Vector Graph RAG
Milvus-based multi-hop graph RAG library where two-stage Jev Noul scoring selects relations and reranks passages.
Overview
Vector Graph RAG, from Zilliz, does graph-style retrieval-augmented generation with pure vector search in Milvus, without a graph database. Entities and relations are encoded as vectors, and multi-hop evidence is retrieved via subgraph expansion, with either single-pass generative reranking or two-stage Jev scoring.
In the two-stage mode, Jev Noul judgments first select relevant graph relations and then rerank full source passages alongside vector-search candidates. The list reports 86.07 percent average Recall@5 across 1,000 MuSiQue and 1,000 2Wiki queries, and the README says this mode leads the compared baselines and reaches the quality-latency Pareto frontier. It uses Milvus Lite by default, works out of the box and is MIT licensed on PyPI.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is this graph relation relevant to the multi-hop question? | yes · no |
| Yes / no | Does this passage help answer the question? | yes · no |
How it works
- Index entities, relations and passages as vectors in Milvus.
- Retrieve candidates and expand a subgraph.
- Stage 1: Jev selects relevant relations.
- Stage 2: Jev reranks full passages; return top results.
Strengths
- No graph database needed.
- Reported 86.07 percent average Recall@5 on MuSiQue and 2Wiki.
- Zero-config default with Milvus Lite.
Caveats
- Benchmark results are the authors'.
- Jev mode needs a TypeSafe key.
Getting started
pip install vector-graph-ragAt a glance
- Type
- Library
- Status
- beta
- Language
- Python
- Runtime
- library (Milvus)
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Multi-hop retrieval: uses Jev Noul judgments to select graph relations and rerank full source passages alongside vector-search candidates, achieving 86.07% average Recall@5 across 1,000 MuSiQue and 1,000 2Wiki queries in the [two-stage evaluation](https://github.com/zilliztech/vector-graph-rag/blob/main/evaluation/jev/two-stage/README.md).
yibie/awesome-jevMilvus-based multi-hop retrieval library with two-stage Jev relation selection and full-passage reranking, achieving 86.07% average Recall@5 on MuSiQue and 2Wiki.
logicrw/awesome-jev-projects
Frequently asked questions
What does Jev do here?
It selects relevant relations and reranks passages using yes/no judgments.
Do I need Neo4j?
No, everything runs on vector search in Milvus.
How well does it retrieve?
86.07 percent average Recall@5 on MuSiQue and 2Wiki, per the reported evaluation.
Related projects
- jev-rerank★ 17
Jev as a calibrated reranker: query plus up to 30 documents in one call, one relevance Noul per document, benchmarked on BEIR.
ProjectbetaPython Terminal tool that finds link paths between Wikipedia articles: Jev ranks each page's outgoing links while Python runs the search.
ProjectbetaPythonCompares Jev's zero-shot reranking with specialist rerankers on MS MARCO passage and document benchmarks and a TREC WSJ transfer test.
ProjectresearchPythonGraph-database-agnostic GraphRAG framework where swappable Laya or Jev decisions score edges, route intents and gate hallucinations.
ServicebetaPythonLocal-first knowledge search with BM25 shortlisting and Jev Noul reranking by default, seven selectable pipelines and cited answers.
AppbetaPythonCLI and agent skill that shortlist notes in Obsidian or Logseq locally, then have Jev rerank them, returning original passages.
DemobetaJavaScript