Hey HN, we want to share HelixDB, a project our team is building to solve the fragmentation in AI memory layers. Why a hybrid? Traditional RAG pipelines rely solely on vector similarity, which works for finding snippets but fails when agents need to reason about explicit relationships or multi-hop dependencies. By combining property graphs with vector embeddings, HelixDB enables agents to traverse causal chains while simultaneously performing semantic lookups.

Why HelixDB?

• Knowledge Graph Construction: Extract entities from documents to build a map of dependencies, improving retrieval accuracy.
• Complex Agent Reasoning: Enable multi-hop queries where agents navigate relationships, such as finding project dependencies with high semantic similarity to specific requirements.
• Molecule Discovery: Map biological interactions using graph traversals while querying molecular structural similarity for drug discovery pipelines.
• Codebase Analysis: Index code structure as a graph to understand function dependencies alongside semantic documentation search.

Technical Deep Dive

Many developers have a thing against "yet another query language," but we implemented a custom graph-vector traversal syntax because it makes working with combined multidimensional data significantly more intuitive than forced SQL joins. We chose Rust for the underlying engine specifically to achieve memory safety and sub-millisecond execution speeds, addressing the skepticism regarding the overhead of graph engines in production pipelines. By avoiding the overhead of external SQL translation, we keep the data flow tight and performant.

Performance Benchmarking

Our internal benchmarking shows that we are on par with Pinecone and Qdrant for vector similarity performance, while our graph traversal capabilities are up to three orders of magnitude faster than traditional Neo4j deployments when integrated into high-concurrency RAG workflows. You can see our initial performance benchmarks and architecture docs here: introduction. We have also documented the specific graph-augmented RAG patterns we use here: graph rag.

If you would like to try it out in a simple RAG demo, you can follow this guide: rag demo. Many thanks! Comments and feedback welcome!