Code-only knowledge graph with vector search. A simplified version of Graphify focused on analyzing codebases using tree-sitter AST extraction and LanceDB vector embeddings.

• This is graphify clone
• Data structure is much more simplified than Graphify
• Only ingests code
• No other formats such as video is supported
• You can use Semantic search

• Multi-language code analysis: Supports 26+ languages via tree-sitter (Python, JavaScript, TypeScript, Go, Rust, Java, C/C++, Ruby, Scala, PHP, etc.)
• Deterministic AST extraction: No LLM calls, fast and reproducible
• Community detection: Groups related code via Leiden/Louvain clustering
• Vector search: Find code by natural language queries, powered by sentence-transformers + LanceDB
• Graph traversal: Explore relationships: calls, imports, data flow

# Install dependencies in a virtual environment
uv sync --all-extras

# Or without Leiden (fallback to Louvain)
uv sync

# (Optional) Download embedding model for offline use
./download-model.sh           # Linux/macOS
powershell -File download-model.ps1  # Windows

cd code-knowledge
pip install -e ".[leiden]"

This installs:

• networkx ≥3.0 — graph algorithms
• tree-sitter ≥0.23.0 + 21 language parsers
• lancedb — vector database
• sentence-transformers — embeddings (all-MiniLM-L6-v2)
• graspologic (optional leiden extra) — Leiden clustering (Linux/macOS, Python <3.13; Windows falls back to Louvain)

Analyze a codebase and build the graph + vector store:

uv run code-knowledge update /path/to/code

Outputs:

• code-knowledge-out/graph.json — 6-field minimal schema for nodes and edges
• code-knowledge-out/vectors/ — LanceDB vector store

uv run code-knowledge query "what handles authentication"

Returns top-10 matching nodes with metadata and neighbors.

uv run code-knowledge path "Session" "AuthToken"

uv run code-knowledge explain "ValidateToken"

Shows metadata, outgoing calls, and incoming references.

uv run code-knowledge index

# Add to PATH for direct execution
uv tool install --editable .

code-knowledge update /path/to/code
code-knowledge query "what handles authentication"

By default, code-knowledge downloads the embedding model (22MB) from Hugging Face Hub on first use. To avoid the HF warning and use offline:

# Download model once (requires internet)
./download-model.sh           # Linux/macOS
powershell -File download-model.ps1  # Windows

# Then all subsequent uses work offline
uv run code-knowledge update /path/to/code
uv run code-knowledge query "..."

The model is cached in the project at .cache/all-MiniLM-L6-v2/ and reused for all future runs. This makes the project portable and CI/CD-friendly:

• For CI/CD: Commit .cache/ to git or download it as part of setup
• For development: Run download-model.sh once, then work offline
• Size: ~350 MB (all-MiniLM-L6-v2 + tokenizers)

Built-in relation types: calls, imports, imports_from, contains, inherits, extends, implements, references, cites, conceptually_related_to, shares_data_with, semantically_similar_to, rationale_for

Each node is embedded as:

"{label} community {community_id} {relation} {neighbor_label} {relation} {neighbor_label} ..."

Example:

"Session community 0 imports authutil imports crypto contains validatetoken"

The embedding captures semantic context: the node's purpose (label), its community cohort, and its direct relationships.

detect(root)                  → find code files
    ↓
extract(files)               → AST extraction via tree-sitter
    ↓
build_graph(extractions)     → merge into NetworkX DiGraph
    ↓
cluster(G)                   → community detection (Leiden/Louvain)
    ↓
export(G, communities)       → graph.json
    ↓
sync_vectors(G)              → LanceDB vector store

Each stage is independent; you can skip vector syncing if you only need the graph.

uv sync --all-extras
uv run code-knowledge update .
uv run code-knowledge query "what extracts code"
uv run code-knowledge explain "extract"

• detect.py — code file discovery and filtering
• extract.py — tree-sitter AST walkers (21 languages)
• build.py — merge extractions into NetworkX graph
• cluster.py — Leiden/Louvain community detection
• export.py — export to graph.json
• vector.py — LanceDB sync and semantic search
• cache.py — per-file extraction cache (by SHA256)
• security.py — path and label sanitization
• validate.py — schema validation
• __main__.py — CLI commands

1. Install tree-sitter-<lang> package
2. Add extract_<lang>(path) function in extract.py
3. Register in the _DISPATCH dict
4. Add to CODE_EXTENSIONS in detect.py

Typical results on a 50K-file Python monorepo:

Total: ~2 minutes for 50K files / 100K nodes.

• Tree-sitter coverage: Functional code idioms (closures, pipes, higher-order functions) are partially supported; anonymous lambdas are skipped. Semantic extraction is code-based only (no LLM).
• Large graphs: Vector search scales to ~1M nodes; UI graph visualization is limited to 5K nodes.
• Precision: Extraction is AST-based (deterministic but not perfect). Cross-file call resolution is heuristic.

MIT