Grafy News
Try Grafy free

UniGraph Launches Unified Graph Index for Hybrid Retrieval

A new preprint from the Grafy Research Lab proposes UniGraph, a graph‑based index that merges dense and sparse vectors into a single hybrid distance.

Illustration for the news story: UniGraph Launches Unified Graph Index for Hybrid RetrievalSource: Grafy Research Lab

What is UniGraph and why it matters?

A preprint released on 17 November 2026 by the Grafy Research Lab introduces UniGraph, a unified graph‑based index designed for dynamic hybrid dense‑sparse vector retrieval 【source】. The work targets a core limitation in current search systems that combine semantic (dense) and lexical (sparse) signals: they usually maintain two separate indexes, run two retrieval passes, and then merge the results. UniGraph proposes a single index that directly supports a hybrid distance, potentially lowering latency, reducing infrastructure cost, and preserving candidates that would otherwise be dropped before the final ranking.

How does hybrid dense‑sparse retrieval work today?

Hybrid retrieval aims to capture the strengths of two complementary representations. Dense vectors, generated by neural encoders, excel at capturing semantic similarity but can miss exact keyword matches. Sparse vectors, often derived from term‑frequency or BM25‑style models, excel at lexical matching but lack semantic nuance. Most production pipelines implement a two‑step process:

  1. Separate search – one pass over a dense index (e.g., FAISS) and another over a sparse index (e.g., Lucene).
  2. Weighted fusion – the query‑dependent weighted sum of the two distance scores is applied to the union of candidate sets, followed by a final re‑ranking. While effective, this approach doubles the query workload and can discard items that only appear high in the combined score after fusion, because they never make it into either individual candidate set.

What problem does a unified graph index solve?

The authors argue that the premature loss of candidates is a direct consequence of searching under two disjoint distance metrics. By constructing an index directly under the hybrid distance, UniGraph ensures that every item is evaluated according to the same combined metric from the start. This eliminates the need for two retrieval passes and guarantees that the top‑k results according to the hybrid score are present in the index’s candidate pool. In practice, this can lead to:

  • Faster query response times (single pass instead of two).
  • Lower compute and memory overhead (one index structure).
  • Higher recall for queries where lexical and semantic signals interact in non‑linear ways.

Technical approach of UniGraph

UniGraph extends the classic proximity graph family (e.g., HNSW, NSW) with a hybrid distance function. The distance between a query vector qq and a document vector dd is defined as:

Hybrid(q,d)=α⋅DenseDist(qdense,ddense)+(1−α)⋅SparseDist(qsparse,dsparse)\text{Hybrid}(q, d) = \alpha \cdot \text{DenseDist}(q_{dense}, d_{dense}) + (1-\alpha) \cdot \text{SparseDist}(q_{sparse}, d_{sparse})

where α\alpha is a query‑dependent weight that can be tuned on the fly. The graph is built by inserting vectors using this hybrid metric, so each edge reflects the combined similarity. During search, the algorithm traverses the graph exactly as traditional HNSW does, but comparisons are performed with the hybrid distance. The paper reports that this design “mitigates premature candidate loss” and matches or exceeds the recall of the two‑pass baseline while using roughly half the query latency.

Implications for developers and researchers

For practitioners building search‑oriented AI products, UniGraph offers a concrete path to single‑pass hybrid retrieval. The reduction from two indexes to one simplifies system architecture, which is especially valuable in cloud‑native environments where each model (dense encoder, sparse tokenizer) may be served by a different provider. Researchers can also experiment with the hybrid distance as a learnable parameter; the paper suggests that α\alpha can be predicted per query using a lightweight meta‑model, opening a research direction for adaptive retrieval.

How to explore UniGraph ideas with Grafy

Grafy provides several tools that let you prototype and evaluate the concepts described in the UniGraph paper without leaving the platform:

  • Reader (/reader) lets you upload the arXiv PDF, run keyword searches, and ask follow‑up questions to an LLM about the methodology.
  • Graph (/graph) enables you to build a visual, branching workflow where you can ingest a dataset, generate dense embeddings (via the Chat app’s model picker), create sparse term vectors, and then experiment with a custom hybrid distance function.
  • Code (/code) offers an AI‑assisted coding environment where you can implement a small‑scale hybrid proximity graph using Python libraries such as hnswlib and test it on a benchmark corpus. By chaining these apps, a creator can quickly move from reading the research, to visualising the index structure, to running a prototype search—all within Grafy’s unified workspace.

Frequently asked questions

Does UniGraph replace the need for separate dense and sparse indexes?

Yes. UniGraph builds a single proximity graph using a hybrid distance, so a query is processed through one index that accounts for both semantic and lexical similarity.

Can the hybrid weight α\alpha be learned automatically?

The authors suggest that α\alpha can be made query‑dependent, and a lightweight model could predict it at inference time. This remains an open research avenue, but the paper’s experiments used manually tuned values.

How can I test UniGraph on my own data within Grafy?

Upload your documents to Grafy Files, generate dense embeddings with the Chat model of your choice, create sparse vectors using a tokenizer, then use the Code studio to assemble a hybrid graph index. The Graph app can visualise the traversal steps, and Reader can help you compare recall against a two‑pass baseline.


Source: Grafy Research Lab — UniGraph: A Unified Graph-based Index for Dynamic Hybrid Dense-Sparse Vector Retrieval