ladybug-rag-rs

Hybrid Graph RAG in Rust — combining vector search with graph intelligence for retrieval that actually finds what matters.

Built on LadybugDB concepts and powered by icebug graph algorithms.

Traditional RAG embeds documents into vectors and retrieves the nearest neighbors. This works for simple lookups but fails when answers require connecting information across documents, understanding entity relationships, or grasping the structural themes of a knowledge base.

Hybrid Graph RAG adds a knowledge graph layer and fuses four retrieval signals:

Results are merged using Reciprocal Rank Fusion (RRF) — chunks found by both vector and graph search receive the highest scores.

Benchmarks (on 500 technical documents vs vector-only RAG):

Three Rust crates, layered from C++ to application:

1. Ingest — split documents into chunks, extract entities, find relations
2. Index — build HNSW vector index + knowledge graph edges
3. Analyze — compute PageRank (importance) and Louvain (communities) via icebug
4. Retrieve — vector search + graph expansion, fused with RRF
5. Assemble — build LLM-ready context with sources, entities, and neighbors

LadybugDB is an embedded property graph database that unifies vector indexes, graph queries, and graph algorithms in a single .lbug file. It provides the conceptual foundation for this project's hybrid retrieval approach.

Learn more: LadybugDB — The Embedded Graph Database covers graph modeling, Cypher queries, vector indexing, and algorithm-powered applications in depth.

icebug is a high-performance graph analysis library (NetworKit fork) with 200+ algorithms and OpenMP parallelism. This project binds three key algorithms:

• PageRank — identifies structurally important entities, separating signal from noise
• Louvain — discovers topic communities for scoped retrieval
• BFS — powers multi-hop graph expansion from seed entities

• Rust toolchain (1.70+)
• CMake (3.14+)
• C++20 compiler (GCC 8+, Clang 6+, or AppleClang)
• OpenMP (brew install libomp on macOS)
• Apache Arrow (brew install apache-arrow on macOS, or via conda/system package manager)

git clone https://github.com/Volland/ladybug-rag-rs.git
cd ladybug-rag-rs
git submodule update --init --recursive
cargo build

cargo run -p ladybug-rag --example demo

This ingests two sample documents, builds a knowledge graph with PageRank and Louvain communities, and queries it:

=== Ladybug Hybrid Graph RAG Demo ===

Ingesting 2 documents...
Computing graph scores (PageRank + Louvain)...

Knowledge base statistics:
  Chunks:      16
  Entities:    49
  Mentions:    74
  Relations:   46
  Communities: 7

--- Query: "How does Rust ensure memory safety?" ---
  [1] score=0.0318 method=Hybrid source=rust_overview.md
  [2] score=0.0318 method=Hybrid source=rust_overview.md
  ...

cargo test

33 tests across all three crates: graph operations, PageRank, Louvain, BFS, chunking, entities, embeddings, vector search, RRF fusion, and end-to-end RAG queries.

The interactive Jupyter notebook walks through the full system with diagrams and live Rust execution.

1. Install Python dependencies:

pip install jupyter notebook

2. Build the Rust project first (the notebook calls cargo run):

cargo build
cargo build -p ladybug-rag --example demo

3. Start Jupyter from the project root:

cd ladybug-rag-rs
jupyter notebook

4. Open notebooks/hybrid_graph_rag_article.ipynb

5. Run all cells — the notebook will:

   • Execute the Rust demo via subprocess and display the output
   • Run the full test suite
   • Display architecture diagrams, algorithm explanations, and code samples

Note: If running from a different directory, the notebook auto-detects the project root. Make sure cargo is available on your PATH.

Read the full technical article explaining how graph algorithms solve the information overload problem:

Finding What Matters: How Graph Algorithms Tame the Information Explosion

Covers:

• Why vectors alone hit a wall at scale
• LadybugDB's unified vector + graph architecture
• How PageRank separates signal from noise
• How Louvain discovers topic structure
• The five-stage hybrid RAG pipeline
• When to use graph RAG vs. vector-only

All diagrams are available as editable Mermaid files (.mmd) and rendered PNGs:

To re-render after editing .mmd files:

for f in diagrams/*.mmd; do
  mmdc -i "$f" -o "${f%.mmd}.png" -b transparent -w 1200 -s 2
done

ladybug-rag-rs/
├── crates/
│   ├── icebug-sys/          # C FFI wrapper + CMake build
│   ├── icebug/              # Safe Rust API (Graph, PageRank, Louvain, BFS)
│   └── ladybug-rag/         # RAG engine
│       ├── src/
│       │   ├── chunker.rs       # Paragraph-aware text splitting
│       │   ├── embeddings.rs    # Embedder trait + SimpleEmbedder
│       │   ├── entities.rs      # Entity/relation extraction
│       │   ├── vector_store.rs  # Cosine similarity search
│       │   ├── graph_store.rs   # Knowledge graph (icebug-backed)
│       │   └── rag.rs           # HybridGraphRag + RRF fusion
│       ├── examples/            # Demo binary
│       └── tests/               # Integration tests
├── vendor/icebug/           # Git submodule
├── diagrams/                # Architecture diagrams (.mmd + .png)
├── article/                 # Technical article + illustrations
└── notebooks/               # Jupyter notebook walkthrough

• LadybugDB — The Embedded Graph Database — the book on building graph + vector applications
• icebug — 200+ graph algorithms, OpenMP parallelism, Apache Arrow integration
• ladybug-rag (Python) — the original Python implementation

MIT