Skip to content

doubao-seed-final-pack.md

Latest commit

 

History

History
136 lines (97 loc) · 6 KB

doubao-seed-final-pack.md

File metadata and controls

136 lines (97 loc) · 6 KB

Example Output: Doubao Seed Search / Ranking Final Pack

This is the expected style for 10_final_pack.md when the candidate has weak-to-medium evidence.

1. Fit Verdict

Verdict: risky fit
Why: The JD is search/ranking + LLM research oriented. The candidate has adjacent RAG and small search demo evidence, but lacks ranking metrics, strong algorithm internship evidence, and doc-understanding experiments.
Strongest match: Mini search demo with query-doc examples and bad-case notes.
Biggest gap: No NDCG/MRR report and no LLM-for-ranking experiment.
Highest interview risk: Claiming "优化搜索相关性" without being able to explain recall/rank/rerank and metrics.
Fastest useful upgrade: Convert the mini search demo into a reproducible BM25 vs embedding vs rerank comparison with NDCG@10/MRR and bad-case analysis.

2. JD Analysis

JD Requirement Hidden Interview Test User Evidence Strength Strategy
搜索相关性 / Ranking Explain query-doc relevance, recall/rank/rerank, NDCG/MRR 30 query-doc examples, no metric table medium Lead with offline analysis, not "improved ranking"
DOC 理解 Quality, authority, freshness, long-doc structure 12 bad cases, no rubric weak Add doc quality labels before writing strong claim
LLM + deep learning Knows LLM basics and eval notes only weak Say "学习/复现理解", not "训练大模型"
Code ability Python scripts, reproducibility scripts exist, README weak medium Add run command and clean README

3. Truth Boundary

Claim / Experience Status Can Write? Evidence Risk
优化知识库问答效果 补证据后写 Not yet RAG demo, no eval Sounds like fake metric
完成企业文档问答系统 谨慎写 Yes, as demo code + screenshot Avoid production wording
搜索排序相关性分析 可以写 Yes query-doc examples + bad cases Add metric to strengthen
熟悉 LLM 算法 谨慎写 Notes only learning notes Do not imply training experience

4. Evidence Contract

Resume Claim Evidence Status Safer Wording Interview Proof
围绕搜索相关性整理长尾 query-doc bad case 30 examples + 12 bad cases 可以写 Keep Show bad-case table and label rules
复现 BM25/embedding/rerank baseline notebook exists, rerank missing 补证据后写 复现 BM25 与 embedding retrieval baseline Run rerank and save metrics first
优化 RAG 效果 no eval 不能写 as strong result 梳理 RAG 召回错误与引用错位问题 Show retrieved chunks and failure cases

5. Targeted Resume Bullets

Conservative

围绕搜索相关性场景整理 30 组 query-doc 样例和 12 类长尾查询 bad case，分析歧义查询、时效性不足、低权威 DOC 等失败原因，梳理召回、粗排、重排阶段的输入输出与评估口径。

完成企业文档问答 demo 的文档切分、向量检索和 Prompt 模板配置，记录召回错误、引用错位和无依据回答等问题，为后续 RAG 评估集建设沉淀样例。

Standard After 1-Day Upgrade

基于小规模 query-doc 样例集对比 BM25、embedding 召回与 cross-encoder rerank baseline，使用 NDCG@10 / MRR 评估相关性表现，并输出长尾查询失败案例分析报告。

Stronger After Evidence

构建可复现的搜索相关性离线评估流程，围绕长尾查询、DOC 质量与时效性特征设计标注样例和 rerank 对比实验，沉淀指标报告、bad-case taxonomy 与后续优化方向。

6. Interview Grilling

  1. 召回、粗排、精排、rerank 分别解决什么问题？你的 demo 属于哪一层？
  2. 你怎么定义 query-doc 相关性？三档标签的边界是什么？
  3. 为什么 NDCG@10 比 accuracy 更适合排序？
  4. BM25 和 embedding retrieval 在长尾 query 上分别容易错在哪里？
  5. DOC 权威性和时效性怎么变成可计算特征？
  6. LLM 可以在哪些位置帮助搜索？哪些位置成本太高？
  7. 你的 RAG demo 有没有 citation？无答案时会不会拒答？
  8. 如果线上 CTR 上升但人工相关性下降，你怎么排查？

7. Answer Card

Question:

你说你优化了搜索排序，指标提升多少？

Dangerous answer:

提升挺明显的，用户看到的结果更准了。

Passable answer:

我现在不会写线上提升，因为没有线上流量和 CTR 记录。我做的是离线 query-doc 分析和 baseline 复现。

Strong answer after upgrade:

我用固定的 query-doc 样例集比较 BM25、embedding retrieval 和 cross-encoder rerank，并用 NDCG@10 / MRR 记录结果。我的重点是长尾 query bad-case taxonomy 和可复现评估流程，不把它包装成线上优化。

Evidence needed:

metric table, run command, labels, bad-case report, notebook or script

Resume wording to downgrade:

Remove "提升搜索排序效果". Use "复现并分析搜索相关性 baseline".

8. Upgrade Plan

Time Task Evidence Produced Resume Claim Unlocked
Half day Draw search pipeline and define labels diagram + label guide Can explain recall/rank/rerank
1 day Add NDCG@10/MRR for BM25 and embedding metric table + notebook Baseline comparison
3 days Add cross-encoder rerank and doc quality labels rerank report + bad cases LLM-assisted rerank analysis
1 week Publish reproducible mini search repo README + scripts + examples Strong personal project

9. Project Scout

Recommended directions:

  1. MiniMind-style from-scratch LLM repo for LLM algorithm basics.
  2. Small search/rerank baseline repo for BM25, embedding retrieval, and cross-encoder rerank.
  3. RAG evaluation repo focused on citation accuracy and answer grounding.

Hard boundary:

Do not write these as internship experience. Write them only after reproduction, modification, logs, and a short report exist.