--- schema: 1 type: video title: Stop Fine-Tuning to Fix Retrieval Problems — Anant Srivastava description: >- Most AI teams never decide where their knowledge lives. Six months of normal product work decides for them. Anant Srivastava, principal technologist for data and AI platforms at Oracle, argues that prompt, memory and weights aren't a ladder you climb when answers go wrong. They're three tools for three jobs. Through a support assistant that kept inventing product names after a fine-tune, he shows how architecture builds up by accident. He then gives a simple diagnostic for each: the prompt is for small, stable behavior; memory is for knowledge that's current, large, citable or access-controlled; weights are only for reflexes that have stopped changing. He closes with a circulating architecture where the agent gets better by doing its job. In this talk: • How every prompt edit, indexed doc and training sample is an architecture decision • The diagnostic question for prompt, memory and weights • Why you fine-tune reflexes, not facts • Moving knowledge between context, memory and weights over time SPEAKER Anant Srivastava, Principal Technologist, Data and AI Platforms, Oracle LinkedIn: https://www.linkedin.com/in/anantds/ CHAPTERS 0:00 Where should knowledge live? 0:15 The real engineering problem 0:50 Three tools, not a ladder 1:20 How teams decide by accident 2:00 Every edit is an architecture decision 2:40 A support assistant goes wrong 4:34 Accumulated, not designed 5:14 Prompt: behavior, not facts 5:59 Example: a support agent prompt 6:59 Memory: current, large, citable 8:09 The wrong job for memory 8:49 Access control belongs in memory 9:14 Example: a code assistant 10:08 Chunking and metadata for code RAG 11:23 Weights: what has stopped changing 12:28 The fine-tuning mistake 13:53 When fine-tuning works 15:08 Medical coding: fine-tune the reflex 16:13 Capability or cost? 16:58 The decision table 17:33 A circulating architecture 19:27 The model is the easy part Recorded at the AI Engineer World's Fair 2026 in San Francisco. Subscribe for more talks from the engineers building with AI. AI Engineer: https://ai.engineer YouTube: https://www.youtube.com/@aiDotEngineer X: https://x.com/aiDotEngineer LinkedIn: https://www.linkedin.com/company/aidotengineer/ #RAG #FineTuning #AIEngineer canonical: >- https://liquidgold.cc/channel/aidotengineer/explore/video/stop-fine-tuning-to-fix-retrieval-problems-anant-sr-11b6d54c81d7 markdown: >- https://liquidgold.cc/channel/aidotengineer/video/stop-fine-tuning-to-fix-retrieval-problems-anant-sr-11b6d54c81d7.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-23T22:20:03.395Z' language: null children: [] videoId: qflLT3SoVbw durationSeconds: 1212 publishedAt: '2026-10-04T00:00:35.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 36 endSeconds: 61 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=36' heldReason: rights - ordinal: 1 text: null startSeconds: 61 endSeconds: 72 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=61' heldReason: rights - ordinal: 2 text: null startSeconds: 93 endSeconds: 126 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=93' heldReason: rights - ordinal: 3 text: null startSeconds: 137 endSeconds: 163 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=137' heldReason: rights - ordinal: 4 text: null startSeconds: 173 endSeconds: 276 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=173' heldReason: rights - ordinal: 5 text: null startSeconds: 286 endSeconds: 321 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=286' heldReason: rights - ordinal: 6 text: null startSeconds: 331 endSeconds: 365 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=331' heldReason: rights - ordinal: 7 text: null startSeconds: 378 endSeconds: 421 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=378' heldReason: rights - ordinal: 8 text: null startSeconds: 421 endSeconds: 431 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=421' heldReason: rights - ordinal: 9 text: null startSeconds: 431 endSeconds: 499 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=431' heldReason: rights - ordinal: 10 text: null startSeconds: 510 endSeconds: 534 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=510' heldReason: rights - ordinal: 11 text: null startSeconds: 544 endSeconds: 558 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=544' heldReason: rights - ordinal: 12 text: null startSeconds: 558 endSeconds: 573 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=558' heldReason: rights - ordinal: 13 text: null startSeconds: 616 endSeconds: 627 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=616' heldReason: rights - ordinal: 14 text: null startSeconds: 627 endSeconds: 684 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=627' heldReason: rights - ordinal: 15 text: null startSeconds: 696 endSeconds: 753 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=696' heldReason: rights - ordinal: 16 text: null startSeconds: 765 endSeconds: 835 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=765' heldReason: rights - ordinal: 17 text: null startSeconds: 846 endSeconds: 914 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=846' heldReason: rights - ordinal: 18 text: null startSeconds: 926 endSeconds: 975 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=926' heldReason: rights - ordinal: 19 text: null startSeconds: 985 endSeconds: 1018 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=985' heldReason: rights - ordinal: 20 text: null startSeconds: 1030 endSeconds: 1065 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=1030' heldReason: rights - ordinal: 21 text: null startSeconds: 1075 endSeconds: 1178 watch: 'https://liquidgold.cc/watch/qflLT3SoVbw?t=1075' heldReason: rights --- # Stop Fine-Tuning to Fix Retrieval Problems — Anant Srivastava Most AI teams never decide where their knowledge lives. Six months of normal product work decides for them. Anant Srivastava, principal technologist for data and AI platforms at Oracle, argues that prompt, memory and weights aren't a ladder you climb when answers go wrong. They're three tools for three jobs. Through a support assistant that kept inventing product names after a fine-tune, he shows how architecture builds up by accident. He then gives a simple diagnostic for each: the prompt is for small, stable behavior; memory is for knowledge that's current, large, citable or access-controlled; weights are only for reflexes that have stopped changing. He closes with a circulating architecture where the agent gets better by doing its job. In this talk: • How every prompt edit, indexed doc and training sample is an architecture decision • The diagnostic question for prompt, memory and weights • Why you fine-tune reflexes, not facts • Moving knowledge between context, memory and weights over time SPEAKER Anant Srivastava, Principal Technologist, Data and AI Platforms, Oracle LinkedIn: https://www.linkedin.com/in/anantds/ CHAPTERS 0:00 Where should knowledge live? 0:15 The real engineering problem 0:50 Three tools, not a ladder 1:20 How teams decide by accident 2:00 Every edit is an architecture decision 2:40 A support assistant goes wrong 4:34 Accumulated, not designed 5:14 Prompt: behavior, not facts 5:59 Example: a support agent prompt 6:59 Memory: current, large, citable 8:09 The wrong job for memory 8:49 Access control belongs in memory 9:14 Example: a code assistant 10:08 Chunking and metadata for code RAG 11:23 Weights: what has stopped changing 12:28 The fine-tuning mistake 13:53 When fine-tuning works 15:08 Medical coding: fine-tune the reflex 16:13 Capability or cost? 16:58 The decision table 17:33 A circulating architecture 19:27 The model is the easy part Recorded at the AI Engineer World's Fair 2026 in San Francisco. Subscribe for more talks from the engineers building with AI. AI Engineer: https://ai.engineer YouTube: https://www.youtube.com/@aiDotEngineer X: https://x.com/aiDotEngineer LinkedIn: https://www.linkedin.com/company/aidotengineer/ #RAG #FineTuning #AIEngineer Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/stop-fine-tuning-to-fix-retrieval-problems-anant-sr-11b6d54c81d7 · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks --- #### Chunk 0: Document excerpt (0:13 to 1:22) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (1:22 to 2:43) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (2:43 to 4:46) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 3: Document excerpt (4:46 to 6:05) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 4: Document excerpt (6:05 to 7:01) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 5: Document excerpt (7:01 to 8:30) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 6: Document excerpt (8:30 to 9:18) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 7: Document excerpt (9:18 to 10:16) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 8: Document excerpt (10:16 to 11:24) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 9: Document excerpt (11:24 to 12:33) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 10: Document excerpt (12:33 to 13:55) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 11: Document excerpt (13:55 to 15:14) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 12: Document excerpt (15:14 to 16:15) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 13: Document excerpt (16:15 to 16:58) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 14: Document excerpt (16:58 to 17:45) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 15: Document excerpt (17:45 to 20:12) > Held: transcript rights do not permit this chunk's text.