@king1818888: Just the day before yesterday, I went for an interview with the Agent implementation team at a certain AI unicorn compa…

X AI KOLs Timeline News
rag hybrid-search reranker bm25 embeddings vector-search

Summary

一个求职者分享了面试 AI Agent 实施团队时被 RAG 检索细节问题难住的经历，并总结出生产级 RAG 需结合向量检索与 BM25 关键词搜索的混合检索，再用 Cross-encoder Reranker 进行二次精排，以解决语义相近但实体错误的检索问题。

Just the day before yesterday, I went for an interview with the Agent implementation team at a certain AI unicorn company, and ended up completely stumped by a question on RAG retrieval details. The interviewer's challenge: The interviewer leaned back in his chair: "A user asks the intelligent customer service, 'How much is the latest Apple phone?' But vector retrieval ranks 'wholesale price per jin of Red Fuji apples' from the knowledge base as the top result. How would you thoroughly solve this problem?" What I was thinking inside: At the time, I was silently chanting to myself in my head: "Big bro, how is this my fault? This is clearly because OpenAI or the open-source Embedding model's word vector clustering isn't accurate enough. Go argue with the algorithm scientists about it—why take it out on me, just an application developer?!" What I actually said out loud: "Uh... switch to an Embedding model with more parameters? Or make the text chunking granularity a bit smaller?" The interviewer's reaction: The interviewer smiled: "Even if you switch to the biggest model, an apple is still an apple, and the semantics will still diverge. Go take a look at how industry pros twist keyword search and semantic search together." That very evening, I posted in a tech community to consult a few senior folks, and suddenly realized I'd been thinking about RAG way too simplistically. Pure Dense vector retrieval is inherently insensitive to proper nouns and specific models. Production-grade solutions have to use hybrid search (Hybrid Search: vector cosine similarity + BM25 keyword exact matching), then pair it with a post-retrieval Reranker (cross-encoder reranking model) for a second round of fine-grained scoring—that's the standard way to wipe out the "Red Fuji phone" issue.
Original Article
View Cached Full Text

Cached at: 10/12/26, 03:29 PM

Just the day before yesterday, I went for an interview with the Agent implementation team at a certain AI unicorn company, and ended up completely stumped by a question on RAG retrieval details.

The interviewer’s challenge: The interviewer leaned back in his chair: “A user asks the intelligent customer service, ‘How much is the latest Apple phone?’ But vector retrieval ranks ‘wholesale price per jin of Red Fuji apples’ from the knowledge base as the top result. How would you thoroughly solve this problem?”

What I was thinking inside: At the time, I was silently chanting to myself in my head: “Big bro, how is this my fault? This is clearly because OpenAI or the open-source Embedding model’s word vector clustering isn’t accurate enough. Go argue with the algorithm scientists about it—why take it out on me, just an application developer?!”

What I actually said out loud: “Uh… switch to an Embedding model with more parameters? Or make the text chunking granularity a bit smaller?”

The interviewer’s reaction: The interviewer smiled: “Even if you switch to the biggest model, an apple is still an apple, and the semantics will still diverge. Go take a look at how industry pros twist keyword search and semantic search together.”

That very evening, I posted in a tech community to consult a few senior folks, and suddenly realized I’d been thinking about RAG way too simplistically. Pure Dense vector retrieval is inherently insensitive to proper nouns and specific models. Production-grade solutions have to use hybrid search (Hybrid Search: vector cosine similarity + BM25 keyword exact matching), then pair it with a post-retrieval Reranker (cross-encoder reranking model) for a second round of fine-grained scoring—that’s the standard way to wipe out the “Red Fuji phone” issue.

Similar Articles

@knoYee_: https://x.com/knoYee_/status/2062780637677752366

X AI KOLs Timeline

The author reviews three months of experience using multi-agent collaboration, summarizing five main pain points (such as conflicts between agents, ignoring boundary conditions, self-censorship failure, difficulty in merging decisions, and exposing harder problems after compressed execution) and two insights (the high value of read-only review agents, and that agent conflicts expose ambiguous requirements), emphasizing the core decision-making role of humans in AI collaboration.

@cxjwin: https://x.com/cxjwin/status/2112056767702286396

X AI KOLs Timeline

An in-depth article on a high-frequency Agent interview question — "Walk me through your RAG pipeline" — systematically covering the principles, trade-offs, and common follow-up questions across the full RAG workflow: offline indexing (parsing, chunking, vectorization, index building) and online QA (retrieval, reranking, context construction).

RAG Retrieval Deep Dive: BM25, Embeddings, and the Power of Agentic Search

YouTube AI Channels

This article provides an in-depth comparison of the advantages and disadvantages of BM25 lexical search and embedding semantic search in RAG retrieval, offers a practical framework for selecting retrieval methods based on query type and system trade-offs, and emphasizes the importance of treating RAG as a system rather than a simple component.

@blanplan: https://x.com/blanplan/status/2056614589236998459

X AI KOLs Timeline

DeepSeek's job description analysis for the 'Agent Harness R&D Engineer' position reveals the company's pursuit of a 'researcher + engineer + community operations' triple-threat talent, as well as its in-depth usage requirements for AI Agent tools.

@ando_w: Today I interviewed a candidate whose resume said "Proficient in RAG". I asked about the recall rate, he said "It's pretty good." I asked how it was evaluated, he said "I asked a few questions and it felt fine." This is the current state of the industry: 90% of RAG projects have no evaluation set, all relying on gut feeling. Those who can say Recall and Precision are already...

X AI KOLs Timeline

An interviewer complained that a candidate claiming to be "proficient in RAG" could not state the recall rate and precision, pointing out that 90% of RAG projects in the industry lack standardized evaluation sets and rely entirely on subjective judgment.