RAG is dead, right?? — Kuba Rogut, Turbopuffer

What was stored

cues

114

whisperx 114

chunks

19

from 114 cues

keyframes

17

kept of 30 captured

frames with text

17

471 lines read

chapters

8

from the source metadata

keyframe bytes

3.6 MB

word timings on 114 cues

Provenance

Frames, and what the machine read

• 

  0:03 #0 done2 line(s)

  shot 0·sharpness 662.1

  1. AlEngineer0.98
  2. EUROPE1.00
• 

  0:08 #1 done2 line(s)

  shot 1·sharpness 816.0

  1. PRESENTINGSPONSOR1.00
  2. Google DeepMind1.00
• 

  0:12 #2 done3 line(s)

  shot 2·sharpness 905.6

  1. PLATINUM SPONSORS0.98
  2. # Braintrust0.96
  3. WorkOS OpenAI0.94
• 

  0:18 #3 done4 line(s)

  shot 3·sharpness 271.3

  1. turbopuffer1.00
  2. Rag is dead, right??0.98
  3. Kuba Rogut1.00
  4. AIF0.86
• 

  0:23 #4 done12 line(s)

  shot 4·sharpness 1303.3

  1. turbopuffer1.00
  2. Rag is dead, right??0.99
  3. AIE1.00
  4. How hybrid, tool-rich retrieval is becoming the default1.00
  5. ★1.00
  6. ★1.00
  7. ★0.99
  8. for serious agentic search1.00
  9. Kuba Rogut0.99
  10. Deployed Engineer1.00
  11. April 9, 20260.97
  12. Google DeepMind1.00
• 

  1:02 #5 done108 line(s)

  shot 5·sharpness 3541.0

  1. "RAG is dead"0.99
  2. turbopuffer1.00
  3. RAG is dead that's why we actually made sure the model could simply use1.00
  4. Replying to @dinOs1.00
  5. Antoine Chaffin@antoine_chaffin · Feb 120.97
  6. Pratik Desai@chheplo - Apr 29, 20240.97
  7. RAG is a makeshift solution.1.00
  8. RAG is bad.1.00
  9. σ ...0.52
  10. swyx@swyx·Oct 13, 20250.95
  11. relatively consensus now that "traditional embeddings RAG" is dead.0.99
  12. @jerryjliu0 called it first at the 2024 AIE WF - Agentic RAG was clearly0.96
  13. grep instead of leveraging our models0.98
  14. RAG will die soon.1.00
  15. better in almost every way except speed.1.00
  16. Praveen Naik @p_naix · Dec 29, 20250.95
  17. O10.68
  18. That way we made sure RAG wouldn't try to revive1.00
  19. O70.65
  20. d 2220.86
  21. 口0.54
  22. 651.00
  23. Manthan Gupta@manthanguptaa · Feb 250.95
  24. After "RAG is dead* we are now in the "Agents only need filesystems* Al0.98
  25. t2 250.83
  26. 3060.99
  27. th1 64K0.81
  28. 口0.57
  29. Q ….0.52
  30. will be sharing more about the journey to Agentic Search and what's next at1.00
  31. @elastic's sf closing0.99
  32. Show more1.00
  33. It's funny how we jokingly said RAG is dead and it actuall died0.98
  34. hype cycle1.00
  35. ★0.76
  36. ★1.00
  37. ★1.00
  38. AIE1.00
  39. ★1.00
  40. ★1.00
  41. ★1.00
  42. MANTU KUMAR0.97
  43. @Mantu_kumar911.00
  44. O10.67
  45. t20.91
  46. h 1020.83
  47. 口0.72
  48. Eli Mernit@mernit · Feb 240.91
  49. JUST USE FILES.1.00
  50. SAN FRANCISCO . OCT 210.95
  51. Keynote: The0.99
  52. elastic0.99
  53. Year Agents Ate0.99
  54. ★1.00
  55. ★1.00
  56. ★1.00
  57. Skill sikhte hi dead ho jaata hai0.98
  58. 0.03-0.93
  59. Search1.00
  60. RAGisDead1.00
  61. SHAWN WANG1.00
  62. X Article0.98
  63. Today, nearly every agent framework is using some mix of Postgres,0.99
  64. Redis, and a vector DB to manage context. This is a mistake.1.00
  65. Agents Don't Need Databases0.99
  66. Rag is dead. Spinning up 20 Codex spark subagents to search your0.99
  67. jason liu@jxnlco - Mar 50.96
  68. filesystem is the solution.0.98
  69. 601.00
  70. t7 140.88
  71. 4081.00
  72. da 44K0.84
  73. Q0.53
  74. Boris Cherny1.00
  75. 0 .0.50
  76. @bcherny1.00
  77. Doug Turnbull@softwaredoug · Mar 30.95
  78. 0 .0.66
  79. Early versions of Claude Code used RAG + a local vector db, but we0.99
  80. RAG is dead again1.00
  81. Piyush Garg1.00
  82. found pretty quickly that agentic search generally works better. It is also1.00
  83. simpler and doesn't have the same issues around security, privacy,1.00
  84. staleness, and relliability.0.99
  85. GPT-5.4 leak: 2M token context + persistent state = KV cache explosion0.99
  86. Ben Pouladian@benitoz ·Mar 10.97
  87. hesam@Hesamation · Oct 15, 20250.95
  88. people are posting "RAG is dead" and half are constantly talking about it.0.99
  89. I don't understand why everyone is so obsessed with RAG. Half of the0.98
  90. It's 2025. Shouldn't RAG be done-and-dusted by now?0.99
  91. 9:56 PM - Jan 31, 2026 - 1.1M Views0.97
  92. Q 1510.79
  93. t23770.91
  94. 5x0.65
  95. 口2.6K0.86
  96. This is the Memory Wars in real time0.97
  97. HBM for weights. SRAM for latency-critical inference. Optical..0.98
  98. Q50.85
  99. t0.86
  100. ○90.66
  101. dt 1.8K0.92
  102. 口0.64
  103. Relevant1.00
  104. View quotes>0.95
  105. RAGis Dead0.96
  106. Braintrust1.00
  107. WorkOS1.00
  108. OpenAI0.97
• 

  1:14 #6 done24 line(s)

  shot 6·sharpness 1656.5

  1. "RAG is dead"?0.99
  2. turbopuffer1.00
  3. Retrieval Augmented Generation1.00
  4. Search term1.00
  5. +0.96
  6. United States1.00
  7. Mar 1, 2021 - Feb 27, 20260.99
  8. Web Search1.00
  9. AIE1.00
  10. ★1.00
  11. ★1.00
  12. ★1.00
  13. ★1.00
  14. Interest over time ①0.95
  15. United States - Mar 1, 2021 - Feb 27, 20260.98
  16. 1001.00
  17. MGoogle Tronds0.94
  18. 20221.00
  19. 20231.00
  20. 20240.99
  21. 20250.99
  22. 20261.00
  23. AlEngineer0.97
  24. EUROPE1.00
• 

  1:59 #7 done29 line(s)

  shot 7·sharpness 4241.3

  1. Let's clarify: RAG vs Agentic Search0.99
  2. turbopuffer1.00
  3. RAG1.00
  4. AgenticSearch1.00
  5. What people *think* it is:0.99
  6. What people *think* it is:0.98
  7. 4 Vector search0.96
  8. 4Filesystem grep0.99
  9. ★0.90
  10. AIE1.00
  11. ★1.00
  12. What it *actually* is:0.97
  13. What it *actually* is:0.99
  14. ★1.00
  15. ★1.00
  16. Retrieval Augmented Generation1.00
  17. giving agents a set of tools to0.99
  18. progressively find and reason over1.00
  19. vector1.00
  20. LLM1.00
  21. external context0.97
  22. bm251.00
  23. harness1.00
  24. grep1.00
  25. glob1.00
  26. regex1.00
  27. filter...0.99
  28. AlEngineer0.96
  29. EUROPE1.00
• 

  2:32 #8 skipped

  shot 8·duplicate of #7

• 

  3:04 #9 skipped

  shot 9·duplicate of #7

• 

  3:41 #10 done55 line(s)

  shot 10·sharpness 4577.8

  1. cursor.com/blog/secure-codebase-indexing1.00
  2. turbopuffer1.00
  3. CURSOR1.00
  4. Merkle trees enable fast, incremental sync1.00
  5. Securely indexing large codebases0.99
  6. & Client0.92
  7. Repository service1.00
  8. Jan 27, 2026 by Jeremy Stribling in Research1.00
  9. /docs #g7n2d20.97
  10. compare hashes1.00
  11. of entries1.00
  12. /docs #w5g3u20.97
  13. ★1.00
  14. guide.md #b8c9d00.96
  15. sync0.93
  16. guide.md#x9y8z70.93
  17. AIE1.00
  18. ★0.99
  19. diagram.py#z8s7130.97
  20. …no updates needed…0.93
  21. diagram,py#z8s7130.97
  22. ★1.00
  23. Semantic search is one of the biggest drivers of agent performance. In our recent1.00
  24. ★1.00
  25. ★1.00
  26. ★1.00
  27. evaluation, it improved response accuracy by 12.5% on average, produced code changes0.99
  28. that were more likely to be retained in codebases, and raised overall request satisfaction.1.00
  29. No matching on client1.00
  30. × delete0.94
  31. config.json #z1x2y30.91
  32. To power semantic search, Cursor builds a searchable index of your codebase when you0.99
  33. open a project. For small projects, this happens almost instantly. But large repositories with1.00
  34. Secure index sharing enables fast, safe onboarding0.98
  35. tens of thousands of files can take hours to process if indexed naively, and semantic0.99
  36. if hashes differ0.98
  37. search isn't available until at least 80% of that work is finished.0.99
  38. & Client0.96
  39. Indexing worker1.00
  40. 目 Vector DB0.90
  41. We looked for ways to speed up indexing based on the simple observation that most teams1.00
  42. work from near-identical copies of the same codebase. In fact, clones of the same0.99
  43. Merkle trees0.98
  44. Embedding model1.00
  45. → DynamoDB cache0.94
  46. codebase average 92% similarity across users within an organization.0.99
  47. Repository service1.00
  48. localCodebase DB0.98
  49. This means that rather than rebuilding every index from scratch when someone joins or0.99
  50. Secure index copies0.98
  51. switches machines, we can securely reuse a teammate's existing index. This cuts time-to-0.99
  52. if hashes match0.95
  53. first-query from hours to seconds on the largest repos.1.00
  54. AlEngineer0.96
  55. EUROPE1.00
• 

  4:08 #11 skipped

  shot 11·duplicate of #10

• 

  4:27 #12 done55 line(s)

  shot 12·sharpness 4602.9

  1. cursor.com/blog/secure-codebase-indexing1.00
  2. turbopuffer1.00
  3. CURSOR1.00
  4. Merkle trees enable fast, incremental sync1.00
  5. Securely indexing large codebases0.99
  6. & Client0.92
  7. Repository service1.00
  8. ★1.00
  9. Jan 27, 2026 by Jeremy Stribling in Research1.00
  10. /docs #g7n2d20.97
  11. compare hashes1.00
  12. of entries0.96
  13. /docs #w5g3u20.97
  14. ★1.00
  15. guide.md #b8c9d00.96
  16. sync0.94
  17. guide.md #x9y8z70.93
  18. AIE1.00
  19. ★1.00
  20. diagram.py#z8s7130.99
  21. …no updates needed0.95
  22. diagram.py #z8s7130.94
  23. Semantic search is one of the biggest drivers of agent performance. In our recent0.99
  24. ★1.00
  25. ★1.00
  26. ★1.00
  27. ★1.00
  28. evaluation, it improved response accuracy by 12.5% on average, produced code changes0.99
  29. that were more likely to be retained in codebases, and raised overall request satisfaction.1.00
  30. No matching on client1.00
  31. × delete0.95
  32. config.json #z1x2y30.94
  33. To power semantic search, Cursor builds a searchable index of your codebase when you0.99
  34. open a project. For small projects, this happens almost instantly. But large repositories with1.00
  35. Secure index sharing enables fast, safe onboarding0.98
  36. tens of thousands of files can take hours to process if indexed naively, and semantic0.99
  37. if hashes differ0.98
  38. search isn't available until at least 80% of that work is finished.0.99
  39. & Client0.95
  40. Indexing worker0.99
  41. 目 Vector DB0.94
  42. We looked for ways to speed up indexing based on the simple observation that most teams1.00
  43. work from near-identical copies of the same codebase. In fact, clones of the same0.99
  44. Merkle troes0.96
  45. Embedding model1.00
  46. → 目 DynamoDB cache0.93
  47. codebase average 92% similarity across users within an organization.0.99
  48. Repository service1.00
  49. localCodebase DB0.98
  50. This means that rather than rebuilding every index from scratch when someone joins or0.99
  51. Secure index copies0.99
  52. switches machines, we can securely reuse a teammate's existing index. This cuts time-to-0.99
  53. it hashes match0.95
  54. first-query from hours to seconds on the largest repos.1.00
  55. Engineering the future of Al1.00
• 

  5:07 #13 done52 line(s)

  shot 13·sharpness 3859.1

  1. cursor.com/blog/semsearch1.00
  2. turbopuffer1.00
  3. CURSOR1.00
  4. All models improve with semantic search0.99
  5. Improving agent with semantic search1.00
  6. Model (alphabetical)1.00
  7. Relative improvement (Cursor Context Bench)0.99
  8. Nov 6, 2025 by Stefan Heule, Emily Jia & Naman Jain in Research0.99
  9. Composer1.00
  10. 23.5%1.00
  11. ★1.00
  12. Table of Contents1.00
  13. Gemini 2.5 Pro1.00
  14. 8.7%1.00
  15. AIE1.00
  16. ★1.00
  17. Offline evals1.00
  18. GPT-51.00
  19. 6.5%1.00
  20. ★1.00
  21. Online A/B tests0.94
  22. Custom retrieval models0.99
  23. Grok Code1.00
  24. 11.9%1.00
  25. ★1.00
  26. ★1.00
  27. ★1.00
  28. Conclusion1.00
  29. Sonnet 4.51.00
  30. 14.7%1.00
  31. When coding agents receive a prompt, returning the right answer requires building an0.99
  32. understanding of the codebase by reading files and searching for relevant information.0.99
  33. Semantic search improves code retention1.00
  34. matching natural language queries, such as "where do we handle authentication?", in0.99
  35. One tool Cursor's agent uses is semantic search, which retrieves segments of code0.99
  36. and reduces dissatisfied user requests0.99
  37. addition to the regex-based searching provided by a tool like grep.1.00
  38. Code Retention1.00
  39. +0.3%1.00
  40. pipelines for fast retrieval. While you could rely exclusively on grep and similar command-1.00
  41. To support semantic search, we've trained our own embedding model and built indexing0.99
  42. Code Retention (large codebases)1.00
  43. +2.6%1.00
  44. line tools for search, we've found that semantic search significantly improves agent1.00
  45. Dissatisfied User Requests1.00
  46. -2.2%1.00
  47. performance, especially over large codebases:0.99
  48. Achieving on average 12.5% higher accuracy in answering questions (6.5%-23.5%0.99
  49. Improvement of having semantic search (vs same model without semantic search).0.99
  50. For "Code Retention" higher is better, and for "Dissatisfied User Requests" lower is better.0.99
  51. depending on the model).1.00
  52. Engineering the future of Al1.00
• 

  5:16 #14 skipped

  shot 14·duplicate of #13

• 

  5:46 #15 skipped

  shot 15·duplicate of #13

• 

  6:12 #16 skipped

  shot 16·duplicate of #6

• 

  6:51 #17 done41 line(s)

  shot 17·sharpness 4348.6

  1. embeddings are cached compute1.00
  2. per-sessiondiscovery1.00
  3. amortized understanding1.00
  4. Agent greps, reads, assesses, repeats0.99
  5. index once, retrieve at runtime1.00
  6. grep "metadata filter" src/1.00
  7. +820 tokens1.00
  8. Index time:1.00
  9. ★1.00
  10. read indexing.py (wrong section)1.00
  11. +2,106 tokens1.00
  12. - Codebase chunked, embedded, indexed0.98
  13. ★1.00
  14. AIE1.00
  15. grep "ingest pipeline" src/1.00
  16. +640 tokens1.00
  17. - Semantic meaning encoded upfront1.00
  18. ★1.00
  19. ★1.00
  20. read api_client.ts:280-4001.00
  21. +1,450 tokens1.00
  22. - Model has already "read" every file0.99
  23. ★1.00
  24. ★1.00
  25. read types.ts (for context)1.00
  26. +1,298 tokens1.00
  27. - One-time cost, amortized across all sessions0.99
  28. Runtime:1.00
  29. Repeated every session by every0.99
  30. Agent query: "How is metadata filtered?"0.98
  31. agent, across every task1.00
  32. Results(ranked chunks):1.00
  33. → 6,314 tokens0.96
  34. - indexing.py1.00
  35. +142 tokens1.00
  36. - api_client.ts0.99
  37. +188 tokens0.96
  38. types.ts1.00
  39. +94 tokens1.00
  40. → 424 tokens0.97
  41. Engineering the future of Al1.00
• 

  7:00 #18 skipped

  shot 18·duplicate of #17

• 

  7:40 #19 skipped

  shot 19·duplicate of #17

• 

  7:59 #20 skipped

  shot 20·duplicate of #17

• 

  8:27 #21 done43 line(s)

  shot 21·sharpness 4325.7

  1. embeddings are cached compute1.00
  2. per-session discovery1.00
  3. amortized understanding1.00
  4. Agent greps, reads, assesses, repeats0.99
  5. index once, retrieve at runtime0.99
  6. grep "metadata filter" src/1.00
  7. +820 tokens1.00
  8. Index time:1.00
  9. ★1.00
  10. ★1.00
  11. read indexing.py (wrong section)0.99
  12. +2,106 tokens1.00
  13. - Codebase chunked, embedded, indexed0.98
  14. ★1.00
  15. AIE1.00
  16. grep "ingest pipeline" src/0.99
  17. +640 tokens1.00
  18. - Semantic meaning encoded upfront0.99
  19. ★1.00
  20. ★1.00
  21. read api_client.ts:280-4000.98
  22. +1,450 tokens0.98
  23. - Model has already "read" every file0.99
  24. ★1.00
  25. ★1.00
  26. read types.ts (for context)1.00
  27. +1,298 tokens1.00
  28. - One-time cost, amortized across all sessions0.99
  29. Runtime:1.00
  30. Repeated every session by every0.99
  31. Agent query: "How is metadata filtered?"1.00
  32. agent, across every task1.00
  33. Results(ranked chunks):1.00
  34. → 6,314 tokens0.97
  35. - indexing.py1.00
  36. +142 tokens0.99
  37. - api_client.ts0.99
  38. +188 tokens0.96
  39. types.ts1.00
  40. +94 tokens1.00
  41. → 424 tokens0.98
  42. AlEngineer0.95
  43. EUROPE1.00
• 

  8:57 #22 done28 line(s)

  shot 22·sharpness 3853.2

  1. from RAG to agentic retrieval0.99
  2. turbopuffer1.00
  3. Haider.1.00
  4. Subscribe1.00
  5. @slow_developer1.00
  6. pattern1.00
  7. Google Jeff Dean says bigger context windows alone are not enough0.99
  8. From RAG → Agentic Retrieval0.98
  9. What matters is staged retrieval: lightweight mechanisms that narrow a0.99
  10. trillion tokens down to 10 million, then to the million you actually need1.00
  11. AIE1.00
  12. old1.00
  13. new1.00
  14. "you don't need a trillion at once, you need the right million"0.99
  15. ★1.00
  16. ★1.00
  17. ★1.00
  18. ★1.00
  19. @slow_developer1.00
  20. Retrieve once1.00
  21. Reason in steps1.00
  22. Stuff the prompt1.00
  23. Search as needed1.00
  24. Cross fingers1.00
  25. - Fetch what is useful0.99
  26. Retrieval is now iterative with tools0.99
  27. AlEngineer0.97
  28. EUROPE1.00
• 

  9:19 #23 skipped

  shot 23·duplicate of #22

Transcript

114 cues· 2,014 words· 10,912 chars

Chapters

1. 0:00 Introduction to the "RAG is dead" discourse
2. 1:12 Google search volume trends for RAG
3. 1:39 Defining RAG vs. Agentic Search
4. 3:15 Cursor's indexing and semantic search approach
5. 6:10 Contrasting Claude Code (grep) vs. Cursor (indexed)
6. 6:40 The concept of embeddings as cached compute
7. 8:38 The shift from simple RAG to Agentic Retrieval
8. 9:44 Jeff Dean on context windows and stage retrieval