code search --regex "pattern"	Lexical-only search
sia-code search --semantic-only "query"	Semantic-only search
sia-code research "question"	Multi-hop code discovery
sia-code status	Index health and staleness metrics
sia-code compact	Remove stale chunks
sia-code memory list	List timeline/changelogs/decisions
sia-code memory changelog	Generate changelog from git
sia-code memory sync-git	Import events from git history
sia-code config show	Display configuration
sia-code interactive	Live search mode

See docs/CLI_FEATURES.md for complete command reference with all options and examples.

Configuration

Recommended: Lexical-only search (best performance, no API key needed)

sia-code init
sia-code index .
# Search uses BM25 by default (89.9% Recall@5)

Optional: Hybrid search (adds semantic embeddings):

export OPENAI_API_KEY=sk-your-key-here
sia-code config set embedding.enabled true
sia-code config set search.vector_weight 0.0  # 0.0 = lexical-only (recommended!)
sia-code index --clean

Edit config at .sia-code/config.json to:

• Set vector_weight (0.0 = lexical-only, 0.5 = hybrid, 1.0 = semantic-only)
• Change embedding model (BAAI/bge-small-en-v1.5, openai-small)
• Exclude patterns (node_modules/, __pycache__/, etc.)
• Adjust chunk sizes (max_chunk_size, min_chunk_size)

View config: sia-code config show

AI Summarization (optional, enhances git changelogs):

{
  "summarization": {
    "enabled": true,
    "model": "google/flan-t5-base",
    "max_commits": 20
  }
}

Output Formats

sia-code search "query" --format json            # JSON output
sia-code search "query" --format table           # Rich table
sia-code search "query" --format csv             # CSV for Excel
sia-code search "query" --output results.json    # Save to file

Supported Languages

Full AST Support (12): Python, JavaScript, TypeScript, JSX, TSX, Go, Rust, Java, C, C++, C#, Ruby, PHP

Recognized: Kotlin, Groovy, Swift, Bash, Vue, Svelte, and more (indexed as text)

Troubleshooting

Issue	Solution
No API key warning	Normal - searches fallback to lexical mode
Index growing large	Run sia-code compact to remove stale chunks
Slow indexing	Use sia-code index --update for incremental
Stale search results	Run sia-code index --clean to rebuild

How It Works

1. Parse - Tree-sitter generates language-agnostic AST for each file
2. Chunk - AST-aware chunking preserves function/class boundaries (max 1200 chars)
3. Index - Usearch HNSW (vectors) + SQLite FTS5 (lexical search with BM25)
4. Store - Portable .sia-code/ directory (17-25 MB per repo)
5. Search - Lexical-first (BM25) with optional hybrid fusion (RRF)

Key Innovation: Lexical-only search (BM25) outperforms hybrid (BM25+embeddings) for code queries because code contains precise identifiers that benefit from exact keyword matching.

Documentation

Architecture & Implementation

• docs/ARCHITECTURE.md - System design, data structures, and technology stack
• docs/CODE_STRUCTURE.md - Codebase organization and key classes
• docs/INDEXING.md - Indexing pipeline and AST-aware chunking
• docs/QUERYING.md - Search methods and hybrid fusion

Benchmark Results

• docs/BENCHMARK_RESULTS.md - 89.9% Recall@5 full results and analysis
• docs/BENCHMARK_METHODOLOGY.md - RepoEval benchmark setup
• docs/PERFORMANCE_ANALYSIS.md - Why sia-code outperforms cAST by +12.9 pts

Usage & Configuration

• docs/CLI_FEATURES.md - Complete CLI reference and examples
• examples/ - Test results and usage examples
• ROADMAP.md - Development progress
• KNOWN_LIMITATIONS.md - Current limitations and workarounds

License

MIT