Inconsistent Query Results Based on Output Fields Selection in Milvus Dashboard

**题意:**在Milvus仪表盘中基于输出字段选择的不一致查询结果

问题背景:

I'm experiencing an issue with the Milvus dashboard where the search results change based on the selected output fields.

I'm working on a RAG project using text data converted into embeddings, stored in a Milvus collection with around 8000 elements. Last week, my retrieval results matched my expectations ("good" results), however, this week, the results have degraded ("bad" results).

I found that when I exclude the embeddings_vector field from the output fields in the Milvus dashboard, I get the "good" results; Including the embeddings_vector field in the output changes the results to "bad".

I've attached two screenshots showing the difference in the results based on the selected output fields.

Any ideas on what's causing this or how to fix it?

Environment:

Python 3.11 pymilvus 2.3.2 llama_index 0.8.64

Thanks in advance!

python 复制代码
from llama_index.vector_stores import MilvusVectorStore
from llama_index import ServiceContext, VectorStoreIndex

# Some other lines..

# Setup for MilvusVectorStore and query execution
vector_store = MilvusVectorStore(uri=MILVUS_URI,
                                 token=MILVUS_API_KEY,
                                 collection_name=collection_name,
                                 embedding_field='embeddings_vector',
                                 doc_id_field='chunk_id',
                                 similarity_metric='IP',
                                 text_key='chunk_text')

embed_model = get_embeddings()
service_context = ServiceContext.from_defaults(embed_model=embed_model, llm=llm)
index = VectorStoreIndex.from_vector_store(vector_store=vector_store, service_context=service_context)
query_engine = index.as_query_engine(similarity_top_k=5, streaming=True)

rag_result = query_engine.query(prompt)

Here is the "good" result: "good" result And here is the "bad" result: "bad" result

问题解决:

I would like to suggest you to follow below considerations.

  • Ensure that your Milvus collection is correctly indexed. Indexing plays a crucial role in how search results are retrieved and ordered. If the index configuration has changed or is not optimized, it might affect the retrieval quality.
  • In your screenshots, the consistency level is set to "Bounded". Try experimenting with different consistency levels (e.g., "Strong" or "Eventually") to see if it impacts the results. Consistency settings can influence the real-time availability of the indexed data.
  • Review the query parameters, especially the similarity_metric. Since you're using IP (Inner Product) as the similarity metric, ensure that your embedding vectors are normalized correctly. Inner Product search works best with normalized vectors.
  • Verify that the embedding vectors are of consistent quality and scale. If there were changes in the embedding model or preprocessing steps, it could lead to variations in the search results.
  • The inclusion of the embeddings_vector field in the output might affect the way Milvus scores and ranks the results. It's possible that returning the raw embeddings affects the internal ranking logic. Ensure that including this field does not inadvertently alter the search behavior.
  • Check the Milvus server logs and performance metrics to identify any anomalies or changes in the search behavior. This might provide insights into why the results differ when the embeddings_vector field is included.
  • Ensure that there are no version mismatches between the client (pymilvus) and the Milvus server. Sometimes, discrepancies between versions can cause unexpected behavior.
  • As a last resort, try modifying your code to exclude the embeddings_vector field programmatically during retrieval and compare the results. This can help isolate whether the issue is indeed caused by including the embeddings in the output.
  • Please try out this code if it helps.
相关推荐
七夜zippoe17 小时前
Agent 告警设计:分级告警、智能降噪与值班排班
开发语言·ai·agent·告警·智能降噪·值班排班
wflynn1 天前
AI 产品 ChatPRD 遭遇 token 欺诈,滥用者用免费试用构建游戏地图
ai·技术
万里侯1 天前
日志聚类之后别把 noise 扔掉:降噪应该改变视图,不抹掉证据
人工智能·微服务·ai·容器·k8s
澹锦汐1 天前
纯前端秋日银杏与枫叶落叶色板智能聚类器:基于 K-Means 色彩量化与直方图自动配方提取
前端·人工智能·ai·语言模型·vue·react·k-means
最强小杰2 天前
某 OpenAI 兼容网关上的 `kimi-k3.1-pro`:Cline、Zed 接入与 SSE 排错示例
java·服务器·网络·ai
Maynor9962 天前
Trymodel 使用介绍:一个平台同时用 Claude、GPT、Gemini 和 Grok
人工智能·gpt·ai·程序员·教程·claude·语音
七夜zippoe2 天前
多 Agent 通信灾难:状态不一致导致雪崩故障的排查与防护
ai·agent·通信灾难·雪崩故障·排查与防护
Forget the Dream2 天前
Spring AI 2.0 系列(二):第一个 ChatClient——从注入到一次完整调用
java·人工智能·spring boot·ai·springai
zdr尽职尽责2 天前
UnityCli 兼容2022版本
ui·unity·ai·c#·游戏引擎
Sharewinfo_BJ2 天前
从“看数据”到“做决策”:AI时代,BI的下一个形态是什么?
ai·数据分析·bi·智信bi