{"version":2,"relationshipURLBase":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html","relationshipResolution":"Resolve relative relationship URLs against relationshipURLBase; complete records and absolute URLs are in blocks.","title":"匹配还可以变成图文检索","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html","markdown":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.md","pages":[{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","unit":"已知图文配对怎样变成可用的文字候选接口？","unitId":"clip-pairing-to-candidates","chapter":"III","number":1},{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","unit":"检索到底复用了什么，又省去了什么？","unitId":"retrieval-cache-and-evaluation","chapter":"III","number":2},{"id":"retrieval-metric-questions","title":"让每个检索指标回答自己的问题","unit":"检索到底复用了什么，又省去了什么？","unitId":"retrieval-cache-and-evaluation","chapter":"III","number":3},{"id":"retrieval-search-and-interaction","title":"选读：省搜索与增加图文交互分别改变什么","unit":"检索到底复用了什么，又省去了什么？","unitId":"retrieval-cache-and-evaluation","chapter":"III","number":4}],"aliases":[{"oldId":"explanation","targetId":"retrieval-reuse-encoding","operation":"split-explanation-range"},{"oldId":"b-a80ceeaaa6b1c404fd16dcce","targetId":"b-a80ceeaaa6b1c404fd16dcce","page":"clip-lab04-candidate-controls","operation":"move","beforeSHA256":"8ef459f480d8633165535c202fae7d876f4ff69a4a5b1505c6fdca66a6f48ffb","afterSHA256":"8ef459f480d8633165535c202fae7d876f4ff69a4a5b1505c6fdca66a6f48ffb","reason":"Same source body in authored semantic page"},{"oldId":"b-be090f9a5879761cebf4410a","targetId":"b-be090f9a5879761cebf4410a","page":"clip-lab04-candidate-controls","operation":"move","beforeSHA256":"08592cac5e85947afd40374d30d6030489704a2787422ccca7752470cb945f62","afterSHA256":"08592cac5e85947afd40374d30d6030489704a2787422ccca7752470cb945f62","reason":"Same source body in authored semantic page"},{"oldId":"b-d4a21190f02429031007c5a3","targetId":"b-d4a21190f02429031007c5a3","page":"clip-lab04-candidate-controls","operation":"move","beforeSHA256":"c7061a8ad0794772fc5fd4b5f0bdb6e66e098c0aaa2732e4c041ff8bb739810e","afterSHA256":"c7061a8ad0794772fc5fd4b5f0bdb6e66e098c0aaa2732e4c041ff8bb739810e","reason":"Same source body in authored semantic page"},{"oldId":"b-f93fec5c8f3e1a87b2bfd6cf","targetId":"b-f93fec5c8f3e1a87b2bfd6cf","page":"clip-lab04-candidate-controls","operation":"move","beforeSHA256":"1fb9bd54f6a6d811bd1b68919726fd1177e6a5264ddb1afa54180876084827ba","afterSHA256":"1fb9bd54f6a6d811bd1b68919726fd1177e6a5264ddb1afa54180876084827ba","reason":"Same source body in authored semantic page"},{"oldId":"b-8353d70533fe6b5fa5e69cd1","targetId":"b-8353d70533fe6b5fa5e69cd1","page":"clip-lab04-candidate-controls","operation":"move","beforeSHA256":"3fd55da4e643785a3ea3b5949b03d73e798ccf3ae1b5812997a45c182c9d7898","afterSHA256":"3fd55da4e643785a3ea3b5949b03d73e798ccf3ae1b5812997a45c182c9d7898","reason":"Same source body in authored semantic page"},{"oldId":"b-5eba727afa355c8a7b2f6b74","targetId":"b-5eba727afa355c8a7b2f6b74","page":"retrieval-reuse-encoding","operation":"move","beforeSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","afterSHA256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","reason":"Same source body in authored semantic page"},{"oldId":"b-b88d7043312de2f60feaaa04","targetId":"b-b88d7043312de2f60feaaa04","page":"retrieval-reuse-encoding","operation":"move","beforeSHA256":"b27dee396a8a35bd527c5e887f7ba085dba60a4f65ea2b692d86c22b0ae82b1b","afterSHA256":"b27dee396a8a35bd527c5e887f7ba085dba60a4f65ea2b692d86c22b0ae82b1b","reason":"Same source body in authored semantic page"},{"oldId":"b-7d727467abe508e04d558282","targetId":"b-7d727467abe508e04d558282","page":"retrieval-reuse-encoding","operation":"move","beforeSHA256":"f726d4dfd12721dd615d742d7afe75d70e549a32421c9fb296bd5365d8942054","afterSHA256":"f726d4dfd12721dd615d742d7afe75d70e549a32421c9fb296bd5365d8942054","reason":"Same source body in authored semantic page"},{"oldId":"b-04edaf888621515863711528","targetId":"b-04edaf888621515863711528","page":"retrieval-reuse-encoding","operation":"rewrite","beforeSHA256":"6afd5e4847ef02a2d444a78bb69ce6fd8abcef6c4ddc3133c1d5e69c6114485e","afterSHA256":"81ec72c83b991a733b7123ff02b7894e9891ec958450e90bc7828c1fed97f784","reason":"Same source identity with figure semantics or reading context enriched"},{"oldId":"b-dc3c47c9ff8463dbe975970f","targetId":"b-dc3c47c9ff8463dbe975970f","page":"retrieval-reuse-encoding","operation":"move","beforeSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","afterSHA256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","reason":"Same source body in authored semantic page"},{"oldId":"b-51ee8395325449e7fd8c6719","targetId":"b-51ee8395325449e7fd8c6719","page":"retrieval-reuse-encoding","operation":"move","beforeSHA256":"c838cd522b3b8bb0b056105565f1663b24283deeb732859aadf4cfa50d4a8a54","afterSHA256":"c838cd522b3b8bb0b056105565f1663b24283deeb732859aadf4cfa50d4a8a54","reason":"Same source body in authored semantic page"},{"oldId":"b-b7a85d47252d737cfb0277fb","targetId":"b-b7a85d47252d737cfb0277fb","page":"retrieval-reuse-encoding","operation":"move","beforeSHA256":"eca861f17392b57ec2be38794daf945be60f1ec3fc9d01c83eff205c5039ba89","afterSHA256":"eca861f17392b57ec2be38794daf945be60f1ec3fc9d01c83eff205c5039ba89","reason":"Same source body in authored semantic page"},{"oldId":"b-80907ac5f1234aede1d5d070","targetId":"b-80907ac5f1234aede1d5d070","page":"retrieval-reuse-encoding","operation":"move","beforeSHA256":"0d52d45b91d64683943b3870efe42eb2211cb249baaa5d700568f83f16ac824f","afterSHA256":"0d52d45b91d64683943b3870efe42eb2211cb249baaa5d700568f83f16ac824f","reason":"Same source body in authored semantic page"},{"oldId":"b-59c0250b7b94373443fe912d","targetId":"b-59c0250b7b94373443fe912d","page":"retrieval-reuse-encoding","operation":"move","beforeSHA256":"548903116868e99139fe6cb659aa140a85b965dcc46a1ad6a79904bd982e3beb","afterSHA256":"548903116868e99139fe6cb659aa140a85b965dcc46a1ad6a79904bd982e3beb","reason":"Same source body in authored semantic page"},{"oldId":"b-cccb07251450676e0446a9bc","targetId":"b-cccb07251450676e0446a9bc","page":"retrieval-reuse-encoding","operation":"move","beforeSHA256":"a0b9c4c8d9eebe422e20a0d9595195dc58efc43923716e264bd41dc0bfb60c5d","afterSHA256":"a0b9c4c8d9eebe422e20a0d9595195dc58efc43923716e264bd41dc0bfb60c5d","reason":"Same source body in authored semantic page"},{"oldId":"b-526b538610a36604cd9f8957","targetId":"b-526b538610a36604cd9f8957","page":"retrieval-reuse-encoding","operation":"move","beforeSHA256":"92827721d51ccba74677884e288a5b789ab5ded7a0520fdc0992941218d8ee81","afterSHA256":"92827721d51ccba74677884e288a5b789ab5ded7a0520fdc0992941218d8ee81","reason":"Same source body in authored semantic page"},{"oldId":"b-33fd9610782b2b8229cca2f3","targetId":"b-33fd9610782b2b8229cca2f3","page":"retrieval-reuse-encoding","operation":"move","beforeSHA256":"e06ae1708d414ac9f8f5b7ac04ac212f1b0939da9c7729b3878ec8b958a74b66","afterSHA256":"e06ae1708d414ac9f8f5b7ac04ac212f1b0939da9c7729b3878ec8b958a74b66","reason":"Same source body in authored semantic page"},{"oldId":"b-ad95390cfb85292761b64306","targetId":"b-ad95390cfb85292761b64306","page":"retrieval-reuse-encoding","operation":"move","beforeSHA256":"ac480cd10ae33ad33bc08b2666dd0b6adbf60adc29c82b8aa1cc91945d5eac3b","afterSHA256":"ac480cd10ae33ad33bc08b2666dd0b6adbf60adc29c82b8aa1cc91945d5eac3b","reason":"Same source body in authored semantic page"},{"oldId":"b-113676cae8a960111acc5927","targetId":"b-113676cae8a960111acc5927","page":"retrieval-reuse-encoding","operation":"move","beforeSHA256":"f19687a8a051cd6500e71ee01a8bc63867f5079ecca32db524f538b4a2c35153","afterSHA256":"f19687a8a051cd6500e71ee01a8bc63867f5079ecca32db524f538b4a2c35153","reason":"Same source body in authored semantic page"},{"oldId":"b-b2782b587cc8dd08dfe6a3db","targetId":"b-b2782b587cc8dd08dfe6a3db","page":"retrieval-reuse-encoding","operation":"move","beforeSHA256":"51500a1bc1ac7c5e8538e5aeea6fb4c8f32e8312edae075d520a40cf6d9eb707","afterSHA256":"51500a1bc1ac7c5e8538e5aeea6fb4c8f32e8312edae075d520a40cf6d9eb707","reason":"Same source body in authored semantic page"},{"oldId":"b-7a12fbe59a5e5835131be287","targetId":"b-7a12fbe59a5e5835131be287","page":"retrieval-reuse-encoding","operation":"move","beforeSHA256":"291861c96834ab4b94163ee8a3ce8ef3cc98d01bc52cdc2c4363af1cfc9b253b","afterSHA256":"291861c96834ab4b94163ee8a3ce8ef3cc98d01bc52cdc2c4363af1cfc9b253b","reason":"Same source body in authored semantic page"},{"oldId":"b-fb4d2e33c9b5cd0c39989ee7","targetId":"b-fb4d2e33c9b5cd0c39989ee7","page":"retrieval-reuse-encoding","operation":"move","beforeSHA256":"90c9bc3365b9cf34f307a6fd25c68d00a9215c3147c22e327f049f444c7caf86","afterSHA256":"90c9bc3365b9cf34f307a6fd25c68d00a9215c3147c22e327f049f444c7caf86","reason":"Same source body in authored semantic page"},{"oldId":"b-fe24cefdb9636fe783dd9392","targetId":"b-fe24cefdb9636fe783dd9392","page":"retrieval-reuse-encoding","operation":"move","beforeSHA256":"ec4455aa1077a8444959c6a0665d97b8fad7cf29b957a5e954edcf4bf7683d1d","afterSHA256":"ec4455aa1077a8444959c6a0665d97b8fad7cf29b957a5e954edcf4bf7683d1d","reason":"Same source body in authored semantic page"},{"oldId":"b-254dd3e8402500f81270c047","targetId":"b-254dd3e8402500f81270c047","page":"retrieval-metric-questions","operation":"move","beforeSHA256":"a32f625f24bd569879229f77efbce5dc69b1243961d1cc94f086b20af7fd2e60","afterSHA256":"a32f625f24bd569879229f77efbce5dc69b1243961d1cc94f086b20af7fd2e60","reason":"Same source body in authored semantic page"},{"oldId":"b-acc1d5d73f777002d50a8553","targetId":"b-acc1d5d73f777002d50a8553","page":"retrieval-metric-questions","operation":"move","beforeSHA256":"7d1e522bf167385f95073a108d6651706c7ab2f25d134521e17b40fe2992bc59","afterSHA256":"7d1e522bf167385f95073a108d6651706c7ab2f25d134521e17b40fe2992bc59","reason":"Same source body in authored semantic page"},{"oldId":"b-18fc10821fe6af84adf28108","targetId":"b-18fc10821fe6af84adf28108","page":"retrieval-metric-questions","operation":"move","beforeSHA256":"e60979fc29d0931ab0fa6565feb357a2b9d204702fac523f27809082dbf318a7","afterSHA256":"e60979fc29d0931ab0fa6565feb357a2b9d204702fac523f27809082dbf318a7","reason":"Same source body in authored semantic page"},{"oldId":"b-d283b5fafa4b4af876c178e7","targetId":"b-d283b5fafa4b4af876c178e7","page":"retrieval-metric-questions","operation":"move","beforeSHA256":"e6be9a71494e6e8c8f730b05c8896471180ddb04def85ade8acb5519b43a466f","afterSHA256":"e6be9a71494e6e8c8f730b05c8896471180ddb04def85ade8acb5519b43a466f","reason":"Same source body in authored semantic page"},{"oldId":"b-10ec0695f64ff08ec26356c6","targetId":"b-10ec0695f64ff08ec26356c6","page":"retrieval-metric-questions","operation":"move","beforeSHA256":"c1afda030628985fb13e6d5c22f9cfe7d7b0fe2cab59e7e3f191abb1dcef9feb","afterSHA256":"c1afda030628985fb13e6d5c22f9cfe7d7b0fe2cab59e7e3f191abb1dcef9feb","reason":"Same source body in authored semantic page"},{"oldId":"b-b7039f5e8869fce0a5ed309d","targetId":"b-b7039f5e8869fce0a5ed309d","page":"retrieval-search-and-interaction","operation":"move","beforeSHA256":"225cf2646d1bbc81fd3f8ed64c5eac6bf960b0c41b919b6f24554236d701a629","afterSHA256":"225cf2646d1bbc81fd3f8ed64c5eac6bf960b0c41b919b6f24554236d701a629","reason":"Same source body in authored semantic page"},{"oldId":"b-d4d365ef6131f1bec7a8a0ee","targetId":"b-d4d365ef6131f1bec7a8a0ee","page":"retrieval-search-and-interaction","operation":"move","beforeSHA256":"1766e359ec4d31dd1478ef54c245a79bf9e3cd2bcb748e45f6e24d363cc4b064","afterSHA256":"1766e359ec4d31dd1478ef54c245a79bf9e3cd2bcb748e45f6e24d363cc4b064","reason":"Same source body in authored semantic page"},{"oldId":"b-daa988ff5623cbdd887cd4df","targetId":"b-daa988ff5623cbdd887cd4df","page":"retrieval-search-and-interaction","operation":"move","beforeSHA256":"dfb749be9ea401298591c13840469fdf0010697e73c244978111ebf419a560a3","afterSHA256":"dfb749be9ea401298591c13840469fdf0010697e73c244978111ebf419a560a3","reason":"Same source body in authored semantic page"},{"oldId":"b-d083e84b19466f7c176c46d8","targetId":"b-d083e84b19466f7c176c46d8","page":"retrieval-search-and-interaction","operation":"move","beforeSHA256":"6424f8550f5b5c6628f1965c5c1b4aafa939764425130e58f1a53d555fcca12a","afterSHA256":"6424f8550f5b5c6628f1965c5c1b4aafa939764425130e58f1a53d555fcca12a","reason":"Same source body in authored semantic page"},{"oldId":"b-afe92d2fc8854ed5d7b12714","targetId":"b-afe92d2fc8854ed5d7b12714","page":"retrieval-search-and-interaction","operation":"move","beforeSHA256":"4317d93511e6997230f0f90631867ff02fa98c7c130893b0bfe4a4e45f7238d1","afterSHA256":"4317d93511e6997230f0f90631867ff02fa98c7c130893b0bfe4a4e45f7238d1","reason":"Same source body in authored semantic page"},{"oldId":"b-729752ad3d2427f4e9c7375d","targetId":"b-729752ad3d2427f4e9c7375d","page":"retrieval-search-and-interaction","operation":"move","beforeSHA256":"f16be0424762a44344338696f06ef266c4571d15367ae2cbbd28fac98ac7e9cb","afterSHA256":"f16be0424762a44344338696f06ef266c4571d15367ae2cbbd28fac98ac7e9cb","reason":"Same source body in authored semantic page"},{"oldId":"b-f6052c27e01d5aaa6d99537c","targetId":"b-f6052c27e01d5aaa6d99537c","page":"retrieval-search-and-interaction","operation":"move","beforeSHA256":"b25685f04cd1b885f3a5999e6a912460c7424021cdc5b2190584242086a82681","afterSHA256":"b25685f04cd1b885f3a5999e6a912460c7424021cdc5b2190584242086a82681","reason":"Same source body in authored semantic page"},{"oldId":"b-dfb9ff8ddaa8f556aee5d733","targetId":"b-dfb9ff8ddaa8f556aee5d733","page":"retrieval-search-and-interaction","operation":"move","beforeSHA256":"817e631e55b07e09213be086a884c4de66045705f64e23106ed9b30aebdf157d","afterSHA256":"817e631e55b07e09213be086a884c4de66045705f64e23106ed9b30aebdf157d","reason":"Same source body in authored semantic page"},{"oldId":"b-f1422c87e4d51f608a7df583","targetId":"b-f1422c87e4d51f608a7df583","page":"retrieval-search-and-interaction","operation":"move","beforeSHA256":"d51727108cd12db21b736dc34162c5db4a173ebea0e9fbb86f3d2ee40c0911b7","afterSHA256":"d51727108cd12db21b736dc34162c5db4a173ebea0e9fbb86f3d2ee40c0911b7","reason":"Same source body in authored semantic page"},{"oldId":"b-ab6c729c8025f6765e30696d","targetId":"b-ab6c729c8025f6765e30696d","page":"retrieval-search-and-interaction","operation":"move","beforeSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","afterSHA256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","reason":"Same source body in authored semantic page"},{"oldId":"b-f0df6a6a8d74c4d50ead2cd7","targetId":"b-f0df6a6a8d74c4d50ead2cd7","page":"retrieval-search-and-interaction","operation":"move","beforeSHA256":"0b7bf6c3e73b32d9d7f76b97606179fde1b53cca0b81d8e498a24853b0999b89","afterSHA256":"0b7bf6c3e73b32d9d7f76b97606179fde1b53cca0b81d8e498a24853b0999b89","reason":"Same source body in authored semantic page"},{"oldId":"clip-lab04-candidate-controls-title","targetId":"b-1bf4ca37f8f7e87ce45520f4","operation":"retain-source-identity"},{"oldId":"clip-lab04-candidate-controls-bridge","targetId":"b-66d7253ca68ee0adcdce7eae","operation":"retain-source-identity"},{"oldId":"clip-lab04-saved-figure","targetId":"b-f2280c99540cc27d10e8b19d","operation":"retain-source-identity"},{"oldId":"retrieval-reuse-encoding-title","targetId":"b-12fcaca6d722fb54e64bb3e2","operation":"retain-source-identity"},{"oldId":"retrieval-metric-questions-title","targetId":"b-826a8147b8b09509f27af15b","operation":"retain-source-identity"},{"oldId":"retrieval-metric-questions-bridge","targetId":"b-cbd572bc6c5fc2b76c156c2c","operation":"retain-source-identity"},{"oldId":"retrieval-search-and-interaction-title","targetId":"b-7002abd47beba8b7772509e3","operation":"retain-source-identity"},{"oldId":"retrieval-search-and-interaction-bridge","targetId":"b-d5a2b891222095ad5a946190","operation":"retain-source-identity"}],"blocks":[{"id":"b-1bf4ca37f8f7e87ce45520f4","type":"h2","scope":"clip-lab04-candidate-controls","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-1bf4ca37f8f7e87ce45520f4","content":"## 在真实猫图上核对变化的量","sha256":"d539bcc5fdead43ca3a9580eb18311ece3c7510d98ea03c9f9ee02373a0161cd","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":1},"sectionTitle":"在真实猫图上核对变化的量","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-66d7253ca68ee0adcdce7eae","prerequisites":[{"id":"b-66d7253ca68ee0adcdce7eae","url":"#b-66d7253ca68ee0adcdce7eae","role":"problem-context"}],"relatedBlocks":[{"id":"b-f2280c99540cc27d10e8b19d","url":"#b-f2280c99540cc27d10e8b19d","relation":"figure-and-explanation"}]},{"id":"b-66d7253ca68ee0adcdce7eae","type":"p","scope":"clip-lab04-candidate-controls","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-66d7253ca68ee0adcdce7eae","content":"教学算例已经预测：共享图文的余弦可保持，而候选集合内的相对分数改变。","sha256":"8bd7ef97e9461420d2cafa3f45217d624050f98a53a498b584a756d60b71cde4","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":1},"sectionTitle":"在真实猫图上核对变化的量","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-1bf4ca37f8f7e87ce45520f4","nextBlock":"b-f2280c99540cc27d10e8b19d","prerequisites":[{"id":"b-1bf4ca37f8f7e87ce45520f4","url":"#b-1bf4ca37f8f7e87ce45520f4","role":"section-title"}],"relatedBlocks":[{"id":"b-f2280c99540cc27d10e8b19d","url":"#b-f2280c99540cc27d10e8b19d","relation":"figure-and-explanation"}]},{"id":"b-f2280c99540cc27d10e8b19d","type":"figure","scope":"clip-lab04-candidate-controls","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-f2280c99540cc27d10e8b19d","content":"![Chelsea 猫图的已有 CLIP 四臂候选实验结果](/assets/display/fed4bc77235bae1509d5761b.webp)LAB04 已保存的真实猫图推理结果。A 为 cat/dog/horse；B 增加 tiger/fox；C 保留 A 的候选、改为 cropped 模板；D 从 B 去掉 cat。结果来自既有记录，本轮没有运行模型。","sha256":"04c5777d12b2c7948be6a306438962b1b6426b8b0d48b4f5720ebdd8fbf6f007","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":1},"sectionTitle":"在真实猫图上核对变化的量","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"LAB04 已保存的真实猫图推理结果。A 为 cat/dog/horse；B 增加 tiger/fox；C 保留 A 的候选、改为 cropped 模板；D 从 B 去掉 cat。结果来自既有记录，本轮没有运行模型。","labels":[],"description":"Chelsea 猫图的已有 CLIP 四臂候选实验结果","evidenceIdentity":"LAB04 已保存的真实猫图推理结果。A 为 cat/dog/horse；B 增加 tiger/fox；C 保留 A 的候选、改为 cropped 模板；D 从 B 去掉 cat。结果来自既有记录，本轮没有运行模型。","images":[{"src":"/assets/display/fed4bc77235bae1509d5761b.webp","alt":"Chelsea 猫图的已有 CLIP 四臂候选实验结果"}]},"previousBlock":"b-66d7253ca68ee0adcdce7eae","nextBlock":"b-a80ceeaaa6b1c404fd16dcce","prerequisites":[{"id":"b-1bf4ca37f8f7e87ce45520f4","url":"#b-1bf4ca37f8f7e87ce45520f4","role":"section-title"},{"id":"b-66d7253ca68ee0adcdce7eae","url":"#b-66d7253ca68ee0adcdce7eae","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-a80ceeaaa6b1c404fd16dcce","type":"h3","scope":"clip-real-checkpoint-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-a80ceeaaa6b1c404fd16dcce","content":"### 真实猫图结果：保留模型、照片与历史身份","sha256":"8ef459f480d8633165535c202fae7d876f4ff69a4a5b1505c6fdca66a6f48ffb","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":1},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":"clip-real-checkpoint-evidence","formulas":[],"previousBlock":"b-f2280c99540cc27d10e8b19d","nextBlock":"b-be090f9a5879761cebf4410a","prerequisites":[{"id":"b-1bf4ca37f8f7e87ce45520f4","url":"#b-1bf4ca37f8f7e87ce45520f4","role":"section-title"},{"id":"b-66d7253ca68ee0adcdce7eae","url":"#b-66d7253ca68ee0adcdce7eae","role":"problem-context"}],"relatedBlocks":[{"id":"b-f2280c99540cc27d10e8b19d","url":"#b-f2280c99540cc27d10e8b19d","relation":"figure-and-explanation"},{"id":"b-be090f9a5879761cebf4410a","url":"#b-be090f9a5879761cebf4410a","relation":"same-reasoning"},{"id":"b-d4a21190f02429031007c5a3","url":"#b-d4a21190f02429031007c5a3","relation":"same-reasoning"},{"id":"b-f93fec5c8f3e1a87b2bfd6cf","url":"#b-f93fec5c8f3e1a87b2bfd6cf","relation":"same-reasoning"},{"id":"b-8353d70533fe6b5fa5e69cd1","url":"#b-8353d70533fe6b5fa5e69cd1","relation":"same-reasoning"}]},{"id":"b-be090f9a5879761cebf4410a","type":"p","scope":"clip-real-checkpoint-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-be090f9a5879761cebf4410a","content":"当前 LAB04 保存真实 Chelsea 猫照片与 OpenAI CLIP ViT-B/32 实验。2026-09-22 的本地复核使用固定 checkpoint revision `3d74acf9a28c67741b2f4f2ea7635f0aaf6f0268`，没有重新训练 CLIP；当前整理也没有重跑该推理。图片先转 RGB、缩放并中心裁剪到 $224\\times224$，再按模型配置归一化。","sha256":"08592cac5e85947afd40374d30d6030489704a2787422ccca7752470cb945f62","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":1},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":"clip-real-checkpoint-evidence","formulas":[{"latex":"224\\times224","definitionContext":"当前 LAB04 保存真实 Chelsea 猫照片与 OpenAI CLIP ViT-B/32 实验。2026-09-22 的本地复核使用固定 checkpoint revision `3d74acf9a28c67741b2f4f2ea7635f0aaf6f0268`，没有重新训练 CLIP；当前整理也没有重跑该推理。图片先转 RGB、缩放并中心裁剪到 $224\\times224$，再按模型配置归一化。\n基础 cat/dog/horse 候选下，cat 余弦约 0.28691、候选 softmax 约 0.99577；添加 tiger 与 fox 后，共享 cat 的余弦保持不变，softmax 降至约 0.98195；再删除 cat，tiger 成为剩余候选的最高项。新旧猫图余弦最大差约 $1.04\\times10^{-7}$，把后来的复核与历史材料连接起来。\n正面结果是模型可以在这张真实照片上把图像与合理文字对应，并且其候选接口的变化符合预期。数值 0.99577 是给定候选集合内的相对分配，不是“现实中有猫的校准概率”。一张猫图也不能替代数据集级泛化评测。\nLAB04 负责模型输入、候选、模板和关系检查；LAB12 复用同一模型产生的完整 512 维向量，研究图库缓存与搜索。不要把四个候选分数拼成所谓图像表示；真正可复用的是编码器输出的向量。先在 LAB04 理解分数身份，再在 LAB12 看系统组织，就无需重复训练或建立两份同义实验。"}],"previousBlock":"b-a80ceeaaa6b1c404fd16dcce","nextBlock":"b-d4a21190f02429031007c5a3","prerequisites":[{"id":"b-a80ceeaaa6b1c404fd16dcce","url":"#b-a80ceeaaa6b1c404fd16dcce","role":"section-title"},{"id":"b-66d7253ca68ee0adcdce7eae","url":"#b-66d7253ca68ee0adcdce7eae","role":"problem-context"}],"relatedBlocks":[{"id":"b-f2280c99540cc27d10e8b19d","url":"#b-f2280c99540cc27d10e8b19d","relation":"figure-and-explanation"},{"id":"b-a80ceeaaa6b1c404fd16dcce","url":"#b-a80ceeaaa6b1c404fd16dcce","relation":"same-reasoning"},{"id":"b-d4a21190f02429031007c5a3","url":"#b-d4a21190f02429031007c5a3","relation":"same-reasoning"},{"id":"b-f93fec5c8f3e1a87b2bfd6cf","url":"#b-f93fec5c8f3e1a87b2bfd6cf","relation":"same-reasoning"},{"id":"b-8353d70533fe6b5fa5e69cd1","url":"#b-8353d70533fe6b5fa5e69cd1","relation":"same-reasoning"}]},{"id":"b-d4a21190f02429031007c5a3","type":"p","scope":"clip-real-checkpoint-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-d4a21190f02429031007c5a3","content":"基础 cat/dog/horse 候选下，cat 余弦约 0.28691、候选 softmax 约 0.99577；添加 tiger 与 fox 后，共享 cat 的余弦保持不变，softmax 降至约 0.98195；再删除 cat，tiger 成为剩余候选的最高项。新旧猫图余弦最大差约 $1.04\\times10^{-7}$，把后来的复核与历史材料连接起来。","sha256":"c7061a8ad0794772fc5fd4b5f0bdb6e66e098c0aaa2732e4c041ff8bb739810e","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":1},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":"clip-real-checkpoint-evidence","formulas":[{"latex":"1.04\\times10^{-7}","definitionContext":"当前 LAB04 保存真实 Chelsea 猫照片与 OpenAI CLIP ViT-B/32 实验。2026-09-22 的本地复核使用固定 checkpoint revision `3d74acf9a28c67741b2f4f2ea7635f0aaf6f0268`，没有重新训练 CLIP；当前整理也没有重跑该推理。图片先转 RGB、缩放并中心裁剪到 $224\\times224$，再按模型配置归一化。\n基础 cat/dog/horse 候选下，cat 余弦约 0.28691、候选 softmax 约 0.99577；添加 tiger 与 fox 后，共享 cat 的余弦保持不变，softmax 降至约 0.98195；再删除 cat，tiger 成为剩余候选的最高项。新旧猫图余弦最大差约 $1.04\\times10^{-7}$，把后来的复核与历史材料连接起来。\n正面结果是模型可以在这张真实照片上把图像与合理文字对应，并且其候选接口的变化符合预期。数值 0.99577 是给定候选集合内的相对分配，不是“现实中有猫的校准概率”。一张猫图也不能替代数据集级泛化评测。\nLAB04 负责模型输入、候选、模板和关系检查；LAB12 复用同一模型产生的完整 512 维向量，研究图库缓存与搜索。不要把四个候选分数拼成所谓图像表示；真正可复用的是编码器输出的向量。先在 LAB04 理解分数身份，再在 LAB12 看系统组织，就无需重复训练或建立两份同义实验。"}],"previousBlock":"b-be090f9a5879761cebf4410a","nextBlock":"b-f93fec5c8f3e1a87b2bfd6cf","prerequisites":[{"id":"b-a80ceeaaa6b1c404fd16dcce","url":"#b-a80ceeaaa6b1c404fd16dcce","role":"section-title"},{"id":"b-66d7253ca68ee0adcdce7eae","url":"#b-66d7253ca68ee0adcdce7eae","role":"problem-context"}],"relatedBlocks":[{"id":"b-f2280c99540cc27d10e8b19d","url":"#b-f2280c99540cc27d10e8b19d","relation":"figure-and-explanation"},{"id":"b-a80ceeaaa6b1c404fd16dcce","url":"#b-a80ceeaaa6b1c404fd16dcce","relation":"same-reasoning"},{"id":"b-be090f9a5879761cebf4410a","url":"#b-be090f9a5879761cebf4410a","relation":"same-reasoning"},{"id":"b-f93fec5c8f3e1a87b2bfd6cf","url":"#b-f93fec5c8f3e1a87b2bfd6cf","relation":"same-reasoning"},{"id":"b-8353d70533fe6b5fa5e69cd1","url":"#b-8353d70533fe6b5fa5e69cd1","relation":"same-reasoning"}]},{"id":"b-f93fec5c8f3e1a87b2bfd6cf","type":"p","scope":"clip-real-checkpoint-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-f93fec5c8f3e1a87b2bfd6cf","content":"正面结果是模型可以在这张真实照片上把图像与合理文字对应，并且其候选接口的变化符合预期。数值 0.99577 是给定候选集合内的相对分配，不是“现实中有猫的校准概率”。一张猫图也不能替代数据集级泛化评测。","sha256":"1fb9bd54f6a6d811bd1b68919726fd1177e6a5264ddb1afa54180876084827ba","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":1},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":"clip-real-checkpoint-evidence","formulas":[],"previousBlock":"b-d4a21190f02429031007c5a3","nextBlock":"b-8353d70533fe6b5fa5e69cd1","prerequisites":[{"id":"b-a80ceeaaa6b1c404fd16dcce","url":"#b-a80ceeaaa6b1c404fd16dcce","role":"section-title"},{"id":"b-66d7253ca68ee0adcdce7eae","url":"#b-66d7253ca68ee0adcdce7eae","role":"problem-context"}],"relatedBlocks":[{"id":"b-f2280c99540cc27d10e8b19d","url":"#b-f2280c99540cc27d10e8b19d","relation":"figure-and-explanation"},{"id":"b-a80ceeaaa6b1c404fd16dcce","url":"#b-a80ceeaaa6b1c404fd16dcce","relation":"same-reasoning"},{"id":"b-be090f9a5879761cebf4410a","url":"#b-be090f9a5879761cebf4410a","relation":"same-reasoning"},{"id":"b-d4a21190f02429031007c5a3","url":"#b-d4a21190f02429031007c5a3","relation":"same-reasoning"},{"id":"b-8353d70533fe6b5fa5e69cd1","url":"#b-8353d70533fe6b5fa5e69cd1","relation":"same-reasoning"}]},{"id":"b-8353d70533fe6b5fa5e69cd1","type":"p","scope":"clip-real-checkpoint-evidence","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-8353d70533fe6b5fa5e69cd1","content":"LAB04 负责模型输入、候选、模板和关系检查；LAB12 复用同一模型产生的完整 512 维向量，研究图库缓存与搜索。不要把四个候选分数拼成所谓图像表示；真正可复用的是编码器输出的向量。先在 LAB04 理解分数身份，再在 LAB12 看系统组织，就无需重复训练或建立两份同义实验。","sha256":"3fd55da4e643785a3ea3b5949b03d73e798ccf3ae1b5812997a45c182c9d7898","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"clip-pairing-to-candidates","title":"已知图文配对怎样变成可用的文字候选接口？"},"page":{"id":"clip-lab04-candidate-controls","title":"在真实猫图上核对变化的量","number":1},"sectionTitle":"真实猫图结果：保留模型、照片与历史身份","currentQuestion":"从一对输入到双向损失，再到真实候选干预，哪些量由训练决定、哪些由使用者决定？","sourceSlides":["l3-r2-p58","l3-r2-p59"],"knowledgeBlock":"clip-real-checkpoint-evidence","formulas":[],"previousBlock":"b-f93fec5c8f3e1a87b2bfd6cf","nextBlock":null,"prerequisites":[{"id":"b-a80ceeaaa6b1c404fd16dcce","url":"#b-a80ceeaaa6b1c404fd16dcce","role":"section-title"},{"id":"b-66d7253ca68ee0adcdce7eae","url":"#b-66d7253ca68ee0adcdce7eae","role":"problem-context"}],"relatedBlocks":[{"id":"b-f2280c99540cc27d10e8b19d","url":"#b-f2280c99540cc27d10e8b19d","relation":"figure-and-explanation"},{"id":"b-a80ceeaaa6b1c404fd16dcce","url":"#b-a80ceeaaa6b1c404fd16dcce","relation":"same-reasoning"},{"id":"b-be090f9a5879761cebf4410a","url":"#b-be090f9a5879761cebf4410a","relation":"same-reasoning"},{"id":"b-d4a21190f02429031007c5a3","url":"#b-d4a21190f02429031007c5a3","relation":"same-reasoning"},{"id":"b-f93fec5c8f3e1a87b2bfd6cf","url":"#b-f93fec5c8f3e1a87b2bfd6cf","relation":"same-reasoning"}]},{"id":"b-12fcaca6d722fb54e64bb3e2","type":"h2","scope":"retrieval-reuse-encoding","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-12fcaca6d722fb54e64bb3e2","content":"## 先编码图库，再让新查询比较","sha256":"4a5f24eae2039467a71b88a7f6fd7ece422508f1e1c5ce3a27fde827943f0323","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"先编码图库，再让新查询比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-5eba727afa355c8a7b2f6b74","prerequisites":[{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"}]},{"id":"b-5eba727afa355c8a7b2f6b74","type":"h2","scope":"explanation","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-5eba727afa355c8a7b2f6b74","content":"## 解释、算例与证据","sha256":"d71cf15e707c8a3674b4c90a599efee783cb84da5b8c8236a1d01c3add7b4958","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"先编码图库，再让新查询比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-12fcaca6d722fb54e64bb3e2","nextBlock":"b-b88d7043312de2f60feaaa04","prerequisites":[{"id":"b-12fcaca6d722fb54e64bb3e2","url":"#b-12fcaca6d722fb54e64bb3e2","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"}]},{"id":"b-b88d7043312de2f60feaaa04","type":"h1","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-b88d7043312de2f60feaaa04","content":"# P59 · 匹配还可以变成图文检索","sha256":"b27dee396a8a35bd527c5e887f7ba085dba60a4f65ea2b692d86c22b0ae82b1b","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"先编码图库，再让新查询比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-5eba727afa355c8a7b2f6b74","nextBlock":"b-7d727467abe508e04d558282","prerequisites":[{"id":"b-5eba727afa355c8a7b2f6b74","url":"#b-5eba727afa355c8a7b2f6b74","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"}]},{"id":"b-7d727467abe508e04d558282","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-7d727467abe508e04d558282","content":"检索把同一个比较空间用于一组已有图片。图库向量可复用，查询只需编码一次再搜索；页面前三张结果来自已有真实缓存。工作量、相关性指标和近似索引在说明与 LAB12 中展开；检索系统的进一步组织留待第八讲。","sha256":"f726d4dfd12721dd615d742d7afe75d70e549a32421c9fb296bd5365d8942054","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"先编码图库，再让新查询比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b88d7043312de2f60feaaa04","nextBlock":"b-04edaf888621515863711528","prerequisites":[{"id":"b-b88d7043312de2f60feaaa04","url":"#b-b88d7043312de2f60feaaa04","role":"section-title"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"}]},{"id":"b-04edaf888621515863711528","type":"figure","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-04edaf888621515863711528","content":"图内标签与数值：6 张来源照片 × 4 个相关视图 = 24 项；查询：“a photo of a cat.”；文字编码一次；与 24 个已保存图像向量比较；1   chelsea-center90；2   chelsea-full；3   chelsea-right90；已有结果：精确检索的前三项都是猫图变体；复用 2026-09-22 本地真实 CLIP 缓存；24 项含相关裁剪，不是 24 个独立来源。本轮未运行模型。\n\n已有六来源照片的二十四项库，接收猫图文字查询；展示既有真实缓存的精确排序前三项。\n\n- 6 张来源照片 × 4 个相关视图 = 24 项\n- 查询：“a photo of a cat.”\n- 文字编码一次\n- 与 24 个已保存图像向量比较\n- 1 chelsea-center90\n- 2 chelsea-full\n- 3 chelsea-right90\n- 已有结果：精确检索的前三项都是猫图变体\n- 复用 2026-09-22 本地真实 CLIP 缓存；24 项含相关裁剪，不是 24 个独立来源。本轮未运行模型。已有六来源照片的二十四项库，接收猫图文字查询；展示既有真实缓存的精确排序前三项。 [查看对应图解状态](/#/l3-r2-p59/0)[查看完整图解](/course/figures/l3-r2-p59-0.html)","sha256":"81ec72c83b991a733b7123ff02b7894e9891ec958450e90bc7828c1fed97f784","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"先编码图库，再让新查询比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"figure":{"caption":"已有六来源照片的二十四项库，接收猫图文字查询；展示既有真实缓存的精确排序前三项。 查看对应图解状态","labels":["6 张来源照片 × 4 个相关视图 = 24 项","查询：“a photo of a cat.”","文字编码一次","与 24 个已保存图像向量比较","1   chelsea-center90","2   chelsea-full","3   chelsea-right90","已有结果：精确检索的前三项都是猫图变体","复用 2026-09-22 本地真实 CLIP 缓存；24 项含相关裁剪，不是 24 个独立来源。本轮未运行模型。"],"description":"已有六来源照片的二十四项库，接收猫图文字查询；展示既有真实缓存的精确排序前三项。6 张来源照片 × 4 个相关视图 = 24 项查询：“a photo of a cat.”文字编码一次与 24 个已保存图像向量比较1 chelsea-center902 chelsea-full3 chelsea-right90已有结果：精确检索的前三项都是猫图变体复用 2026-09-22 本地真实 CLIP 缓存；24 项含相关裁剪，不是 24 个独立来源。本轮未运行模型。","evidenceIdentity":"已有六来源照片的二十四项库，接收猫图文字查询；展示既有真实缓存的精确排序前三项。 查看对应图解状态","images":[{"src":"/assets/display/61b4b65c7be015d923cad24b.webp","alt":""}]},"previousBlock":"b-7d727467abe508e04d558282","nextBlock":"b-dc3c47c9ff8463dbe975970f","prerequisites":[{"id":"b-b88d7043312de2f60feaaa04","url":"#b-b88d7043312de2f60feaaa04","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-dc3c47c9ff8463dbe975970f","type":"h2","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-dc3c47c9ff8463dbe975970f","content":"## 课堂讲述","sha256":"786ce10a6b32a492259a3dccd0a54e61cd41d28a34382fb7b5386c5f03b7ad69","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"先编码图库，再让新查询比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-04edaf888621515863711528","nextBlock":"b-51ee8395325449e7fd8c6719","prerequisites":[{"id":"b-b88d7043312de2f60feaaa04","url":"#b-b88d7043312de2f60feaaa04","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"}]},{"id":"b-51ee8395325449e7fd8c6719","type":"p","scope":"narrative","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-51ee8395325449e7fd8c6719","content":"同一对编码器还可以用于检索。图片库先得到表示，新的文字查询来了，再比较相似度并返回相关图片。反过来，也可以用图片查找描述。\n这里和第二部分有一个计算思想上的呼应：把可以复用的计算先完成，再让不同查询使用结果。但这只是计算组织的类比，不是说检测研究直接导致了向量检索。\n分类候选和检索库也不能用同一次小演示混为一谈。检索需要一组查询和明确的相关结果来评价。今天看到任务接口即可，近似索引、重排序和完整检索工程留到后面。这样既把 CLIP 能做什么讲清，又不把第三讲变成第八讲的提前版。","sha256":"c838cd522b3b8bb0b056105565f1663b24283deeb732859aadf4cfa50d4a8a54","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"先编码图库，再让新查询比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-dc3c47c9ff8463dbe975970f","nextBlock":"b-b7a85d47252d737cfb0277fb","prerequisites":[{"id":"b-dc3c47c9ff8463dbe975970f","url":"#b-dc3c47c9ff8463dbe975970f","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"}]},{"id":"b-b7a85d47252d737cfb0277fb","type":"h3","scope":"retrieval-cache-interface","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-b7a85d47252d737cfb0277fb","content":"### 检索为什么能缓存：每张图编码一次，查询再比较","sha256":"eca861f17392b57ec2be38794daf945be60f1ec3fc9d01c83eff205c5039ba89","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"检索为什么能缓存：每张图编码一次，查询再比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":"retrieval-cache-interface","formulas":[],"previousBlock":"b-51ee8395325449e7fd8c6719","nextBlock":"b-80907ac5f1234aede1d5d070","prerequisites":[{"id":"b-dc3c47c9ff8463dbe975970f","url":"#b-dc3c47c9ff8463dbe975970f","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"},{"id":"b-80907ac5f1234aede1d5d070","url":"#b-80907ac5f1234aede1d5d070","relation":"same-reasoning"},{"id":"b-59c0250b7b94373443fe912d","url":"#b-59c0250b7b94373443fe912d","relation":"same-reasoning"},{"id":"b-cccb07251450676e0446a9bc","url":"#b-cccb07251450676e0446a9bc","relation":"same-reasoning"},{"id":"b-526b538610a36604cd9f8957","url":"#b-526b538610a36604cd9f8957","relation":"same-reasoning"}]},{"id":"b-80907ac5f1234aede1d5d070","type":"p","scope":"retrieval-cache-interface","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-80907ac5f1234aede1d5d070","content":"设图库有 $N$ 张图，收到 $Q$ 条文字查询。双编码器的图像输出不依赖当前查询，所以可以先计算图库矩阵 $V\\in\\mathbb R^{N\\times d}$，查询到来时只计算文字向量 $u$，再用 $Vu$ 得分。这把“理解输入的深度编码”和“在现有表示中比较”分成两段。","sha256":"0d52d45b91d64683943b3870efe42eb2211cb249baaa5d700568f83f16ac824f","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"检索为什么能缓存：每张图编码一次，查询再比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":"retrieval-cache-interface","formulas":[{"latex":"N","definitionContext":"设图库有 $N$ 张图，收到 $Q$ 条文字查询。双编码器的图像输出不依赖当前查询，所以可以先计算图库矩阵 $V\\in\\mathbb R^{N\\times d}$，查询到来时只计算文字向量 $u$，再用 $Vu$ 得分。这把“理解输入的深度编码”和“在现有表示中比较”分成两段。\n若每个图文对都重新分别编码，需 $NQ$ 次图片样本编码和 $NQ$ 次文字样本编码；有缓存则是 $N+Q$。本课程 $N=24,Q=12$，两者为 576 与 36 次样本编码。它是解析工作量，不是秒数；一个批次调用可以同时编码多个样本。\n缓存并没有消除搜索本身。exact 查询仍要与全部 $N$ 向量做点积；float32 图库矩阵占 $4Nd$ 字节。本例 $24\\times512\\times4=49,152$ 字节，不含图片、模型、ID 与索引。改变 checkpoint、预处理或图片内容需要重建对应缓存；只新增文字候选不改变旧图片向量。\nP59 使用已有 24 项缓存显示猫查询的前三项：`chelsea-center90`、`chelsea-full`、`chelsea-right90`。24 项其实来自六张照片的四种相关裁剪，不能当成 24 张独立语义样本。它足以观察排名与缓存复用；更大的索引工程属于第八讲和 LAB12 的选读部分。"},{"latex":"Q","definitionContext":"设图库有 $N$ 张图，收到 $Q$ 条文字查询。双编码器的图像输出不依赖当前查询，所以可以先计算图库矩阵 $V\\in\\mathbb R^{N\\times d}$，查询到来时只计算文字向量 $u$，再用 $Vu$ 得分。这把“理解输入的深度编码”和“在现有表示中比较”分成两段。\n若每个图文对都重新分别编码，需 $NQ$ 次图片样本编码和 $NQ$ 次文字样本编码；有缓存则是 $N+Q$。本课程 $N=24,Q=12$，两者为 576 与 36 次样本编码。它是解析工作量，不是秒数；一个批次调用可以同时编码多个样本。\n缓存并没有消除搜索本身。exact 查询仍要与全部 $N$ 向量做点积；float32 图库矩阵占 $4Nd$ 字节。本例 $24\\times512\\times4=49,152$ 字节，不含图片、模型、ID 与索引。改变 checkpoint、预处理或图片内容需要重建对应缓存；只新增文字候选不改变旧图片向量。\nP59 使用已有 24 项缓存显示猫查询的前三项：`chelsea-center90`、`chelsea-full`、`chelsea-right90`。24 项其实来自六张照片的四种相关裁剪，不能当成 24 张独立语义样本。它足以观察排名与缓存复用；更大的索引工程属于第八讲和 LAB12 的选读部分。"},{"latex":"V\\in\\mathbb R^{N\\times d}","definitionContext":"设图库有 $N$ 张图，收到 $Q$ 条文字查询。双编码器的图像输出不依赖当前查询，所以可以先计算图库矩阵 $V\\in\\mathbb R^{N\\times d}$，查询到来时只计算文字向量 $u$，再用 $Vu$ 得分。这把“理解输入的深度编码”和“在现有表示中比较”分成两段。\n若每个图文对都重新分别编码，需 $NQ$ 次图片样本编码和 $NQ$ 次文字样本编码；有缓存则是 $N+Q$。本课程 $N=24,Q=12$，两者为 576 与 36 次样本编码。它是解析工作量，不是秒数；一个批次调用可以同时编码多个样本。\n缓存并没有消除搜索本身。exact 查询仍要与全部 $N$ 向量做点积；float32 图库矩阵占 $4Nd$ 字节。本例 $24\\times512\\times4=49,152$ 字节，不含图片、模型、ID 与索引。改变 checkpoint、预处理或图片内容需要重建对应缓存；只新增文字候选不改变旧图片向量。\nP59 使用已有 24 项缓存显示猫查询的前三项：`chelsea-center90`、`chelsea-full`、`chelsea-right90`。24 项其实来自六张照片的四种相关裁剪，不能当成 24 张独立语义样本。它足以观察排名与缓存复用；更大的索引工程属于第八讲和 LAB12 的选读部分。"},{"latex":"u","definitionContext":"设图库有 $N$ 张图，收到 $Q$ 条文字查询。双编码器的图像输出不依赖当前查询，所以可以先计算图库矩阵 $V\\in\\mathbb R^{N\\times d}$，查询到来时只计算文字向量 $u$，再用 $Vu$ 得分。这把“理解输入的深度编码”和“在现有表示中比较”分成两段。\n若每个图文对都重新分别编码，需 $NQ$ 次图片样本编码和 $NQ$ 次文字样本编码；有缓存则是 $N+Q$。本课程 $N=24,Q=12$，两者为 576 与 36 次样本编码。它是解析工作量，不是秒数；一个批次调用可以同时编码多个样本。\n缓存并没有消除搜索本身。exact 查询仍要与全部 $N$ 向量做点积；float32 图库矩阵占 $4Nd$ 字节。本例 $24\\times512\\times4=49,152$ 字节，不含图片、模型、ID 与索引。改变 checkpoint、预处理或图片内容需要重建对应缓存；只新增文字候选不改变旧图片向量。\nP59 使用已有 24 项缓存显示猫查询的前三项：`chelsea-center90`、`chelsea-full`、`chelsea-right90`。24 项其实来自六张照片的四种相关裁剪，不能当成 24 张独立语义样本。它足以观察排名与缓存复用；更大的索引工程属于第八讲和 LAB12 的选读部分。"},{"latex":"Vu","definitionContext":"设图库有 $N$ 张图，收到 $Q$ 条文字查询。双编码器的图像输出不依赖当前查询，所以可以先计算图库矩阵 $V\\in\\mathbb R^{N\\times d}$，查询到来时只计算文字向量 $u$，再用 $Vu$ 得分。这把“理解输入的深度编码”和“在现有表示中比较”分成两段。\n若每个图文对都重新分别编码，需 $NQ$ 次图片样本编码和 $NQ$ 次文字样本编码；有缓存则是 $N+Q$。本课程 $N=24,Q=12$，两者为 576 与 36 次样本编码。它是解析工作量，不是秒数；一个批次调用可以同时编码多个样本。\n缓存并没有消除搜索本身。exact 查询仍要与全部 $N$ 向量做点积；float32 图库矩阵占 $4Nd$ 字节。本例 $24\\times512\\times4=49,152$ 字节，不含图片、模型、ID 与索引。改变 checkpoint、预处理或图片内容需要重建对应缓存；只新增文字候选不改变旧图片向量。\nP59 使用已有 24 项缓存显示猫查询的前三项：`chelsea-center90`、`chelsea-full`、`chelsea-right90`。24 项其实来自六张照片的四种相关裁剪，不能当成 24 张独立语义样本。它足以观察排名与缓存复用；更大的索引工程属于第八讲和 LAB12 的选读部分。"}],"previousBlock":"b-b7a85d47252d737cfb0277fb","nextBlock":"b-59c0250b7b94373443fe912d","prerequisites":[{"id":"b-b7a85d47252d737cfb0277fb","url":"#b-b7a85d47252d737cfb0277fb","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"},{"id":"b-b7a85d47252d737cfb0277fb","url":"#b-b7a85d47252d737cfb0277fb","relation":"same-reasoning"},{"id":"b-59c0250b7b94373443fe912d","url":"#b-59c0250b7b94373443fe912d","relation":"same-reasoning"},{"id":"b-cccb07251450676e0446a9bc","url":"#b-cccb07251450676e0446a9bc","relation":"same-reasoning"},{"id":"b-526b538610a36604cd9f8957","url":"#b-526b538610a36604cd9f8957","relation":"same-reasoning"}]},{"id":"b-59c0250b7b94373443fe912d","type":"p","scope":"retrieval-cache-interface","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-59c0250b7b94373443fe912d","content":"若每个图文对都重新分别编码，需 $NQ$ 次图片样本编码和 $NQ$ 次文字样本编码；有缓存则是 $N+Q$。本课程 $N=24,Q=12$，两者为 576 与 36 次样本编码。它是解析工作量，不是秒数；一个批次调用可以同时编码多个样本。","sha256":"548903116868e99139fe6cb659aa140a85b965dcc46a1ad6a79904bd982e3beb","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"检索为什么能缓存：每张图编码一次，查询再比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":"retrieval-cache-interface","formulas":[{"latex":"NQ","definitionContext":"设图库有 $N$ 张图，收到 $Q$ 条文字查询。双编码器的图像输出不依赖当前查询，所以可以先计算图库矩阵 $V\\in\\mathbb R^{N\\times d}$，查询到来时只计算文字向量 $u$，再用 $Vu$ 得分。这把“理解输入的深度编码”和“在现有表示中比较”分成两段。\n若每个图文对都重新分别编码，需 $NQ$ 次图片样本编码和 $NQ$ 次文字样本编码；有缓存则是 $N+Q$。本课程 $N=24,Q=12$，两者为 576 与 36 次样本编码。它是解析工作量，不是秒数；一个批次调用可以同时编码多个样本。\n缓存并没有消除搜索本身。exact 查询仍要与全部 $N$ 向量做点积；float32 图库矩阵占 $4Nd$ 字节。本例 $24\\times512\\times4=49,152$ 字节，不含图片、模型、ID 与索引。改变 checkpoint、预处理或图片内容需要重建对应缓存；只新增文字候选不改变旧图片向量。\nP59 使用已有 24 项缓存显示猫查询的前三项：`chelsea-center90`、`chelsea-full`、`chelsea-right90`。24 项其实来自六张照片的四种相关裁剪，不能当成 24 张独立语义样本。它足以观察排名与缓存复用；更大的索引工程属于第八讲和 LAB12 的选读部分。"},{"latex":"NQ","definitionContext":"设图库有 $N$ 张图，收到 $Q$ 条文字查询。双编码器的图像输出不依赖当前查询，所以可以先计算图库矩阵 $V\\in\\mathbb R^{N\\times d}$，查询到来时只计算文字向量 $u$，再用 $Vu$ 得分。这把“理解输入的深度编码”和“在现有表示中比较”分成两段。\n若每个图文对都重新分别编码，需 $NQ$ 次图片样本编码和 $NQ$ 次文字样本编码；有缓存则是 $N+Q$。本课程 $N=24,Q=12$，两者为 576 与 36 次样本编码。它是解析工作量，不是秒数；一个批次调用可以同时编码多个样本。\n缓存并没有消除搜索本身。exact 查询仍要与全部 $N$ 向量做点积；float32 图库矩阵占 $4Nd$ 字节。本例 $24\\times512\\times4=49,152$ 字节，不含图片、模型、ID 与索引。改变 checkpoint、预处理或图片内容需要重建对应缓存；只新增文字候选不改变旧图片向量。\nP59 使用已有 24 项缓存显示猫查询的前三项：`chelsea-center90`、`chelsea-full`、`chelsea-right90`。24 项其实来自六张照片的四种相关裁剪，不能当成 24 张独立语义样本。它足以观察排名与缓存复用；更大的索引工程属于第八讲和 LAB12 的选读部分。"},{"latex":"N+Q","definitionContext":"设图库有 $N$ 张图，收到 $Q$ 条文字查询。双编码器的图像输出不依赖当前查询，所以可以先计算图库矩阵 $V\\in\\mathbb R^{N\\times d}$，查询到来时只计算文字向量 $u$，再用 $Vu$ 得分。这把“理解输入的深度编码”和“在现有表示中比较”分成两段。\n若每个图文对都重新分别编码，需 $NQ$ 次图片样本编码和 $NQ$ 次文字样本编码；有缓存则是 $N+Q$。本课程 $N=24,Q=12$，两者为 576 与 36 次样本编码。它是解析工作量，不是秒数；一个批次调用可以同时编码多个样本。\n缓存并没有消除搜索本身。exact 查询仍要与全部 $N$ 向量做点积；float32 图库矩阵占 $4Nd$ 字节。本例 $24\\times512\\times4=49,152$ 字节，不含图片、模型、ID 与索引。改变 checkpoint、预处理或图片内容需要重建对应缓存；只新增文字候选不改变旧图片向量。\nP59 使用已有 24 项缓存显示猫查询的前三项：`chelsea-center90`、`chelsea-full`、`chelsea-right90`。24 项其实来自六张照片的四种相关裁剪，不能当成 24 张独立语义样本。它足以观察排名与缓存复用；更大的索引工程属于第八讲和 LAB12 的选读部分。"},{"latex":"N=24,Q=12","definitionContext":"设图库有 $N$ 张图，收到 $Q$ 条文字查询。双编码器的图像输出不依赖当前查询，所以可以先计算图库矩阵 $V\\in\\mathbb R^{N\\times d}$，查询到来时只计算文字向量 $u$，再用 $Vu$ 得分。这把“理解输入的深度编码”和“在现有表示中比较”分成两段。\n若每个图文对都重新分别编码，需 $NQ$ 次图片样本编码和 $NQ$ 次文字样本编码；有缓存则是 $N+Q$。本课程 $N=24,Q=12$，两者为 576 与 36 次样本编码。它是解析工作量，不是秒数；一个批次调用可以同时编码多个样本。\n缓存并没有消除搜索本身。exact 查询仍要与全部 $N$ 向量做点积；float32 图库矩阵占 $4Nd$ 字节。本例 $24\\times512\\times4=49,152$ 字节，不含图片、模型、ID 与索引。改变 checkpoint、预处理或图片内容需要重建对应缓存；只新增文字候选不改变旧图片向量。\nP59 使用已有 24 项缓存显示猫查询的前三项：`chelsea-center90`、`chelsea-full`、`chelsea-right90`。24 项其实来自六张照片的四种相关裁剪，不能当成 24 张独立语义样本。它足以观察排名与缓存复用；更大的索引工程属于第八讲和 LAB12 的选读部分。"}],"previousBlock":"b-80907ac5f1234aede1d5d070","nextBlock":"b-cccb07251450676e0446a9bc","prerequisites":[{"id":"b-b7a85d47252d737cfb0277fb","url":"#b-b7a85d47252d737cfb0277fb","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"},{"id":"b-b7a85d47252d737cfb0277fb","url":"#b-b7a85d47252d737cfb0277fb","relation":"same-reasoning"},{"id":"b-80907ac5f1234aede1d5d070","url":"#b-80907ac5f1234aede1d5d070","relation":"same-reasoning"},{"id":"b-cccb07251450676e0446a9bc","url":"#b-cccb07251450676e0446a9bc","relation":"same-reasoning"},{"id":"b-526b538610a36604cd9f8957","url":"#b-526b538610a36604cd9f8957","relation":"same-reasoning"}]},{"id":"b-cccb07251450676e0446a9bc","type":"p","scope":"retrieval-cache-interface","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-cccb07251450676e0446a9bc","content":"缓存并没有消除搜索本身。exact 查询仍要与全部 $N$ 向量做点积；float32 图库矩阵占 $4Nd$ 字节。本例 $24\\times512\\times4=49,152$ 字节，不含图片、模型、ID 与索引。改变 checkpoint、预处理或图片内容需要重建对应缓存；只新增文字候选不改变旧图片向量。","sha256":"a0b9c4c8d9eebe422e20a0d9595195dc58efc43923716e264bd41dc0bfb60c5d","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"检索为什么能缓存：每张图编码一次，查询再比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":"retrieval-cache-interface","formulas":[{"latex":"N","definitionContext":"设图库有 $N$ 张图，收到 $Q$ 条文字查询。双编码器的图像输出不依赖当前查询，所以可以先计算图库矩阵 $V\\in\\mathbb R^{N\\times d}$，查询到来时只计算文字向量 $u$，再用 $Vu$ 得分。这把“理解输入的深度编码”和“在现有表示中比较”分成两段。\n若每个图文对都重新分别编码，需 $NQ$ 次图片样本编码和 $NQ$ 次文字样本编码；有缓存则是 $N+Q$。本课程 $N=24,Q=12$，两者为 576 与 36 次样本编码。它是解析工作量，不是秒数；一个批次调用可以同时编码多个样本。\n缓存并没有消除搜索本身。exact 查询仍要与全部 $N$ 向量做点积；float32 图库矩阵占 $4Nd$ 字节。本例 $24\\times512\\times4=49,152$ 字节，不含图片、模型、ID 与索引。改变 checkpoint、预处理或图片内容需要重建对应缓存；只新增文字候选不改变旧图片向量。\nP59 使用已有 24 项缓存显示猫查询的前三项：`chelsea-center90`、`chelsea-full`、`chelsea-right90`。24 项其实来自六张照片的四种相关裁剪，不能当成 24 张独立语义样本。它足以观察排名与缓存复用；更大的索引工程属于第八讲和 LAB12 的选读部分。"},{"latex":"4Nd","definitionContext":"设图库有 $N$ 张图，收到 $Q$ 条文字查询。双编码器的图像输出不依赖当前查询，所以可以先计算图库矩阵 $V\\in\\mathbb R^{N\\times d}$，查询到来时只计算文字向量 $u$，再用 $Vu$ 得分。这把“理解输入的深度编码”和“在现有表示中比较”分成两段。\n若每个图文对都重新分别编码，需 $NQ$ 次图片样本编码和 $NQ$ 次文字样本编码；有缓存则是 $N+Q$。本课程 $N=24,Q=12$，两者为 576 与 36 次样本编码。它是解析工作量，不是秒数；一个批次调用可以同时编码多个样本。\n缓存并没有消除搜索本身。exact 查询仍要与全部 $N$ 向量做点积；float32 图库矩阵占 $4Nd$ 字节。本例 $24\\times512\\times4=49,152$ 字节，不含图片、模型、ID 与索引。改变 checkpoint、预处理或图片内容需要重建对应缓存；只新增文字候选不改变旧图片向量。\nP59 使用已有 24 项缓存显示猫查询的前三项：`chelsea-center90`、`chelsea-full`、`chelsea-right90`。24 项其实来自六张照片的四种相关裁剪，不能当成 24 张独立语义样本。它足以观察排名与缓存复用；更大的索引工程属于第八讲和 LAB12 的选读部分。"},{"latex":"24\\times512\\times4=49,152","definitionContext":"设图库有 $N$ 张图，收到 $Q$ 条文字查询。双编码器的图像输出不依赖当前查询，所以可以先计算图库矩阵 $V\\in\\mathbb R^{N\\times d}$，查询到来时只计算文字向量 $u$，再用 $Vu$ 得分。这把“理解输入的深度编码”和“在现有表示中比较”分成两段。\n若每个图文对都重新分别编码，需 $NQ$ 次图片样本编码和 $NQ$ 次文字样本编码；有缓存则是 $N+Q$。本课程 $N=24,Q=12$，两者为 576 与 36 次样本编码。它是解析工作量，不是秒数；一个批次调用可以同时编码多个样本。\n缓存并没有消除搜索本身。exact 查询仍要与全部 $N$ 向量做点积；float32 图库矩阵占 $4Nd$ 字节。本例 $24\\times512\\times4=49,152$ 字节，不含图片、模型、ID 与索引。改变 checkpoint、预处理或图片内容需要重建对应缓存；只新增文字候选不改变旧图片向量。\nP59 使用已有 24 项缓存显示猫查询的前三项：`chelsea-center90`、`chelsea-full`、`chelsea-right90`。24 项其实来自六张照片的四种相关裁剪，不能当成 24 张独立语义样本。它足以观察排名与缓存复用；更大的索引工程属于第八讲和 LAB12 的选读部分。"}],"previousBlock":"b-59c0250b7b94373443fe912d","nextBlock":"b-526b538610a36604cd9f8957","prerequisites":[{"id":"b-b7a85d47252d737cfb0277fb","url":"#b-b7a85d47252d737cfb0277fb","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"},{"id":"b-b7a85d47252d737cfb0277fb","url":"#b-b7a85d47252d737cfb0277fb","relation":"same-reasoning"},{"id":"b-80907ac5f1234aede1d5d070","url":"#b-80907ac5f1234aede1d5d070","relation":"same-reasoning"},{"id":"b-59c0250b7b94373443fe912d","url":"#b-59c0250b7b94373443fe912d","relation":"same-reasoning"},{"id":"b-526b538610a36604cd9f8957","url":"#b-526b538610a36604cd9f8957","relation":"same-reasoning"}]},{"id":"b-526b538610a36604cd9f8957","type":"p","scope":"retrieval-cache-interface","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-526b538610a36604cd9f8957","content":"P59 使用已有 24 项缓存显示猫查询的前三项：`chelsea-center90`、`chelsea-full`、`chelsea-right90`。24 项其实来自六张照片的四种相关裁剪，不能当成 24 张独立语义样本。它足以观察排名与缓存复用；更大的索引工程属于第八讲和 LAB12 的选读部分。","sha256":"92827721d51ccba74677884e288a5b789ab5ded7a0520fdc0992941218d8ee81","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"检索为什么能缓存：每张图编码一次，查询再比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":"retrieval-cache-interface","formulas":[],"previousBlock":"b-cccb07251450676e0446a9bc","nextBlock":"b-98a0b3e4874726a16a672f8a","prerequisites":[{"id":"b-b7a85d47252d737cfb0277fb","url":"#b-b7a85d47252d737cfb0277fb","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"},{"id":"b-b7a85d47252d737cfb0277fb","url":"#b-b7a85d47252d737cfb0277fb","relation":"same-reasoning"},{"id":"b-80907ac5f1234aede1d5d070","url":"#b-80907ac5f1234aede1d5d070","relation":"same-reasoning"},{"id":"b-59c0250b7b94373443fe912d","url":"#b-59c0250b7b94373443fe912d","relation":"same-reasoning"},{"id":"b-cccb07251450676e0446a9bc","url":"#b-cccb07251450676e0446a9bc","relation":"same-reasoning"}]},{"id":"b-98a0b3e4874726a16a672f8a","type":"h4","scope":"retrieval-reuse-encoding","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-98a0b3e4874726a16a672f8a","content":"#### 图示条件与公式约定","sha256":"ef7f915c1ae63ed67bd23fecf050249273ba0efa8965d2c8d1c1c6b2d3178122","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"先编码图库，再让新查询比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-526b538610a36604cd9f8957","nextBlock":"b-33fd9610782b2b8229cca2f3","prerequisites":[{"id":"b-b7a85d47252d737cfb0277fb","url":"#b-b7a85d47252d737cfb0277fb","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"}]},{"id":"b-33fd9610782b2b8229cca2f3","type":"p","scope":"document","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-33fd9610782b2b8229cca2f3","content":"本页只复用前轮已存在的本地 CLIP 结果，不读取、不运行 Deepnote，不新跑模型。缓存模型 openai/clip-vit-base-patch32，查询 a photo of a cat.；已有 24 项图像库的精确排序前三项为 chelsea-center90、chelsea-full、chelsea-right90。24 项来自 6 张照片的相关裁剪，并非独立样本。协议、来源与缓存校验在已有证据包中。只展示精确小库；ANN 和缓存工程留补充。","sha256":"e06ae1708d414ac9f8f5b7ac04ac212f1b0939da9c7729b3878ec8b958a74b66","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"先编码图库，再让新查询比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-98a0b3e4874726a16a672f8a","nextBlock":"b-fc4eec79b355843695cec43f","prerequisites":[{"id":"b-98a0b3e4874726a16a672f8a","url":"#b-98a0b3e4874726a16a672f8a","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"}]},{"id":"b-fc4eec79b355843695cec43f","type":"h4","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-fc4eec79b355843695cec43f","content":"#### 这条线索的展开与后续阅读","sha256":"48b8229f7559b36ed4d5628416e3991e5c73293fbce43374db67386c8736f3bc","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"先编码图库，再让新查询比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-33fd9610782b2b8229cca2f3","nextBlock":"b-ad95390cfb85292761b64306","prerequisites":[{"id":"b-98a0b3e4874726a16a672f8a","url":"#b-98a0b3e4874726a16a672f8a","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"}]},{"id":"b-ad95390cfb85292761b64306","type":"ul","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-ad95390cfb85292761b64306","content":"- [交互发生得早或晚，决定缓存与表达｜问题从哪里出现](/supplement.html?return=l3-r2-p59/0#/l3-encoder-cost-problem/0) 第八讲相关预备阅读\n- [交互发生得早或晚，决定缓存与表达｜核心改动怎样工作](/supplement.html?return=l3-r2-p59/0#/l3-encoder-cost-mechanism/0) 第八讲相关预备阅读\n- [交互发生得早或晚，决定缓存与表达｜一起算一个完整例子](/supplement.html?return=l3-r2-p59/0#/l3-encoder-cost-example/0) 第八讲相关预备阅读\n- [交互发生得早或晚，决定缓存与表达｜怎样检查这项解释](/supplement.html?return=l3-r2-p59/0#/l3-encoder-cost-evidence/0) 第八讲相关预备阅读\n- [交互发生得早或晚，决定缓存与表达｜什么结论还不能推出](/supplement.html?return=l3-r2-p59/0#/l3-encoder-cost-boundary/0) 第八讲相关预备阅读","sha256":"ac480cd10ae33ad33bc08b2666dd0b6adbf60adc29c82b8aa1cc91945d5eac3b","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"先编码图库，再让新查询比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-fc4eec79b355843695cec43f","nextBlock":"b-113676cae8a960111acc5927","prerequisites":[{"id":"b-fc4eec79b355843695cec43f","url":"#b-fc4eec79b355843695cec43f","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"}]},{"id":"b-113676cae8a960111acc5927","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-113676cae8a960111acc5927","content":"- [交互发生得早或晚，决定缓存与表达｜问题从哪里出现](/supplement.html?return=l3-r2-p59/0#/l3-encoder-cost-problem/0) 第八讲相关预备阅读","sha256":"f19687a8a051cd6500e71ee01a8bc63867f5079ecca32db524f538b4a2c35153","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"先编码图库，再让新查询比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ad95390cfb85292761b64306","nextBlock":"b-b2782b587cc8dd08dfe6a3db","prerequisites":[{"id":"b-fc4eec79b355843695cec43f","url":"#b-fc4eec79b355843695cec43f","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"}]},{"id":"b-b2782b587cc8dd08dfe6a3db","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-b2782b587cc8dd08dfe6a3db","content":"- [交互发生得早或晚，决定缓存与表达｜核心改动怎样工作](/supplement.html?return=l3-r2-p59/0#/l3-encoder-cost-mechanism/0) 第八讲相关预备阅读","sha256":"51500a1bc1ac7c5e8538e5aeea6fb4c8f32e8312edae075d520a40cf6d9eb707","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"先编码图库，再让新查询比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-113676cae8a960111acc5927","nextBlock":"b-7a12fbe59a5e5835131be287","prerequisites":[{"id":"b-fc4eec79b355843695cec43f","url":"#b-fc4eec79b355843695cec43f","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"}]},{"id":"b-7a12fbe59a5e5835131be287","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-7a12fbe59a5e5835131be287","content":"- [交互发生得早或晚，决定缓存与表达｜一起算一个完整例子](/supplement.html?return=l3-r2-p59/0#/l3-encoder-cost-example/0) 第八讲相关预备阅读","sha256":"291861c96834ab4b94163ee8a3ce8ef3cc98d01bc52cdc2c4363af1cfc9b253b","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"先编码图库，再让新查询比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-b2782b587cc8dd08dfe6a3db","nextBlock":"b-fb4d2e33c9b5cd0c39989ee7","prerequisites":[{"id":"b-fc4eec79b355843695cec43f","url":"#b-fc4eec79b355843695cec43f","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"}]},{"id":"b-fb4d2e33c9b5cd0c39989ee7","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-fb4d2e33c9b5cd0c39989ee7","content":"- [交互发生得早或晚，决定缓存与表达｜怎样检查这项解释](/supplement.html?return=l3-r2-p59/0#/l3-encoder-cost-evidence/0) 第八讲相关预备阅读","sha256":"90c9bc3365b9cf34f307a6fd25c68d00a9215c3147c22e327f049f444c7caf86","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"先编码图库，再让新查询比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7a12fbe59a5e5835131be287","nextBlock":"b-fe24cefdb9636fe783dd9392","prerequisites":[{"id":"b-fc4eec79b355843695cec43f","url":"#b-fc4eec79b355843695cec43f","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"}]},{"id":"b-fe24cefdb9636fe783dd9392","type":"li","scope":"retained-reading","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-fe24cefdb9636fe783dd9392","content":"- [交互发生得早或晚，决定缓存与表达｜什么结论还不能推出](/supplement.html?return=l3-r2-p59/0#/l3-encoder-cost-boundary/0) 第八讲相关预备阅读","sha256":"ec4455aa1077a8444959c6a0665d97b8fad7cf29b957a5e954edcf4bf7683d1d","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-reuse-encoding","title":"先编码图库，再让新查询比较","number":2},"sectionTitle":"先编码图库，再让新查询比较","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-fb4d2e33c9b5cd0c39989ee7","nextBlock":null,"prerequisites":[{"id":"b-fc4eec79b355843695cec43f","url":"#b-fc4eec79b355843695cec43f","role":"section-title"},{"id":"b-7d727467abe508e04d558282","url":"#b-7d727467abe508e04d558282","role":"problem-context"}],"relatedBlocks":[{"id":"b-04edaf888621515863711528","url":"#b-04edaf888621515863711528","relation":"figure-and-explanation"}]},{"id":"b-826a8147b8b09509f27af15b","type":"h2","scope":"retrieval-metric-questions","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-826a8147b8b09509f27af15b","content":"## 让每个检索指标回答自己的问题","sha256":"1ecef71d4a57b1fa987665b39fb54887896eb7e5ffd21892a4c2b86d0917d6c6","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-metric-questions","title":"让每个检索指标回答自己的问题","number":3},"sectionTitle":"让每个检索指标回答自己的问题","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-cbd572bc6c5fc2b76c156c2c","prerequisites":[{"id":"b-cbd572bc6c5fc2b76c156c2c","url":"#b-cbd572bc6c5fc2b76c156c2c","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-cbd572bc6c5fc2b76c156c2c","type":"p","scope":"retrieval-metric-questions","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-cbd572bc6c5fc2b76c156c2c","content":"缓存改变计算组织；质量判断仍需明确查询、相关集合和 K。","sha256":"f687449dd129fdd11e40e7d693f066aa1b4283bf1cc04d5968eef03db520470e","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-metric-questions","title":"让每个检索指标回答自己的问题","number":3},"sectionTitle":"让每个检索指标回答自己的问题","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-826a8147b8b09509f27af15b","nextBlock":"b-254dd3e8402500f81270c047","prerequisites":[{"id":"b-826a8147b8b09509f27af15b","url":"#b-826a8147b8b09509f27af15b","role":"section-title"}],"relatedBlocks":[]},{"id":"b-254dd3e8402500f81270c047","type":"h3","scope":"retrieval-metric-contract","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-254dd3e8402500f81270c047","content":"### Hit、Recall 与 exact 重合率，各自在问什么","sha256":"a32f625f24bd569879229f77efbce5dc69b1243961d1cc94f086b20af7fd2e60","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-metric-questions","title":"让每个检索指标回答自己的问题","number":3},"sectionTitle":"Hit、Recall 与 exact 重合率，各自在问什么","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":"retrieval-metric-contract","formulas":[],"previousBlock":"b-cbd572bc6c5fc2b76c156c2c","nextBlock":"b-acc1d5d73f777002d50a8553","prerequisites":[{"id":"b-826a8147b8b09509f27af15b","url":"#b-826a8147b8b09509f27af15b","role":"section-title"},{"id":"b-cbd572bc6c5fc2b76c156c2c","url":"#b-cbd572bc6c5fc2b76c156c2c","role":"problem-context"}],"relatedBlocks":[{"id":"b-acc1d5d73f777002d50a8553","url":"#b-acc1d5d73f777002d50a8553","relation":"same-reasoning"},{"id":"b-18fc10821fe6af84adf28108","url":"#b-18fc10821fe6af84adf28108","relation":"same-reasoning"},{"id":"b-d283b5fafa4b4af876c178e7","url":"#b-d283b5fafa4b4af876c178e7","relation":"same-reasoning"},{"id":"b-10ec0695f64ff08ec26356c6","url":"#b-10ec0695f64ff08ec26356c6","relation":"same-reasoning"}]},{"id":"b-acc1d5d73f777002d50a8553","type":"p","scope":"retrieval-metric-contract","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-acc1d5d73f777002d50a8553","content":"固定一个查询，相关集合为 $R(q)$，检索前 $K$ 个结果为 $A_K(q)$。三个常见指标不是同义词：Hit@K 问是否至少命中一个相关结果；Recall@K 为 $|A_K\\cap R|/|R|$；ANN overlap 问近似搜索是否复现精确搜索，计算 $|A_K\\cap E_K|/K$，其中 $E_K$ 为 exact 前 K。","sha256":"7d1e522bf167385f95073a108d6651706c7ab2f25d134521e17b40fe2992bc59","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-metric-questions","title":"让每个检索指标回答自己的问题","number":3},"sectionTitle":"Hit、Recall 与 exact 重合率，各自在问什么","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":"retrieval-metric-contract","formulas":[{"latex":"R(q)","definitionContext":"固定一个查询，相关集合为 $R(q)$，检索前 $K$ 个结果为 $A_K(q)$。三个常见指标不是同义词：Hit@K 问是否至少命中一个相关结果；Recall@K 为 $|A_K\\cap R|/|R|$；ANN overlap 问近似搜索是否复现精确搜索，计算 $|A_K\\cap E_K|/K$，其中 $E_K$ 为 exact 前 K。\n假设某猫查询有四个猫裁剪算相关，前三名命中两个。Hit@3 是 1，Recall@3 是 $2/4=0.5$。若 exact 的前三名是 a,b,c，ANN 找到 a,b,d，二者 overlap 是 $2/3$；d 即使仍然是猫，语义命中也可能不变。模型相关性与索引逼近的是两个不同问题。\nLAB12 的历史结果正体现这个区别：exact 文搜图 Hit@3=1；IVF 只查一个簇时 exact top-3 重合率为 0.7778，语义 Hit@3 仍为 1。它漏掉部分同源裁剪，却仍保留某个相关裁剪。这不是指标互相矛盾，而是各自回答了自己的问题。\n阅读结果先找查询集合、相关集合和 K，再看成绩。若一个查询只有一个正例，Hit@K 与单正例 Recall@K 恰好同值；换成多正例后必须重新写清口径。保存查询和命中 ID 比只保存平均数字更便于 AI 或同学追溯解释。"},{"latex":"K","definitionContext":"固定一个查询，相关集合为 $R(q)$，检索前 $K$ 个结果为 $A_K(q)$。三个常见指标不是同义词：Hit@K 问是否至少命中一个相关结果；Recall@K 为 $|A_K\\cap R|/|R|$；ANN overlap 问近似搜索是否复现精确搜索，计算 $|A_K\\cap E_K|/K$，其中 $E_K$ 为 exact 前 K。\n假设某猫查询有四个猫裁剪算相关，前三名命中两个。Hit@3 是 1，Recall@3 是 $2/4=0.5$。若 exact 的前三名是 a,b,c，ANN 找到 a,b,d，二者 overlap 是 $2/3$；d 即使仍然是猫，语义命中也可能不变。模型相关性与索引逼近的是两个不同问题。\nLAB12 的历史结果正体现这个区别：exact 文搜图 Hit@3=1；IVF 只查一个簇时 exact top-3 重合率为 0.7778，语义 Hit@3 仍为 1。它漏掉部分同源裁剪，却仍保留某个相关裁剪。这不是指标互相矛盾，而是各自回答了自己的问题。\n阅读结果先找查询集合、相关集合和 K，再看成绩。若一个查询只有一个正例，Hit@K 与单正例 Recall@K 恰好同值；换成多正例后必须重新写清口径。保存查询和命中 ID 比只保存平均数字更便于 AI 或同学追溯解释。"},{"latex":"A_K(q)","definitionContext":"固定一个查询，相关集合为 $R(q)$，检索前 $K$ 个结果为 $A_K(q)$。三个常见指标不是同义词：Hit@K 问是否至少命中一个相关结果；Recall@K 为 $|A_K\\cap R|/|R|$；ANN overlap 问近似搜索是否复现精确搜索，计算 $|A_K\\cap E_K|/K$，其中 $E_K$ 为 exact 前 K。\n假设某猫查询有四个猫裁剪算相关，前三名命中两个。Hit@3 是 1，Recall@3 是 $2/4=0.5$。若 exact 的前三名是 a,b,c，ANN 找到 a,b,d，二者 overlap 是 $2/3$；d 即使仍然是猫，语义命中也可能不变。模型相关性与索引逼近的是两个不同问题。\nLAB12 的历史结果正体现这个区别：exact 文搜图 Hit@3=1；IVF 只查一个簇时 exact top-3 重合率为 0.7778，语义 Hit@3 仍为 1。它漏掉部分同源裁剪，却仍保留某个相关裁剪。这不是指标互相矛盾，而是各自回答了自己的问题。\n阅读结果先找查询集合、相关集合和 K，再看成绩。若一个查询只有一个正例，Hit@K 与单正例 Recall@K 恰好同值；换成多正例后必须重新写清口径。保存查询和命中 ID 比只保存平均数字更便于 AI 或同学追溯解释。"},{"latex":"|A_K\\cap R|/|R|","definitionContext":"固定一个查询，相关集合为 $R(q)$，检索前 $K$ 个结果为 $A_K(q)$。三个常见指标不是同义词：Hit@K 问是否至少命中一个相关结果；Recall@K 为 $|A_K\\cap R|/|R|$；ANN overlap 问近似搜索是否复现精确搜索，计算 $|A_K\\cap E_K|/K$，其中 $E_K$ 为 exact 前 K。\n假设某猫查询有四个猫裁剪算相关，前三名命中两个。Hit@3 是 1，Recall@3 是 $2/4=0.5$。若 exact 的前三名是 a,b,c，ANN 找到 a,b,d，二者 overlap 是 $2/3$；d 即使仍然是猫，语义命中也可能不变。模型相关性与索引逼近的是两个不同问题。\nLAB12 的历史结果正体现这个区别：exact 文搜图 Hit@3=1；IVF 只查一个簇时 exact top-3 重合率为 0.7778，语义 Hit@3 仍为 1。它漏掉部分同源裁剪，却仍保留某个相关裁剪。这不是指标互相矛盾，而是各自回答了自己的问题。\n阅读结果先找查询集合、相关集合和 K，再看成绩。若一个查询只有一个正例，Hit@K 与单正例 Recall@K 恰好同值；换成多正例后必须重新写清口径。保存查询和命中 ID 比只保存平均数字更便于 AI 或同学追溯解释。"},{"latex":"|A_K\\cap E_K|/K","definitionContext":"固定一个查询，相关集合为 $R(q)$，检索前 $K$ 个结果为 $A_K(q)$。三个常见指标不是同义词：Hit@K 问是否至少命中一个相关结果；Recall@K 为 $|A_K\\cap R|/|R|$；ANN overlap 问近似搜索是否复现精确搜索，计算 $|A_K\\cap E_K|/K$，其中 $E_K$ 为 exact 前 K。\n假设某猫查询有四个猫裁剪算相关，前三名命中两个。Hit@3 是 1，Recall@3 是 $2/4=0.5$。若 exact 的前三名是 a,b,c，ANN 找到 a,b,d，二者 overlap 是 $2/3$；d 即使仍然是猫，语义命中也可能不变。模型相关性与索引逼近的是两个不同问题。\nLAB12 的历史结果正体现这个区别：exact 文搜图 Hit@3=1；IVF 只查一个簇时 exact top-3 重合率为 0.7778，语义 Hit@3 仍为 1。它漏掉部分同源裁剪，却仍保留某个相关裁剪。这不是指标互相矛盾，而是各自回答了自己的问题。\n阅读结果先找查询集合、相关集合和 K，再看成绩。若一个查询只有一个正例，Hit@K 与单正例 Recall@K 恰好同值；换成多正例后必须重新写清口径。保存查询和命中 ID 比只保存平均数字更便于 AI 或同学追溯解释。"},{"latex":"E_K","definitionContext":"固定一个查询，相关集合为 $R(q)$，检索前 $K$ 个结果为 $A_K(q)$。三个常见指标不是同义词：Hit@K 问是否至少命中一个相关结果；Recall@K 为 $|A_K\\cap R|/|R|$；ANN overlap 问近似搜索是否复现精确搜索，计算 $|A_K\\cap E_K|/K$，其中 $E_K$ 为 exact 前 K。\n假设某猫查询有四个猫裁剪算相关，前三名命中两个。Hit@3 是 1，Recall@3 是 $2/4=0.5$。若 exact 的前三名是 a,b,c，ANN 找到 a,b,d，二者 overlap 是 $2/3$；d 即使仍然是猫，语义命中也可能不变。模型相关性与索引逼近的是两个不同问题。\nLAB12 的历史结果正体现这个区别：exact 文搜图 Hit@3=1；IVF 只查一个簇时 exact top-3 重合率为 0.7778，语义 Hit@3 仍为 1。它漏掉部分同源裁剪，却仍保留某个相关裁剪。这不是指标互相矛盾，而是各自回答了自己的问题。\n阅读结果先找查询集合、相关集合和 K，再看成绩。若一个查询只有一个正例，Hit@K 与单正例 Recall@K 恰好同值；换成多正例后必须重新写清口径。保存查询和命中 ID 比只保存平均数字更便于 AI 或同学追溯解释。"}],"previousBlock":"b-254dd3e8402500f81270c047","nextBlock":"b-18fc10821fe6af84adf28108","prerequisites":[{"id":"b-254dd3e8402500f81270c047","url":"#b-254dd3e8402500f81270c047","role":"section-title"},{"id":"b-cbd572bc6c5fc2b76c156c2c","url":"#b-cbd572bc6c5fc2b76c156c2c","role":"problem-context"}],"relatedBlocks":[{"id":"b-254dd3e8402500f81270c047","url":"#b-254dd3e8402500f81270c047","relation":"same-reasoning"},{"id":"b-18fc10821fe6af84adf28108","url":"#b-18fc10821fe6af84adf28108","relation":"same-reasoning"},{"id":"b-d283b5fafa4b4af876c178e7","url":"#b-d283b5fafa4b4af876c178e7","relation":"same-reasoning"},{"id":"b-10ec0695f64ff08ec26356c6","url":"#b-10ec0695f64ff08ec26356c6","relation":"same-reasoning"}]},{"id":"b-18fc10821fe6af84adf28108","type":"p","scope":"retrieval-metric-contract","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-18fc10821fe6af84adf28108","content":"假设某猫查询有四个猫裁剪算相关，前三名命中两个。Hit@3 是 1，Recall@3 是 $2/4=0.5$。若 exact 的前三名是 a,b,c，ANN 找到 a,b,d，二者 overlap 是 $2/3$；d 即使仍然是猫，语义命中也可能不变。模型相关性与索引逼近的是两个不同问题。","sha256":"e60979fc29d0931ab0fa6565feb357a2b9d204702fac523f27809082dbf318a7","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-metric-questions","title":"让每个检索指标回答自己的问题","number":3},"sectionTitle":"Hit、Recall 与 exact 重合率，各自在问什么","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":"retrieval-metric-contract","formulas":[{"latex":"2/4=0.5","definitionContext":"固定一个查询，相关集合为 $R(q)$，检索前 $K$ 个结果为 $A_K(q)$。三个常见指标不是同义词：Hit@K 问是否至少命中一个相关结果；Recall@K 为 $|A_K\\cap R|/|R|$；ANN overlap 问近似搜索是否复现精确搜索，计算 $|A_K\\cap E_K|/K$，其中 $E_K$ 为 exact 前 K。\n假设某猫查询有四个猫裁剪算相关，前三名命中两个。Hit@3 是 1，Recall@3 是 $2/4=0.5$。若 exact 的前三名是 a,b,c，ANN 找到 a,b,d，二者 overlap 是 $2/3$；d 即使仍然是猫，语义命中也可能不变。模型相关性与索引逼近的是两个不同问题。\nLAB12 的历史结果正体现这个区别：exact 文搜图 Hit@3=1；IVF 只查一个簇时 exact top-3 重合率为 0.7778，语义 Hit@3 仍为 1。它漏掉部分同源裁剪，却仍保留某个相关裁剪。这不是指标互相矛盾，而是各自回答了自己的问题。\n阅读结果先找查询集合、相关集合和 K，再看成绩。若一个查询只有一个正例，Hit@K 与单正例 Recall@K 恰好同值；换成多正例后必须重新写清口径。保存查询和命中 ID 比只保存平均数字更便于 AI 或同学追溯解释。"},{"latex":"2/3","definitionContext":"固定一个查询，相关集合为 $R(q)$，检索前 $K$ 个结果为 $A_K(q)$。三个常见指标不是同义词：Hit@K 问是否至少命中一个相关结果；Recall@K 为 $|A_K\\cap R|/|R|$；ANN overlap 问近似搜索是否复现精确搜索，计算 $|A_K\\cap E_K|/K$，其中 $E_K$ 为 exact 前 K。\n假设某猫查询有四个猫裁剪算相关，前三名命中两个。Hit@3 是 1，Recall@3 是 $2/4=0.5$。若 exact 的前三名是 a,b,c，ANN 找到 a,b,d，二者 overlap 是 $2/3$；d 即使仍然是猫，语义命中也可能不变。模型相关性与索引逼近的是两个不同问题。\nLAB12 的历史结果正体现这个区别：exact 文搜图 Hit@3=1；IVF 只查一个簇时 exact top-3 重合率为 0.7778，语义 Hit@3 仍为 1。它漏掉部分同源裁剪，却仍保留某个相关裁剪。这不是指标互相矛盾，而是各自回答了自己的问题。\n阅读结果先找查询集合、相关集合和 K，再看成绩。若一个查询只有一个正例，Hit@K 与单正例 Recall@K 恰好同值；换成多正例后必须重新写清口径。保存查询和命中 ID 比只保存平均数字更便于 AI 或同学追溯解释。"}],"previousBlock":"b-acc1d5d73f777002d50a8553","nextBlock":"b-d283b5fafa4b4af876c178e7","prerequisites":[{"id":"b-254dd3e8402500f81270c047","url":"#b-254dd3e8402500f81270c047","role":"section-title"},{"id":"b-cbd572bc6c5fc2b76c156c2c","url":"#b-cbd572bc6c5fc2b76c156c2c","role":"problem-context"}],"relatedBlocks":[{"id":"b-254dd3e8402500f81270c047","url":"#b-254dd3e8402500f81270c047","relation":"same-reasoning"},{"id":"b-acc1d5d73f777002d50a8553","url":"#b-acc1d5d73f777002d50a8553","relation":"same-reasoning"},{"id":"b-d283b5fafa4b4af876c178e7","url":"#b-d283b5fafa4b4af876c178e7","relation":"same-reasoning"},{"id":"b-10ec0695f64ff08ec26356c6","url":"#b-10ec0695f64ff08ec26356c6","relation":"same-reasoning"}]},{"id":"b-d283b5fafa4b4af876c178e7","type":"p","scope":"retrieval-metric-contract","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-d283b5fafa4b4af876c178e7","content":"LAB12 的历史结果正体现这个区别：exact 文搜图 Hit@3=1；IVF 只查一个簇时 exact top-3 重合率为 0.7778，语义 Hit@3 仍为 1。它漏掉部分同源裁剪，却仍保留某个相关裁剪。这不是指标互相矛盾，而是各自回答了自己的问题。","sha256":"e6be9a71494e6e8c8f730b05c8896471180ddb04def85ade8acb5519b43a466f","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-metric-questions","title":"让每个检索指标回答自己的问题","number":3},"sectionTitle":"Hit、Recall 与 exact 重合率，各自在问什么","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":"retrieval-metric-contract","formulas":[],"previousBlock":"b-18fc10821fe6af84adf28108","nextBlock":"b-10ec0695f64ff08ec26356c6","prerequisites":[{"id":"b-254dd3e8402500f81270c047","url":"#b-254dd3e8402500f81270c047","role":"section-title"},{"id":"b-cbd572bc6c5fc2b76c156c2c","url":"#b-cbd572bc6c5fc2b76c156c2c","role":"problem-context"}],"relatedBlocks":[{"id":"b-254dd3e8402500f81270c047","url":"#b-254dd3e8402500f81270c047","relation":"same-reasoning"},{"id":"b-acc1d5d73f777002d50a8553","url":"#b-acc1d5d73f777002d50a8553","relation":"same-reasoning"},{"id":"b-18fc10821fe6af84adf28108","url":"#b-18fc10821fe6af84adf28108","relation":"same-reasoning"},{"id":"b-10ec0695f64ff08ec26356c6","url":"#b-10ec0695f64ff08ec26356c6","relation":"same-reasoning"}]},{"id":"b-10ec0695f64ff08ec26356c6","type":"p","scope":"retrieval-metric-contract","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-10ec0695f64ff08ec26356c6","content":"阅读结果先找查询集合、相关集合和 K，再看成绩。若一个查询只有一个正例，Hit@K 与单正例 Recall@K 恰好同值；换成多正例后必须重新写清口径。保存查询和命中 ID 比只保存平均数字更便于 AI 或同学追溯解释。","sha256":"c1afda030628985fb13e6d5c22f9cfe7d7b0fe2cab59e7e3f191abb1dcef9feb","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-metric-questions","title":"让每个检索指标回答自己的问题","number":3},"sectionTitle":"Hit、Recall 与 exact 重合率，各自在问什么","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":"retrieval-metric-contract","formulas":[],"previousBlock":"b-d283b5fafa4b4af876c178e7","nextBlock":null,"prerequisites":[{"id":"b-254dd3e8402500f81270c047","url":"#b-254dd3e8402500f81270c047","role":"section-title"},{"id":"b-cbd572bc6c5fc2b76c156c2c","url":"#b-cbd572bc6c5fc2b76c156c2c","role":"problem-context"}],"relatedBlocks":[{"id":"b-254dd3e8402500f81270c047","url":"#b-254dd3e8402500f81270c047","relation":"same-reasoning"},{"id":"b-acc1d5d73f777002d50a8553","url":"#b-acc1d5d73f777002d50a8553","relation":"same-reasoning"},{"id":"b-18fc10821fe6af84adf28108","url":"#b-18fc10821fe6af84adf28108","relation":"same-reasoning"},{"id":"b-d283b5fafa4b4af876c178e7","url":"#b-d283b5fafa4b4af876c178e7","relation":"same-reasoning"}]},{"id":"b-7002abd47beba8b7772509e3","type":"h2","scope":"retrieval-search-and-interaction","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-7002abd47beba8b7772509e3","content":"## 选读：省搜索与增加图文交互分别改变什么","sha256":"b04ca7f2c77d979aa917eeb222b9c820c3bacd757368df4158ef70a6f3bb5ba3","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-search-and-interaction","title":"选读：省搜索与增加图文交互分别改变什么","number":4},"sectionTitle":"选读：省搜索与增加图文交互分别改变什么","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":null,"nextBlock":"b-d5a2b891222095ad5a946190","prerequisites":[{"id":"b-d5a2b891222095ad5a946190","url":"#b-d5a2b891222095ad5a946190","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-d5a2b891222095ad5a946190","type":"p","scope":"retrieval-search-and-interaction","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-d5a2b891222095ad5a946190","content":"指标已分清语义命中与 exact 复现；进一步只在关心系统成本或关系表达时阅读。","sha256":"2f8771eb19af43f3265be23f63db25705c09f392e182967c7d04fc14645702c8","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-search-and-interaction","title":"选读：省搜索与增加图文交互分别改变什么","number":4},"sectionTitle":"选读：省搜索与增加图文交互分别改变什么","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-7002abd47beba8b7772509e3","nextBlock":"b-b7039f5e8869fce0a5ed309d","prerequisites":[{"id":"b-7002abd47beba8b7772509e3","url":"#b-7002abd47beba8b7772509e3","role":"section-title"}],"relatedBlocks":[]},{"id":"b-b7039f5e8869fce0a5ed309d","type":"h3","scope":"retrieval-approximation-and-interaction","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-b7039f5e8869fce0a5ed309d","content":"### 选读：省搜索和增加图文交互，是不同方向","sha256":"225cf2646d1bbc81fd3f8ed64c5eac6bf960b0c41b919b6f24554236d701a629","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-search-and-interaction","title":"选读：省搜索与增加图文交互分别改变什么","number":4},"sectionTitle":"选读：省搜索和增加图文交互，是不同方向","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":"retrieval-approximation-and-interaction","formulas":[],"previousBlock":"b-d5a2b891222095ad5a946190","nextBlock":"b-d4d365ef6131f1bec7a8a0ee","prerequisites":[{"id":"b-7002abd47beba8b7772509e3","url":"#b-7002abd47beba8b7772509e3","role":"section-title"},{"id":"b-d5a2b891222095ad5a946190","url":"#b-d5a2b891222095ad5a946190","role":"problem-context"}],"relatedBlocks":[{"id":"b-d4d365ef6131f1bec7a8a0ee","url":"#b-d4d365ef6131f1bec7a8a0ee","relation":"same-reasoning"},{"id":"b-daa988ff5623cbdd887cd4df","url":"#b-daa988ff5623cbdd887cd4df","relation":"same-reasoning"},{"id":"b-d083e84b19466f7c176c46d8","url":"#b-d083e84b19466f7c176c46d8","relation":"same-reasoning"},{"id":"b-afe92d2fc8854ed5d7b12714","url":"#b-afe92d2fc8854ed5d7b12714","relation":"same-reasoning"}]},{"id":"b-d4d365ef6131f1bec7a8a0ee","type":"p","scope":"retrieval-approximation-and-interaction","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-d4d365ef6131f1bec7a8a0ee","content":"这部分服务于第八讲的检索预备阅读，不是 P59 必须讲完的技术章。LAB12 的 IVF-Flat 先把图库分成 8 个簇，查询只打开 `nprobe` 个最近簇，再对桶内原向量精确打分。近似出现在没有打开的桶；把所有 8 个桶打开应恢复 exact 排名。","sha256":"1766e359ec4d31dd1478ef54c245a79bf9e3cd2bcb748e45f6e24d363cc4b064","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-search-and-interaction","title":"选读：省搜索与增加图文交互分别改变什么","number":4},"sectionTitle":"选读：省搜索和增加图文交互，是不同方向","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":"retrieval-approximation-and-interaction","formulas":[],"previousBlock":"b-b7039f5e8869fce0a5ed309d","nextBlock":"b-daa988ff5623cbdd887cd4df","prerequisites":[{"id":"b-b7039f5e8869fce0a5ed309d","url":"#b-b7039f5e8869fce0a5ed309d","role":"section-title"},{"id":"b-d5a2b891222095ad5a946190","url":"#b-d5a2b891222095ad5a946190","role":"problem-context"}],"relatedBlocks":[{"id":"b-b7039f5e8869fce0a5ed309d","url":"#b-b7039f5e8869fce0a5ed309d","relation":"same-reasoning"},{"id":"b-daa988ff5623cbdd887cd4df","url":"#b-daa988ff5623cbdd887cd4df","relation":"same-reasoning"},{"id":"b-d083e84b19466f7c176c46d8","url":"#b-d083e84b19466f7c176c46d8","relation":"same-reasoning"},{"id":"b-afe92d2fc8854ed5d7b12714","url":"#b-afe92d2fc8854ed5d7b12714","relation":"same-reasoning"}]},{"id":"b-daa988ff5623cbdd887cd4df","type":"p","scope":"retrieval-approximation-and-interaction","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-daa988ff5623cbdd887cd4df","content":"2026-09-22 保存的本地结果中，nprobe=1 平均扫描 4.17 个图像向量，exact 扫描 24 个；但整批 12 查询中位耗时为约 0.1016 ms 与 0.0456 ms。小库矩阵乘法已经很便宜，Python 路由和排序成本反而更大。扫描更少不等于实测更快。计时排除模型编码、磁盘读取和建索引。","sha256":"dfb749be9ea401298591c13840469fdf0010697e73c244978111ebf419a560a3","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-search-and-interaction","title":"选读：省搜索与增加图文交互分别改变什么","number":4},"sectionTitle":"选读：省搜索和增加图文交互，是不同方向","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":"retrieval-approximation-and-interaction","formulas":[],"previousBlock":"b-d4d365ef6131f1bec7a8a0ee","nextBlock":"b-d083e84b19466f7c176c46d8","prerequisites":[{"id":"b-b7039f5e8869fce0a5ed309d","url":"#b-b7039f5e8869fce0a5ed309d","role":"section-title"},{"id":"b-d5a2b891222095ad5a946190","url":"#b-d5a2b891222095ad5a946190","role":"problem-context"}],"relatedBlocks":[{"id":"b-b7039f5e8869fce0a5ed309d","url":"#b-b7039f5e8869fce0a5ed309d","relation":"same-reasoning"},{"id":"b-d4d365ef6131f1bec7a8a0ee","url":"#b-d4d365ef6131f1bec7a8a0ee","relation":"same-reasoning"},{"id":"b-d083e84b19466f7c176c46d8","url":"#b-d083e84b19466f7c176c46d8","relation":"same-reasoning"},{"id":"b-afe92d2fc8854ed5d7b12714","url":"#b-afe92d2fc8854ed5d7b12714","relation":"same-reasoning"}]},{"id":"b-d083e84b19466f7c176c46d8","type":"p","scope":"retrieval-approximation-and-interaction","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-d083e84b19466f7c176c46d8","content":"另一个方向是让查询参与局部交互。缓存左右区域向量 $r_j$，对查询 $q$ 计算 $u_j=q^Tr_j$、$w_j=\\mathrm{softmax}(u_j/.05)$，再汇总 $\\sum_j w_ju_j$。猫查询可提高猫区域权重，宇航员查询可提高宇航员区域权重；但这个聚合对区域顺序不敏感，交换左右后总分不变。多向量并不自动保存关系。","sha256":"6424f8550f5b5c6628f1965c5c1b4aafa939764425130e58f1a53d555fcca12a","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-search-and-interaction","title":"选读：省搜索与增加图文交互分别改变什么","number":4},"sectionTitle":"选读：省搜索和增加图文交互，是不同方向","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":"retrieval-approximation-and-interaction","formulas":[{"latex":"r_j","definitionContext":"这部分服务于第八讲的检索预备阅读，不是 P59 必须讲完的技术章。LAB12 的 IVF-Flat 先把图库分成 8 个簇，查询只打开 `nprobe` 个最近簇，再对桶内原向量精确打分。近似出现在没有打开的桶；把所有 8 个桶打开应恢复 exact 排名。\n2026-09-22 保存的本地结果中，nprobe=1 平均扫描 4.17 个图像向量，exact 扫描 24 个；但整批 12 查询中位耗时为约 0.1016 ms 与 0.0456 ms。小库矩阵乘法已经很便宜，Python 路由和排序成本反而更大。扫描更少不等于实测更快。计时排除模型编码、磁盘读取和建索引。\n另一个方向是让查询参与局部交互。缓存左右区域向量 $r_j$，对查询 $q$ 计算 $u_j=q^Tr_j$、$w_j=\\mathrm{softmax}(u_j/.05)$，再汇总 $\\sum_j w_ju_j$。猫查询可提高猫区域权重，宇航员查询可提高宇航员区域权重；但这个聚合对区域顺序不敏感，交换左右后总分不变。多向量并不自动保存关系。\n逐对重复调用独立 CLIP 编码器仍是 $f(I)^Tg(T)$，不因此变成交叉编码。联合模型把图文在内部相互条件化，通常需要更多查询相关计算。可用双编码先召回，再少量重排；要判断收益，需要固定检索库、查询和评价协议。这些机制在现有 LAB12 保留为选做，不在主讲另开一章。"},{"latex":"q","definitionContext":"这部分服务于第八讲的检索预备阅读，不是 P59 必须讲完的技术章。LAB12 的 IVF-Flat 先把图库分成 8 个簇，查询只打开 `nprobe` 个最近簇，再对桶内原向量精确打分。近似出现在没有打开的桶；把所有 8 个桶打开应恢复 exact 排名。\n2026-09-22 保存的本地结果中，nprobe=1 平均扫描 4.17 个图像向量，exact 扫描 24 个；但整批 12 查询中位耗时为约 0.1016 ms 与 0.0456 ms。小库矩阵乘法已经很便宜，Python 路由和排序成本反而更大。扫描更少不等于实测更快。计时排除模型编码、磁盘读取和建索引。\n另一个方向是让查询参与局部交互。缓存左右区域向量 $r_j$，对查询 $q$ 计算 $u_j=q^Tr_j$、$w_j=\\mathrm{softmax}(u_j/.05)$，再汇总 $\\sum_j w_ju_j$。猫查询可提高猫区域权重，宇航员查询可提高宇航员区域权重；但这个聚合对区域顺序不敏感，交换左右后总分不变。多向量并不自动保存关系。\n逐对重复调用独立 CLIP 编码器仍是 $f(I)^Tg(T)$，不因此变成交叉编码。联合模型把图文在内部相互条件化，通常需要更多查询相关计算。可用双编码先召回，再少量重排；要判断收益，需要固定检索库、查询和评价协议。这些机制在现有 LAB12 保留为选做，不在主讲另开一章。"},{"latex":"u_j=q^Tr_j","definitionContext":"这部分服务于第八讲的检索预备阅读，不是 P59 必须讲完的技术章。LAB12 的 IVF-Flat 先把图库分成 8 个簇，查询只打开 `nprobe` 个最近簇，再对桶内原向量精确打分。近似出现在没有打开的桶；把所有 8 个桶打开应恢复 exact 排名。\n2026-09-22 保存的本地结果中，nprobe=1 平均扫描 4.17 个图像向量，exact 扫描 24 个；但整批 12 查询中位耗时为约 0.1016 ms 与 0.0456 ms。小库矩阵乘法已经很便宜，Python 路由和排序成本反而更大。扫描更少不等于实测更快。计时排除模型编码、磁盘读取和建索引。\n另一个方向是让查询参与局部交互。缓存左右区域向量 $r_j$，对查询 $q$ 计算 $u_j=q^Tr_j$、$w_j=\\mathrm{softmax}(u_j/.05)$，再汇总 $\\sum_j w_ju_j$。猫查询可提高猫区域权重，宇航员查询可提高宇航员区域权重；但这个聚合对区域顺序不敏感，交换左右后总分不变。多向量并不自动保存关系。\n逐对重复调用独立 CLIP 编码器仍是 $f(I)^Tg(T)$，不因此变成交叉编码。联合模型把图文在内部相互条件化，通常需要更多查询相关计算。可用双编码先召回，再少量重排；要判断收益，需要固定检索库、查询和评价协议。这些机制在现有 LAB12 保留为选做，不在主讲另开一章。"},{"latex":"w_j=\\mathrm{softmax}(u_j/.05)","definitionContext":"这部分服务于第八讲的检索预备阅读，不是 P59 必须讲完的技术章。LAB12 的 IVF-Flat 先把图库分成 8 个簇，查询只打开 `nprobe` 个最近簇，再对桶内原向量精确打分。近似出现在没有打开的桶；把所有 8 个桶打开应恢复 exact 排名。\n2026-09-22 保存的本地结果中，nprobe=1 平均扫描 4.17 个图像向量，exact 扫描 24 个；但整批 12 查询中位耗时为约 0.1016 ms 与 0.0456 ms。小库矩阵乘法已经很便宜，Python 路由和排序成本反而更大。扫描更少不等于实测更快。计时排除模型编码、磁盘读取和建索引。\n另一个方向是让查询参与局部交互。缓存左右区域向量 $r_j$，对查询 $q$ 计算 $u_j=q^Tr_j$、$w_j=\\mathrm{softmax}(u_j/.05)$，再汇总 $\\sum_j w_ju_j$。猫查询可提高猫区域权重，宇航员查询可提高宇航员区域权重；但这个聚合对区域顺序不敏感，交换左右后总分不变。多向量并不自动保存关系。\n逐对重复调用独立 CLIP 编码器仍是 $f(I)^Tg(T)$，不因此变成交叉编码。联合模型把图文在内部相互条件化，通常需要更多查询相关计算。可用双编码先召回，再少量重排；要判断收益，需要固定检索库、查询和评价协议。这些机制在现有 LAB12 保留为选做，不在主讲另开一章。"},{"latex":"\\sum_j w_ju_j","definitionContext":"这部分服务于第八讲的检索预备阅读，不是 P59 必须讲完的技术章。LAB12 的 IVF-Flat 先把图库分成 8 个簇，查询只打开 `nprobe` 个最近簇，再对桶内原向量精确打分。近似出现在没有打开的桶；把所有 8 个桶打开应恢复 exact 排名。\n2026-09-22 保存的本地结果中，nprobe=1 平均扫描 4.17 个图像向量，exact 扫描 24 个；但整批 12 查询中位耗时为约 0.1016 ms 与 0.0456 ms。小库矩阵乘法已经很便宜，Python 路由和排序成本反而更大。扫描更少不等于实测更快。计时排除模型编码、磁盘读取和建索引。\n另一个方向是让查询参与局部交互。缓存左右区域向量 $r_j$，对查询 $q$ 计算 $u_j=q^Tr_j$、$w_j=\\mathrm{softmax}(u_j/.05)$，再汇总 $\\sum_j w_ju_j$。猫查询可提高猫区域权重，宇航员查询可提高宇航员区域权重；但这个聚合对区域顺序不敏感，交换左右后总分不变。多向量并不自动保存关系。\n逐对重复调用独立 CLIP 编码器仍是 $f(I)^Tg(T)$，不因此变成交叉编码。联合模型把图文在内部相互条件化，通常需要更多查询相关计算。可用双编码先召回，再少量重排；要判断收益，需要固定检索库、查询和评价协议。这些机制在现有 LAB12 保留为选做，不在主讲另开一章。"}],"previousBlock":"b-daa988ff5623cbdd887cd4df","nextBlock":"b-afe92d2fc8854ed5d7b12714","prerequisites":[{"id":"b-b7039f5e8869fce0a5ed309d","url":"#b-b7039f5e8869fce0a5ed309d","role":"section-title"},{"id":"b-d5a2b891222095ad5a946190","url":"#b-d5a2b891222095ad5a946190","role":"problem-context"}],"relatedBlocks":[{"id":"b-b7039f5e8869fce0a5ed309d","url":"#b-b7039f5e8869fce0a5ed309d","relation":"same-reasoning"},{"id":"b-d4d365ef6131f1bec7a8a0ee","url":"#b-d4d365ef6131f1bec7a8a0ee","relation":"same-reasoning"},{"id":"b-daa988ff5623cbdd887cd4df","url":"#b-daa988ff5623cbdd887cd4df","relation":"same-reasoning"},{"id":"b-afe92d2fc8854ed5d7b12714","url":"#b-afe92d2fc8854ed5d7b12714","relation":"same-reasoning"}]},{"id":"b-afe92d2fc8854ed5d7b12714","type":"p","scope":"retrieval-approximation-and-interaction","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-afe92d2fc8854ed5d7b12714","content":"逐对重复调用独立 CLIP 编码器仍是 $f(I)^Tg(T)$，不因此变成交叉编码。联合模型把图文在内部相互条件化，通常需要更多查询相关计算。可用双编码先召回，再少量重排；要判断收益，需要固定检索库、查询和评价协议。这些机制在现有 LAB12 保留为选做，不在主讲另开一章。","sha256":"4317d93511e6997230f0f90631867ff02fa98c7c130893b0bfe4a4e45f7238d1","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-search-and-interaction","title":"选读：省搜索与增加图文交互分别改变什么","number":4},"sectionTitle":"选读：省搜索和增加图文交互，是不同方向","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":"retrieval-approximation-and-interaction","formulas":[{"latex":"f(I)^Tg(T)","definitionContext":"这部分服务于第八讲的检索预备阅读，不是 P59 必须讲完的技术章。LAB12 的 IVF-Flat 先把图库分成 8 个簇，查询只打开 `nprobe` 个最近簇，再对桶内原向量精确打分。近似出现在没有打开的桶；把所有 8 个桶打开应恢复 exact 排名。\n2026-09-22 保存的本地结果中，nprobe=1 平均扫描 4.17 个图像向量，exact 扫描 24 个；但整批 12 查询中位耗时为约 0.1016 ms 与 0.0456 ms。小库矩阵乘法已经很便宜，Python 路由和排序成本反而更大。扫描更少不等于实测更快。计时排除模型编码、磁盘读取和建索引。\n另一个方向是让查询参与局部交互。缓存左右区域向量 $r_j$，对查询 $q$ 计算 $u_j=q^Tr_j$、$w_j=\\mathrm{softmax}(u_j/.05)$，再汇总 $\\sum_j w_ju_j$。猫查询可提高猫区域权重，宇航员查询可提高宇航员区域权重；但这个聚合对区域顺序不敏感，交换左右后总分不变。多向量并不自动保存关系。\n逐对重复调用独立 CLIP 编码器仍是 $f(I)^Tg(T)$，不因此变成交叉编码。联合模型把图文在内部相互条件化，通常需要更多查询相关计算。可用双编码先召回，再少量重排；要判断收益，需要固定检索库、查询和评价协议。这些机制在现有 LAB12 保留为选做，不在主讲另开一章。"}],"previousBlock":"b-d083e84b19466f7c176c46d8","nextBlock":"b-729752ad3d2427f4e9c7375d","prerequisites":[{"id":"b-b7039f5e8869fce0a5ed309d","url":"#b-b7039f5e8869fce0a5ed309d","role":"section-title"},{"id":"b-d5a2b891222095ad5a946190","url":"#b-d5a2b891222095ad5a946190","role":"problem-context"}],"relatedBlocks":[{"id":"b-b7039f5e8869fce0a5ed309d","url":"#b-b7039f5e8869fce0a5ed309d","relation":"same-reasoning"},{"id":"b-d4d365ef6131f1bec7a8a0ee","url":"#b-d4d365ef6131f1bec7a8a0ee","relation":"same-reasoning"},{"id":"b-daa988ff5623cbdd887cd4df","url":"#b-daa988ff5623cbdd887cd4df","relation":"same-reasoning"},{"id":"b-d083e84b19466f7c176c46d8","url":"#b-d083e84b19466f7c176c46d8","relation":"same-reasoning"}]},{"id":"b-729752ad3d2427f4e9c7375d","type":"h3","scope":"dialogue","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-729752ad3d2427f4e9c7375d","content":"### 沿着图解追问","sha256":"f16be0424762a44344338696f06ef266c4571d15367ae2cbbd28fac98ac7e9cb","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-search-and-interaction","title":"选读：省搜索与增加图文交互分别改变什么","number":4},"sectionTitle":"选读：省搜索与增加图文交互分别改变什么","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-afe92d2fc8854ed5d7b12714","nextBlock":"b-f6052c27e01d5aaa6d99537c","prerequisites":[{"id":"b-b7039f5e8869fce0a5ed309d","url":"#b-b7039f5e8869fce0a5ed309d","role":"section-title"},{"id":"b-d5a2b891222095ad5a946190","url":"#b-d5a2b891222095ad5a946190","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-f6052c27e01d5aaa6d99537c","type":"p","scope":"dialogue","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-f6052c27e01d5aaa6d99537c","content":"以下是备课用的模拟师生对话，不是课堂实录。","sha256":"b25685f04cd1b885f3a5999e6a912460c7424021cdc5b2190584242086a82681","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-search-and-interaction","title":"选读：省搜索与增加图文交互分别改变什么","number":4},"sectionTitle":"选读：省搜索与增加图文交互分别改变什么","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-729752ad3d2427f4e9c7375d","nextBlock":"b-dfb9ff8ddaa8f556aee5d733","prerequisites":[{"id":"b-729752ad3d2427f4e9c7375d","url":"#b-729752ad3d2427f4e9c7375d","role":"section-title"},{"id":"b-d5a2b891222095ad5a946190","url":"#b-d5a2b891222095ad5a946190","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-dfb9ff8ddaa8f556aee5d733","type":"p","scope":"dialogue","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-dfb9ff8ddaa8f556aee5d733","content":"**学生：**缓存后为什么还需要检索算法？","sha256":"817e631e55b07e09213be086a884c4de66045705f64e23106ed9b30aebdf157d","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-search-and-interaction","title":"选读：省搜索与增加图文交互分别改变什么","number":4},"sectionTitle":"选读：省搜索与增加图文交互分别改变什么","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f6052c27e01d5aaa6d99537c","nextBlock":"b-f1422c87e4d51f608a7df583","prerequisites":[{"id":"b-729752ad3d2427f4e9c7375d","url":"#b-729752ad3d2427f4e9c7375d","role":"section-title"},{"id":"b-d5a2b891222095ad5a946190","url":"#b-d5a2b891222095ad5a946190","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-f1422c87e4d51f608a7df583","type":"p","scope":"dialogue","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-f1422c87e4d51f608a7df583","content":"**教师：**缓存省去了重复深度编码，却还要在 N 个向量里比较和选前 K。小库精确点积可能最简单，大库才需要衡量索引省的工作与遗漏。","sha256":"d51727108cd12db21b736dc34162c5db4a173ebea0e9fbb86f3d2ee40c0911b7","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-search-and-interaction","title":"选读：省搜索与增加图文交互分别改变什么","number":4},"sectionTitle":"选读：省搜索与增加图文交互分别改变什么","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-dfb9ff8ddaa8f556aee5d733","nextBlock":"b-ab6c729c8025f6765e30696d","prerequisites":[{"id":"b-729752ad3d2427f4e9c7375d","url":"#b-729752ad3d2427f4e9c7375d","role":"section-title"},{"id":"b-d5a2b891222095ad5a946190","url":"#b-d5a2b891222095ad5a946190","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-ab6c729c8025f6765e30696d","type":"h2","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-ab6c729c8025f6765e30696d","content":"## 沿这个问题继续","sha256":"cf44fcc62782fb4f6bdc42c16c5187d57556a03cb7ce6065e7bd9c36d8e7d733","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-search-and-interaction","title":"选读：省搜索与增加图文交互分别改变什么","number":4},"sectionTitle":"选读：省搜索与增加图文交互分别改变什么","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f1422c87e4d51f608a7df583","nextBlock":"b-f0df6a6a8d74c4d50ead2cd7","prerequisites":[{"id":"b-729752ad3d2427f4e9c7375d","url":"#b-729752ad3d2427f4e9c7375d","role":"section-title"},{"id":"b-d5a2b891222095ad5a946190","url":"#b-d5a2b891222095ad5a946190","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-f0df6a6a8d74c4d50ead2cd7","type":"p","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-f0df6a6a8d74c4d50ead2cd7","content":"先回看本页画面并指出输入、条件和输出，再按“真实猫图结果：保留模型、照片与历史身份 → 检索为什么能缓存：每张图编码一次，查询再比较 → Hit、Recall 与 exact 重合率，各自在问什么”阅读。相关实验以其已保存结果为起点，需要动手时再顺序执行 Notebook；","sha256":"0b7bf6c3e73b32d9d7f76b97606179fde1b53cca0b81d8e498a24853b0999b89","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-search-and-interaction","title":"选读：省搜索与增加图文交互分别改变什么","number":4},"sectionTitle":"选读：省搜索与增加图文交互分别改变什么","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-ab6c729c8025f6765e30696d","nextBlock":"b-7058b5a610b8d407fe4a8cef","prerequisites":[{"id":"b-ab6c729c8025f6765e30696d","url":"#b-ab6c729c8025f6765e30696d","role":"section-title"},{"id":"b-d5a2b891222095ad5a946190","url":"#b-d5a2b891222095ad5a946190","role":"problem-context"}],"relatedBlocks":[]},{"id":"b-7058b5a610b8d407fe4a8cef","type":"h4","scope":"practice","url":"https://codingai-lec03.pages.dev/course/pages/l3-r2-p59.html#b-7058b5a610b8d407fe4a8cef","content":"#### 按问题回查的讲解与短练习（1）","sha256":"0480869f0a66009b015a757d67aed6061011e378987f98a024f4e8bfc5386152","chapter":{"id":"III","title":"视觉与语言"},"unit":{"id":"retrieval-cache-and-evaluation","title":"检索到底复用了什么，又省去了什么？"},"page":{"id":"retrieval-search-and-interaction","title":"选读：省搜索与增加图文交互分别改变什么","number":4},"sectionTitle":"选读：省搜索与增加图文交互分别改变什么","currentQuestion":"编码缓存、语义相关性与近似搜索要怎样分别检验？","sourceSlides":["l3-r2-p59"],"knowledgeBlock":null,"formulas":[],"previousBlock":"b-f0df6a6a8d74c4d50ead2cd7","nextBlock":null,"prerequisites":[{"id":"b-ab6c729c8025f6765e30696d","url":"#b-ab6c729c8025f6765e30696d","role":"section-title"},{"id":"b-d5a2b891222095ad5a946190","url":"#b-d5a2b891222095ad5a946190","role":"problem-context"}],"relatedBlocks":[]}]}