---
type: "post"
id: "improving-vector-graph-rag-with-two-stage-jev-19aa96"
slug: "improving-vector-graph-rag-with-two-stage-jev-19aa96"
title: "Improving Vector Graph RAG with Two-Stage Jev Reranking"
description: "Milvus says Vector Graph RAG now uses Jev in two retrieval stages: first to select graph relationships, then to rerank source passages alongside direct vector-search results."
retrieval_nugget: "Milvus says Vector Graph RAG now uses Jev in two retrieval stages: first to select graph relationships, then to rerank source passages alongside direct vector-search results. On 1,000 questions each from MuSiQue and 2Wiki, Jev 1.14 reached 86.61% average Recall@5. The result offers builders a measured retrieval trade-off, not evidence about final answer quality."
published_at: "2026-11-03"
updated_at: "2026-11-03"
record_date: "2026-11-03"
date_kind: "published_at"
topics: ["retrieval"]
entities: ["Milvus","Vector Graph RAG","Jev","MuSiQue","2Wiki","GPT-6 Luna","DeepSeek V4.1 Flash"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "Milvus reports a two-stage Jev pipeline for Vector Graph RAG, improving average Recall@5 by 6.03 points; Jev 1.14 reached 86.61%, the best of the configurations compared."
prior_refs: ["https://newruntime.com/posts/jev-vs-llms-in-agentic-search-analysis-acf0ca/"]
source_url: "https://milvus.io/blog/improving-vector-graph-rag-with-two-stage-jev-reranking.md"
source_urls: ["https://milvus.io/blog/improving-vector-graph-rag-with-two-stage-jev-reranking.md"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:improving-vector-graph-rag-with-two-stage-jev-19aa96"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/improving-vector-graph-rag-with-two-stage-jev-19aa96/","markdown":"https://newruntime.com/posts/improving-vector-graph-rag-with-two-stage-jev-19aa96.md","json":"https://newruntime.com/posts/improving-vector-graph-rag-with-two-stage-jev-19aa96.json"}
markdown_url: "https://newruntime.com/posts/improving-vector-graph-rag-with-two-stage-jev-19aa96.md"
---

# Improving Vector Graph RAG with Two-Stage Jev Reranking

## Retrieval answer

Milvus says Vector Graph RAG now uses Jev in two retrieval stages: first to select graph relationships, then to rerank source passages alongside direct vector-search results. On 1,000 questions each from MuSiQue and 2Wiki, Jev 1.14 reached 86.61% average Recall@5. The result offers builders a measured retrieval trade-off, not evidence about final answer quality.

## Summary
Milvus says Vector Graph RAG now uses Jev in two retrieval stages: first to select graph relationships, then to rerank source passages alongside direct vector-search results. On 1,000 questions each from MuSiQue and 2Wiki, Jev 1.14 reached 86.61% average Recall@5. The result offers builders a measured retrieval trade-off, not evidence about final answer quality.

## What happened
In an October 30, 2026 engineering post, Cheney Chen describes an updated Jev integration in Vector Graph RAG, an open-source project for multi-hop retrieval with Milvus. The system stores entities, relationships, and source passages separately in Milvus and links them by IDs. It uses vector similarity to find starting entities and relationships, then ID lookups to expand connected evidence without a separate graph database.

The earlier design used an LLM to select useful relationships after graph expansion. The first Jev integration replaced that relationship-selection step while retaining the searches, expansion, and passage lookups. The new version adds a second judgment stage: Jev selects relationships, their source passages are combined with passages from direct vector search, and Jev scores the complete passages against the question. The scores determine final retrieval order. LLMs remain in use for document triple extraction during ingestion, query entity extraction, and final answer generation.

The post illustrates the process with a question about side effects of a first-line treatment for type 2 diabetes. The question may not name metformin, so graph links can connect diabetes to the drug and then to passages about adverse effects. Relationship sources can also include less useful text, such as a passage about the drug's development history. Passage-level ranking gives the system another way to place directly relevant evidence above that material.

For the frozen benchmark recipe, the system retains up to 64 relationships and expands their sources in round-robin order. It interleaves those passages with the top 10 direct-retrieval results, deduplicates by full text, and caps the candidate pool at 16. Jev scores each passage; ties preserve the existing order.

The current evaluation used 1,000 questions each from MuSiQue and 2Wiki, weighted equally. Recall@5 measures the share of ground-truth supporting evidence in the top five retrieved results, independently of answer generation. Jev 1.13 single-stage averaged 80.04%; Jev 1.13 two-stage reached 86.07% with 3.13 seconds of additional model API time; Jev 1.14 two-stage reached 86.61% with about 4.9 seconds, estimated from direct API timing. The post also reports 82.70% for Vector Graph RAG with GPT-6 Luna and 83.34% with DeepSeek V4.1 Flash. Those comparisons use different additional API times, including 5.35 and 10.97 seconds respectively.

## What the archive adds
[“Jev vs. LLMs in Agentic Search”](https://newruntime.com/posts/jev-vs-llms-in-agentic-search-analysis-acf0ca/) covered Jev as a structured-decision model for bounded tasks in agentic search, including reranking memories. This is a new fact on that known story: Milvus describes a two-stage implementation in multi-hop retrieval and reports results on MuSiQue and 2Wiki. The archive also covered Milvus's SVS Vamana and LeanVec retrieval trade-offs in [“SVS in Milvus”](https://newruntime.com/posts/svs-in-milvus-reshaping-the-recall-memory-speed-0a935b/), a related but distinct retrieval-system development.

## Trend signal
No listed SITE_PATTERNS thesis directly fits this result. It concerns reranking evidence inside a retrieval pipeline and reports recall and latency; it does not establish a broader change in agent economics, protocols, runtime security, or work organization.

## Why now
The post explains the immediate motivation: in the earlier architecture, an LLM had to evaluate potentially hundreds or thousands of expanded relationship candidates, while vector searches and ID lookups were relatively fast. The new experiment tests whether Jev can improve passage ranking after relationship selection while preserving graph-based evidence discovery.

## Connections to other events
The earlier Jev coverage considered structured decisions such as deciding when a search agent should stop and reranking memories. The new Milvus result applies the same bounded-decision framing to two successive choices in one retrieval path: which relationships to keep and which passages should rank highest. It is a specific implementation and benchmark, not evidence that Jev replaces the LLM tasks that remain in the pipeline.

## Why it matters
For teams building retrieval-augmented systems, the reported change is architectural as much as it is model-specific. Graph expansion can discover evidence through relationships even when a query omits a linking entity; reranking the passages afterward addresses a separate problem, because a selected relationship can point to both useful and distracting source text. The reported two-stage results suggest that judging evidence at both levels can improve Recall@5 on the two tested datasets.

The trade-off is measurable but not free: Jev 1.14's reported 86.61% recall comes with approximately 4.9 seconds of additional model API time. Teams should compare that extra latency with the value of retrieving more supporting evidence in their own corpus and query mix. The results do not show that this configuration improves generated answers, lowers total cost, or is best for every retrieval workload. The practical takeaway is to test relationship selection and passage ranking as distinct stages rather than assuming one reranker must do both jobs.

## Limits and uncertainty
The current evaluation covers only MuSiQue and 2Wiki, with 1,000 questions per dataset and equal weighting. Recall@5 measures evidence retrieval, not answer correctness, user satisfaction, or end-to-end latency. The reported API-time figures are not all directly comparable: Jev 1.14's roughly 4.9 seconds is estimated, while some rows use published baseline aggregates and others are measured project runs. The post does not establish performance on other datasets, the effect of corpus size, operating cost, or how the pipeline behaves in production.

## Reader takeaway
The useful result is not simply that Jev 1.14 scores well: it is that Vector Graph RAG separates finding relevant relationships from ranking the passages those relationships uncover. The reported benchmark indicates a recall improvement on two datasets, with additional model API time. Treat it as a promising design to reproduce against your own retrieval workload, not as proof of better final answers.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Retrieval](https://newruntime.com/topics/retrieval.md): Explore the retrieval topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)