Hey HN, we want to share HelixDB, a project a few friends and I are working on to solve the fragmentation in AI data infrastructure.

Why a hybrid? While vector databases are great for similarity queries, many AI agents require both semantic search and complex relationship traversal. Currently, developers are forced to glue together separate vector and graph stores, leading to massive synchronization issues and latency. HelixDB solves this by providing a unified, native Graph-Vector database built in Rust.

Use Cases

• Indexing complex codebases: You can vectorize code snippets for similarity while using the graph engine to map function calls, dependencies, and inheritance structures.
• Agentic Knowledge Graphs: Store conversational history as entities and relationships, allowing agents to perform multi-hop reasoning over past interactions.
• Fraud Detection: Combine vector embeddings of transaction patterns with graph analysis to identify clusters of illicit activity across entity relationships.
• Molecule Discovery: Map biological interactions as a graph while using vector search for molecular structural similarity, as seen in these related approaches.

Technical Deep Dive

Many people have a thing against “yet another database” implemented from scratch, but we chose to write HelixDB in Rust to ensure memory safety and high-concurrency performance. By building the graph and vector engines on a shared storage layer, we avoid the overhead of traditional multi-database middleware. You can review our implementation details here: README.

Performance Benchmarks

In our initial testing, HelixDB shows competitive performance against industry standards. Our vector search latency is on par with Pinecone and Qdrant, while our graph traversal speeds are up to 3x faster than traditional setups like Neo4j because we eliminate the need for cross-database network hops. You can see a performance comparison video here: szgQu3yq.

Learn more about our architecture in our official docs: introduction. If you’d like to try it out in a simple RAG demo, you can follow our getting started guide at README. Many thanks! Comments and feedback welcome!