1import{ab as p,l as S}from"./js.cookie.a9c443ab.js";import{u as P}from"./useEurekaUrl.82a9d372.js";import{u as T,_ as I}from"./_plugin-vue_export-helper.d0b30c10.js";import{u}from"./composables.2acea480.js";import{f as b,r as R,q as m,x as y,z as e,af as d,u as s,ag as h,C as A,D as x,aj as f,al as g,h as C,y as k}from"./runtime-core.esm-bundler.693c71a8.js";import"./state.a5689590.js";const w=n=>(A("data-v-e023de21"),n=n(),x(),n),D={class:"benchmark-detail"},F={class:"bench-banner"},z={class:"w1280"},$=w(()=>e("p",{class:"bench-banner-date"},"July 2026",-1)),O=w(()=>e("h2",null,"Patsnap PatentBench for Novelty Search",-1)),E={class:"bench-list"},B={class:"w1280"},N={id:"benchContent1"},H=f('<div class="banech-content" data-v-e023de21><div class="w1280" data-v-e023de21><div class="banech-pt-100" data-v-e023de21><h2 class="banech-public-h2" data-v-e023de21>Understanding Novelty Search</h2><p class="banech-public-p" data-v-e023de21> Novelty search is a key patent task that involves systematically identifying prior art worldwide to determine whether a technical solution is new and inventive under patent law. </p><p class="banech-public-p bench-pt-24" data-v-e023de21> It plays a critical role throughout the innovation process, including: </p><p class="banech-public-p bench-pt-24" data-v-e023de21><span data-v-e023de21></span>R&D planning: guiding the direction and feasibility of new developments </p><p class="banech-public-p" data-v-e023de21><span data-v-e023de21></span>Pre-filing: verifying that an invention is patentable before submission </p><p class="banech-public-p" data-v-e023de21><span data-v-e023de21></span>Patent examination: helping examiners assess the novelty of applications </p></div><div class="banech-pt-100" data-v-e023de21><h2 class="banech-public-h2" data-v-e023de21>Key Findings</h2><p class="banech-public-p" data-v-e023de21> This benchmark evaluates seven AI tools for patent novelty search: Patsnap's Novelty Search AI Agent, Claude Opus 4.8 (with web search), Perplexity Pro (with web search), Kimi-k3 (with web search), DeepSeek-v4-flash (with web search), ChatGPT 5.6 (with web search), and Gemini 3.1 Pro (with web search). </p><p class="banech-public-p bench-pt-24" data-v-e023de21> The evaluation uses a curated cross-jurisdiction patent family dataset of 340 test samples. Each sample contains a problem statement and a standard answer: the family set of X references cited by examiners across different patent offices. This design creates a practical benchmark answer that closely reflects real-world novelty search requirements. </p><p class="banech-public-p bench-pt-24" style="display:block;" data-v-e023de21> Benchmark results show that Patsnap's Novelty Search AI Agent remains the leading tool, with an <strong style="font-weight:700;" data-v-e023de21>85% X Hit Rate</strong> and a <strong style="font-weight:700;" data-v-e023de21>37% X Recall Rate</strong> within the top 100 results. Among general-purpose AI tools, ChatGPT 5.6 performed best, reaching a <strong style="font-weight:700;" data-v-e023de21>71.76% X Hit Rate</strong> and a <strong style="font-weight:700;" data-v-e023de21>30.51% X Recall Rate</strong>. Overall, domain-specific AI tools still show clear value in patent novelty search. </p><p class="banech-public-p bench-pt-24" data-v-e023de21> The evaluation dataset is evenly distributed across IPC classifications, covering both mainstream technologies and niche domains. In terms of language, 68% of the data is in English and 32% of the data is in Chinese, ensuring the model performs well across multilingual patent content. </p><p class="banech-public-p bench-pt-24" data-v-e023de21> For receiving-office distribution, applications from United States (US) and China (CN) each make up about 32%, while those from the European Patent Office (EP) and WIPO (WO) each account for roughly 18%. This balanced mix reflects the different examination styles across major patent jurisdictions and ensures more realistic, globally representative evaluation. </p><div class="bench-border-div" data-v-e023de21><p data-v-e023de21>Language distribution of patent texts in 340 test samples</p><img style="width:100%;" src="https://static-official.zhihuiya.com/patsnap/image/2026/bench_img1.svg" alt="" data-v-e023de21><p style="margin-top:40px;" data-v-e023de21> Distribution of IPC samples across 340 test samples </p><img style="width:100%;" src="https://static-official.zhihuiya.com/patsnap/image/2026/bench_img2.svg" alt="" data-v-e023de21><p style="margin-top:24px;font-size:14px;" data-v-e023de21> Note: Percentages may not sum to 100% due to rounding to one decimal place. </p><p style="margin-top:40px;" data-v-e023de21>
1 Distribution of receiving offices for 340 test samples </p><img style="width:100%;" src="https://static-official.zhihuiya.com/patsnap/image/2026/bench_img3.svg" alt="" data-v-e023de21></div><p class="banech-color-p bench-pt-24" data-v-e023de21>1) X Hit Rate</p><p class="banech-public-p bench-pt-24" data-v-e023de21> Patsnapâs Novelty Search AI Agent successfully identified at least one relevant X document in 85% of test casesâan essential capability for speeding up decision-making in patent examination and early-stage R&D. </p><img class="banech-mt-60" style="width:580px;max-width:100%;display:block;margin-left:auto;margin-right:auto;" src="https://static-official.zhihuiya.com/patsnap/image/2026/bench_img4_new.svg" alt="X Hit Rate" data-v-e023de21><p class="banech-color-p" data-v-e023de21>2) X Recall Rate</p><p class="banech-public-p bench-pt-24" data-v-e023de21> Patsnapâs Novelty Search AI Agent retrieved 37% of all relevant X documents, enabling more thorough analysis and more informed patent claim drafting.<br data-v-e023de21> A high X Recall Rate is key during R&D planning and before filing a patent. Patsnapâs Novelty Search AI Agent helps teamsâwhether in-house researchers, patent professionals, or external agentsâfind more relevant X documents. This supports better technical decisions and stronger patent claims, increasing the chances of patent approval. </p><img class="banech-mt-60" style="width:580px;max-width:100%;display:block;margin-left:auto;margin-right:auto;" src="https://static-official.zhihuiya.com/patsnap/image/2026/bench_img5_new.png" alt="X Recall Rate" data-v-e023de21><p class="banech-color-p" data-v-e023de21>3) Typical Test Result Sample</p><p class="banech-public-p bench-pt-24" data-v-e023de21> In this test, the patent specification, or problem statement, was submitted to each AI tool. The returned results were evaluated against a predefined standard-answer set of X references. </p><p class="banech-public-p bench-pt-24" data-v-e023de21> The sample below shows how the benchmark evaluates the returned patent references. Green references are hits in the standard-answer family set, and the bottom rows calculate X Hit Rate and X Recall Rate directly from those hits. </p><div class="bench-border-div" style="margin-bottom:0;" data-v-e023de21><h2 style="font-size:40px;font-weight:400;" data-v-e023de21> A single-sample benchmark test </h2><img class="banech-mt-60" style="width:100%;" src="https://static-official.zhihuiya.com/patsnap/image/2026/bench_img9_723.svg" alt="" data-v-e023de21><img class="banech-mt-60" style="width:100%;" src="https://static-official.zhihuiya.com/patsnap/image/2026/bench_img10_new.svg" alt="" data-v-e023de21><p class="bench-sample-copy" data-v-e023de21> In this sample, Patsnap's Novelty Search AI Agent identified three of the four relevant patent families, achieving a 100% X Hit Rate and a 75% X Recall Rate. Claude Opus 4.8 returned eight results in this sample; its two hits appear at positions 3 and 8, identifying two of the four relevant families and achieving a 100% X Hit Rate and a 50% X Recall Rate for this sample. ChatGPT 5.6 identified two of the four relevant families, achieving a 100% X Hit Rate and a 50% X Recall Rate for this sample. Perplexity Pro identified one relevant family, while DeepSeek-v4-flash, Kimi-k3, and Gemini 3.1 Pro did not identify a relevant family in this example. </p><p class="bench-sample-copy bench-sample-copy-last" data-v-e023de21> These sample-level results are illustrative. The overall benchmark conclusions are based on the full dataset of 340 cross-jurisdiction patent family samples. </p></div></div><div class="banech-pt-100" data-v-e023de21><h2 class="banech-public-h2" data-v-e023de21>Future Research</h2><p class="banech-public-p" data-v-e023de21> Future benchmarks will continue to expand the dataset and refine the evaluation methods for greater accuracy, coverage, and representativeness. As the dataset grows, the benchmark will provide a more robust view of how AI tools perform in professional patent novelty search. </p></div></div></div>',1),X=[H],U={id:"benchContent2"},q=f('<div class="banech-content" data-v-e023de21><div class="w1280" data-v-e023de21><div class="banech-pt-100" data-v-e023de21><h2 class="banech-public-h2" data-v-e023de21>Methodology</h2><p class="banech-public-p" data-v-e023de21> The Patsnap PatentBench Novelty Search benchmark encompasses four key dimensions: </p><p class="banech-color-p bench-pt-24" data-v-e023de21> 1) Test samples to establish the âbenchmark valueâ </p><p class="banech-public-p bench-pt-24" data-v-e023de21>
1 Each test sample is a basic unit in the benchmark. It consists of a problem statement and a standard answer. Because patent tasks rarely have perfect one-line answers, domain experts construct standard-answer sets that closely approximate the ideal outcome for professional novelty search. </p><p class="banech-public-p bench-pt-24" data-v-e023de21> For this benchmark, X and Y references cited by examiners at different receiving offices were collected, deduplicated, and normalized by patent family. This creates a reusable reference set for evaluation and comparison. </p><p class="banech-color-p banech-mt-60" data-v-e023de21> 2) Datasets to create reliable and unbiased test results </p><p class="banech-public-p bench-pt-24" data-v-e023de21> The dataset contains 340 carefully selected cross-jurisdiction patent family samples. The samples are controlled for language and IPC distribution, helping the benchmark reflect real-world diversity across patent texts and technical fields. </p><img class="banech-mt-60" style="width:100%;" src="https://static-official.zhihuiya.com/patsnap/image/2026/bench_img11_723.svg" alt="Design Test Samples" data-v-e023de21><p class="banech-color-p banech-mt-60" data-v-e023de21> 3) Evaluation metrics - the core of benchmarking </p><p class="banech-public-p bench-pt-24" data-v-e023de21> Evaluation metrics are used to measure and compare performance. They can be single or combined indicators, carefully designed by experts to reflect the practical needs of patent professionals. </p><p class="banech-public-p bench-pt-24" data-v-e023de21> Novelty search aims to identify relevant prior art to determine whether a patent claim is truly new. It follows the general principles of traditional search logic, but with a specialized focus on patent validity. </p><p class="banech-public-p bench-pt-24" data-v-e023de21> The Patsnap PatentBench uses the following indicators to measure the quality of search results: </p><div class="bench-num-div" data-v-e023de21><img src="https://static-official.zhihuiya.com/patsnap/image/2026/bench_num_icon01.svg" alt="" data-v-e023de21><div data-v-e023de21><h2 data-v-e023de21>X Hit Rate</h2><p data-v-e023de21> Proportion of samples where a correct answer appears among the top 1, 3, or 5 results </p><img style="width:100%;" src="https://static-official.zhihuiya.com/patsnap/image/2026/bench_num_img01.svg" alt="" data-v-e023de21></div></div><p class="banech-public-p" style="justify-content:center;" data-v-e023de21><span data-v-e023de21></span>X Hit Rate: Proportion of samples where a correct answer appears among the top 1, 3, or 5 results. </p><div class="bench-num-div" style="width:605px;" data-v-e023de21><img src="https://static-official.zhihuiya.com/patsnap/image/2026/bench_num_icon02.svg" alt="" data-v-e023de21><div data-v-e023de21><h2 data-v-e023de21>X Recall Rate</h2><p data-v-e023de21> Percentage of correct answers found within the top 100 results </p><img style="width:100%;" src="https://static-official.zhihuiya.com/patsnap/image/2026/bench_num_img02.svg" alt="" data-v-e023de21></div></div><p class="banech-public-p" style="justify-content:center;" data-v-e023de21><span data-v-e023de21></span>X Recall Rate: Percentage of correct answers found within the top 100 results. </p><p class="banech-color-p banech-mt-60" data-v-e023de21> 4) Comparison AI tools with industry experts </p><p class="banech-public-p bench-pt-24" data-v-e023de21> AI tools are designed to support and enhance the work of professionals, so benchmark comparisons should include both specialized AI agents and general-purpose models. This report compares Patsnap's Novelty Search AI Agent with six general-purpose AI tools that support web search. </p></div></div></div>',1),j=[q],G={id:"benchContent3"},M=f('<div class="banech-content" data-v-e023de21><div class="w1280" data-v-e023de21><div class="banech-pt-100" data-v-e023de21><h2 class="banech-public-h2" data-v-e023de21> Patsnap Novelty Search AI Agent <br data-v-e023de21>Use Cases & Its Impact </h2><p class="banech-public-p" data-v-e023de21> Patsnap's Novelty Search AI Agent stands out in benchmark testing thanks to its domain-specific fine-tuning and advanced Retrieval-Augmented Generation (RAG) technology. </p><p class="banech-public-p bench-pt-24" data-v-e023de21> Built on an open-source base model, the agent has been systematically refined with specialized patent knowledge, allowing it to understand patent language and search logic. By integrating RAG, it c
1ombines real-time retrieval with generative capabilities, enabling high-quality and low-hallucination search results. </p><p class="banech-public-p bench-pt-24" data-v-e023de21> For IP professionals in corporations and patent firms, this translates into a productivity boost. Search, filtering, and ranking work that traditionally takes hours can be completed in minutes, allowing experts to spend less time on repetitive search tasks and more time on strategic analysis and decision-making. </p><p class="banech-public-p bench-pt-24" data-v-e023de21> R&D teams also benefit during early-stage project evaluations. Fast and effective novelty search helps teams identify non-novel ideas earlier, reduce wasted resources, and make better innovation decisions. </p></div></div></div>',1),L=[M],K={class:"bench-footer"},J=w(()=>e("p",null,[g(" Discover how the Novelty Search Agent "),e("br"),g("can increase your productivity by 10 fold ")],-1)),V=b({__name:"BenchmarkNoveltySearch",setup(n){const t=R(1),r=P(),{openSignInDialog:v}=T();function l(i){v({startFrom:"eureka_landingpage",registrationRedirectUrl:`${r}${i||"/home?from=eureka_landingpage"}`})}function c(i){t.value=i,import.meta.client&&window.scrollTo({top:0,behavior:"smooth"})}return u({title:"Patsnap PatentBench for Novelty Search"}),(i,a)=>(m(),y("main",D,[e("div",F,[e("div",z,[$,O,e("button",{type:"button",class:"bench-banner-btn",onClick:a[0]||(a[0]=o=>l("/ip/checking/#/novelty-check-report?start_from=eureka_landingpage"))}," Try Novelty Search AI Agent ")])]),e("div",E,[e("div",B,[e("ul",null,[e("li",{id:"benchTabTit1",class:d({"bench-active-li":s(t)===1}),onClick:a[1]||(a[1]=o=>c(1))}," PatentBench-Novelty Search ",2),e("li",{id:"benchTabTit2",class:d({"bench-active-li":s(t)===2}),onClick:a[2]||(a[2]=o=>c(2))}," Methodology ",2),e("li",{id:"benchTabTit3",class:d({"bench-active-li":s(t)===3}),onClick:a[3]||(a[3]=o=>c(3))}," Use Cases ",2)])])]),h(e("div",N,X,512),[[p,s(t)===1]]),h(e("div",U,j,512),[[p,s(t)===2]]),h(e("div",G,L,512),[[p,s(t)===3]]),e("div",K,[e("div",null,[J,e("button",{type:"button",onClick:a[4]||(a[4]=o=>l(""))}," Try Novelty Search AI Agent ")])])]))}});const W=I(V,[["__scopeId","data-v-e023de21"]]),_=n=>(A("data-v-4fffec73"),n=n(),x(),n),Y={class:"benchmark-detail"},Q={class:"bench-banner"},Z={class:"w1280"},ee=_(()=>e("p",{class:"bench-banner-date"},"March 2026",-1)),ae=_(()=>e("h2",null,"Patsnap PatentBench for Design FTO Search",-1)),te={class:"bench-list"},se={class:"w1280"},ie={id:"benchContent1"},ne=f(`<div class="banech-content" data-v-4fffec73><div class="w1280" data-v-4fffec73><div class="banech-pt-100" data-v-4fffec73><h2 class="banech-public-h2" data-v-4fffec73>Understanding Design Patent FTO</h2><p class="banech-public-p" data-v-4fffec73> Design FTO is the process of systematically searching granted design patents worldwide to determine whether a product's appearance infringes existing design patent rights. </p><p class="banech-public-p bench-pt-24" data-v-4fffec73><span data-v-4fffec73></span>R&D Phase â Evaluate design direction, avoid potential infringement risks </p><p class="banech-public-p" data-v-4fffec73><span data-v-4fffec73></span>Pre-Launch â Comprehensive infringement risk screening for products </p><p class="banech-public-p" data-v-4fffec73><span data-v-4fffec73></span>E-commerce Compliance â Verify product designs against target market patents </p><p class="banech-public-p" data-v-4fffec73><span data-v-4fffec73></span>Litigation & Invalidation â Search similar designs as evidence in proceedings </p></div><div class="banech-pt-100" data-v-4fffec73><h2 class="banech-public-h2" data-v-4fffec73>Key Findings</h2><p class="banech-public-p" data-v-4fffec73> This benchmark evaluates the performance of three AI tools: Patsnap's Design FTO Search AI Agent, ChatGPT 5.4 (with web search), and Gemini 3.1 Pro (with web search). The evaluation dataset of 261 samples is distributed across three major patent offices (CN, US, EU) and covers 26 Locarno (LOC) primary classifications. Two sample types are included: e-commerce infringement images (64.8%) and patent invalidation real-object/line-drawing pairs (35.2%), ensuring the model performs well across both high-frequency commercial scenarios and challenging cross-modal patent comparisons. For receiving-office distribution, applications from China (CN), the European Union (EU), and the United States (US) each account for roughly one-third. This balanced mix reflects the different examination standards across major design patent jurisdictions and ensures more realistic, globally representative evaluation. </p><div class="bench-border-div" data-v-4fffec73><p data-v-4fffec73>Image type distribution of 261 test samples</p><img style="width:100%;" src="https://static-official.zhihuiya.com/patsnap/image/2026/designfot_img1.svg" alt="" data-v-4fffec73><p style="margin-top:40px;" data-v-4fffec73>
1 Distribution of receiving offices for 261 test samples </p><img style="width:100%;" src="https://static-official.zhihuiya.com/patsnap/image/2026/designfot_img2.svg" alt="" data-v-4fffec73><p style="margin-top:40px;" data-v-4fffec73> Distribution of LOC classifications across 261 test samples (TOP 10) </p><img style="width:100%;" src="https://static-official.zhihuiya.com/patsnap/image/2026/designfot_img3.svg" alt="" data-v-4fffec73></div><p class="banech-public-p" style="display:block;" data-v-4fffec73> Benchmark results show that Patsnap's Design FTO AI Agent achieved a <strong style="font-weight:700;" data-v-4fffec73>77% High-Risk Patent Hit Rate </strong>and a <strong style="font-weight:700;" data-v-4fffec73>0.7 PRES Score </strong> within the top 200 resultsâsignificantly outperforming two leading general-purpose AI tools. </p><p class="banech-color-p bench-pt-24" data-v-4fffec73> 1) High-Risk Patent Hit Rate </p><p class="banech-public-p bench-pt-24" data-v-4fffec73> Patsnap's Design FTO AI Agent achieved a High-Risk Patent Hit Rate of 77%â21 to 83 times that of general-purpose models (Gemini 3.71%, ChatGPT 0.93%)âan essential capability for speeding up infringement risk identification in product clearance and e-commerce compliance. </p><p class="banech-public-p bench-pt-24" data-v-4fffec73> Performance was strongest at the EU office, where the agent reached a 92% Hit Rate andâthe highest among all three major offices (CN/US/EU)âdemonstrating particularly powerful search capabilities for European design patent scenarios. </p><div class="bench-num-div" data-v-4fffec73><div data-v-4fffec73><h2 data-v-4fffec73>High-Risk Patent Hit Rate</h2><p data-v-4fffec73> The percentage of tests with accurate hits in the top 200 results </p><img style="width:100%;" src="https://static-official.zhihuiya.com/patsnap/image/2026/designfot_img4.svg" alt="" data-v-4fffec73></div></div><div style="display:flex;width:100%;max-width:1000px;padding:38px;flex-direction:column;align-items:flex-start;gap:40px;border-radius:20px;border:1px solid #bcc2cc;background:#fff;margin:60px auto;box-sizing:border-box;" data-v-4fffec73><img style="width:100%;" src="https://static-official.zhihuiya.com/patsnap/image/2026/designfot_img5.png" alt="" data-v-4fffec73></div><p class="banech-color-p" data-v-4fffec73>2) PRES Score</p><p class="banech-public-p bench-pt-24" data-v-4fffec73> The PRES Score measures an AI tool's end-to-end infringement determination capabilityânot only retrieving infringing patents, but also correctly delivering infringement conclusions in the final report. It is critical for real-world business decisions, as it directly reflects the quality of the complete pipeline from search to determination. </p><p class="banech-public-p bench-pt-24" data-v-4fffec73> Patsnap's Design FTO AI Agent achieved a PRES Score of 0.7, helping teamsâwhether in-house IP counsel, patent attorneys, or e-commerce compliance officersâsurface the most relevant infringing patents faster. This supports quicker risk assessment and stronger defensive strategies, reducing the chance of costly infringement disputes. </p><div class="bench-num-div" data-v-4fffec73><div data-v-4fffec73><h2 data-v-4fffec73>PRES Score</h2><p data-v-4fffec73> PRES (Patent Retrieval Evaluation Score) by Magdy & Jones (2010). </p><img style="width:100%;" src="https://static-official.zhihuiya.com/patsnap/image/2026/designfot_img6.svg" alt="" data-v-4fffec73></div></div><p class="banech-color-p" data-v-4fffec73>3) Typical Test Result Sample</p><p class="banech-public-p bench-pt-24" data-v-4fffec73> In this test, a product image (the "test question") was submitted to each AI tool along with the target market (US). Their results were then evaluated against a confirmed infringing design patent USD869847S1 (the "model answer"). </p><p class="banech-public-p bench-pt-24" data-v-4fffec73> Patsnap's Design FTO AI Agent successfully identified the target infringing patent at rank #2 within 200 returned results, achieving a High-Risk Patent Hit Rate of 100% and a determination accuracy of 100%. </p><p class="banech-public-p bench-pt-24" data-v-4fffec73> By comparison, ChatGPT 5.4 returned only 6 results and failed to i
1dentify the target patent, resulting in a Hit Rate of 0%. Gemini 3.1 Pro returned 20 results and placed the target at rank #6, achieving a 100% Hit Rate but with far fewer candidate patents for comprehensive screening. </p><p class="banech-public-p bench-pt-24" data-v-4fffec73> These findings highlight that while general-purpose LLMs can attempt visual patent searches, they struggle with the specialized image-to-patent matching and comprehensive coverage required for design FTO. In comparison, domain-specific AI tools like Patsnap's Design FTO AI Agent offer superior accuracy and relevance, underscoring their essential role in design patent-focused workflows. </p><div class="bench-border-div" style="margin-bottom:0;" data-v-4fffec73><h2 style="color:#000;font-family:'PingFang SC', -apple-system, BlinkMacSystemFont, 2 'Segoe UI', sans-serif;font-size:32px;font-style:normal;font-weight:300;line-height:140%;" data-v-4fffec73> Design Patent Infringement Search Comparison Example </h2><img class="banech-mt-60" style="width:100%;" src="https://static-official.zhihuiya.com/patsnap/image/2026/designfot_img7.svg" alt="" data-v-4fffec73><img class="banech-mt-60" style="width:258px;margin:60px auto;" src="https://static-official.zhihuiya.com/patsnap/image/2026/designfot_img8.svg" alt="" data-v-4fffec73><img class="banech-mt-60" style="width:100%;" src="https://static-official.zhihuiya.com/patsnap/image/2026/designfot_img9-2.svg" alt="" data-v-4fffec73></div></div><div class="banech-pt-100" data-v-4fffec73><h2 class="banech-public-h2" data-v-4fffec73>Future Research</h2><p class="banech-public-p" data-v-4fffec73> Future benchmarks will focus on three key directions: </p><p class="banech-public-p bench-pt-24" data-v-4fffec73><span data-v-4fffec73></span>Cross-modal visual alignment enhancement: Improve hit rates for real-object and line-drawing scenarios and close the gap with e-commerce performance </p><p class="banech-public-p" data-v-4fffec73><span data-v-4fffec73></span>Continued dataset expansion: Add test samples from JP and KR markets with broader LOC classification coverage </p><p class="banech-public-p" data-v-4fffec73><span data-v-4fffec73></span>Wider model comparison: Incorporate more multimodal large models as testing baselines </p></div></div></div>`,1),ce=[ne],oe={id:"benchContent2"},re=f('<div class="banech-content" data-v-4fffec73><div class="w1280" data-v-4fffec73><div class="banech-pt-100" data-v-4fffec73><h2 class="banech-public-h2" data-v-4fffec73>Methodology</h2><p class="banech-public-p" data-v-4fffec73> The Patsnap PatentBench Design FTO benchmark encompasses four key dimensions: </p><p class="banech-color-p bench-pt-24" data-v-4fffec73> 1) Test samples to establish the "benchmark value" </p><p class="banech-public-p bench-pt-24" data-v-4fffec73> Each test sample is a basic unit in the benchmark, consisting of a "test question" (a product image plus its target market) and a "standard answer" (confirmed infringing design patents). Since ideal answers rarely exist in design patent tasksâlike a definitive list of all visually similar patents that constitute infringementâdomain experts must carefully define realistic questions and construct answers that closely approximate ideal outcomes, grounded in real patent examination practice and judicial precedent. </p><p class="banech-color-p banech-mt-60" data-v-4fffec73> 2) Datasets to create reliable and unbiased test results </p><p class="banech-public-p bench-pt-24" data-v-4fffec73> To ensure statistical validity, the benchmark dataset includes 261 carefully selected samples that reflect real-world diversity across receiving offices, image types, and industrial
2classifications. All samples are sourced from real infringement casesâe-commerce platform complaints and patent X-reference relationshipsârather than artificial simulations, and all infringement relationships have been rigorously verified by a human annotation team. </p><p class="banech-color-p banech-mt-60" data-v-4fffec73> 3) Evaluation metrics - the core of benchmarking </p><p class="banech-public-p bench-pt-24" data-v-4fffec73> Evaluation metrics are used to measure and compare performance. They can be single or combined indicators, carefully designed by experts to reflect the practical needs of patent professionals. </p><p class="banech-public-p bench-pt-24" data-v-4fffec73> Design patent FTO aims to identify visually similar patents that may constitute infringement, then determine whether infringement actually exists. It follows the principles of visual similarity search, but with a specialized focus on design patent lawâincluding holistic observation, prioritization of key design features, and the ordinary observer standard. </p><p class="banech-public-p bench-pt-24" data-v-4fffec73> The Patsnap PatentBench-Design FTO uses the following indicators to measure the quality of search results: </p><div class="bench-num-div" data-v-4fffec73><img src="https://static-official.zhihuiya.com/patsnap/image/2026/bench_num_icon01.svg" alt="" data-v-4fffec73><div data-v-4fffec73><h2 data-v-4fffec73>High-Risk Patent Hit Rate</h2><img style="width:100%;margin-top:2vw;" src="https://static-official.zhihuiya.com/patsnap/image/2026/designFto-formula1.png" alt="" data-v-4fffec73></div></div><p class="banech-public-p" style="justify-content:center;" data-v-4fffec73><span data-v-4fffec73></span>High-Risk PatentHit Rate: Proportion of test samples where the confirmed infringing patent appears in the top K results. The standard answer may contain multiple infringing patents; a hit is counted when at least one is found. </p><div class="bench-num-div" data-v-4fffec73><img src="https://static-official.zhihuiya.com/patsnap/image/2026/bench_num_icon02.svg" alt="" data-v-4fffec73><div data-v-4fffec73><h2 data-v-4fffec73>PRES Score</h2><img style="width:100%;margin-top:2vw;" src="https://static-official.zhihuiya.com/patsnap/image/2026/designFto-formula2.png" alt="" data-v-4fffec73></div></div><p class="banech-public-p" style="justify-content:center;" data-v-4fffec73><span data-v-4fffec73></span>PRES Score: Measures end-to-end infringement determination quality. PRES (Patent Retrieval Evaluation Score), proposed by Magdy & Jones (2010), evaluates not only whether the infringing patent was found, but how highly it was ranked in the final outputâdirectly reflecting the complete pipeline from search to determination. </p><p class="banech-color-p banech-mt-60" data-v-4fffec73> 4) Comparison AI tools with industry experts </p><p class="banech-public-p bench-pt-24" data-v-4fffec73> AI tools are designed to support and enhance the work of professionals, so direct comparisons with established baselines are essential. This benchmark compares Patsnap's Design FTO AI Agent with two leading general-purpose modelsâChatGPT 5.4 (OpenAI, with web search) and Gemini 3.1 Pro (Google, with web search)âto establish baselines and better understand the capabilities of specialized versus general-purpose AI in design patent infringement tasks. </p></div></div></div>',1),le=[re],pe={id:"benchContent3"},de=f('<div class="banech-content" data-v-4fffec73><div class="w1280" data-v-4fffec73><div class="banech-pt-100" data-v-4fffec73><h2 class="banech-public-h2" data-v-4fffec73> Patsnap Design FTO AI Agent <br data-v-4fffec73>Use Cases & Its Impact </h2><p class="banech-public-p" data-v-4fffec73> This benchmark makes one thing clear: design patent infringement analysis is a fundamentally different challenge from what general-purpose AI was built to solve. Patsnap's Design FTO AI Agent achieves its leading performance by combining visual and textual feature fusion, iterative multi-round retrieval, and infringement reasoning grounded in design patent law. These tightly integrated capabilities create a level of domain expertise that general-purpose models cannot easily replicate. </p><p class="banech-public-p bench-pt-24" data-v-4fffec73> For IP professionals in corporations, law firms, and e-commerce operations, this translates into a powerful productivity boost. Tasks like screening product images against global design patent databasesâonce requiring days of manual visual comparisonâcan now be completed in minutes with significantly higher accuracy. This shift allows experts to spend less time on repetitive search work and more on strategic risk assessment and decision-making. </p><p class="banech-public-p bench-pt-24" data-v-4fffec73>
2 PatentBench-Design FTO will continue to evolve alongside the technology it measuresâgrowing its dataset, covering more jurisdictions, and refining its evaluation criteria to keep pace with the industry's advancing capabilities. </p></div></div></div>',1),he=[de],fe={class:"bench-footer"},me=_(()=>e("p",null,[g(" The Design FTO Agent catches 77% of high-risk"),e("br"),g("patents that general-purpose AI misses ")],-1)),ve=b({__name:"BenchmarkDesignFto",setup(n){const t=R(1),r=P(),{openSignInDialog:v}=T();function l(i){v({startFrom:"eureka_landingpage",registrationRedirectUrl:`${r}${i||"/home?from=eureka_landingpage"}`})}function c(i){t.value=i,import.meta.client&&window.scrollTo({top:0,behavior:"smooth"})}return u({title:"Patsnap PatentBench for Design FTO Search"}),(i,a)=>(m(),y("main",Y,[e("div",Q,[e("div",Z,[ee,ae,e("button",{type:"button",class:"bench-banner-btn",onClick:a[0]||(a[0]=o=>l("/ip/checking/#/design-fto?start_from=eureka_landingpage"))}," Try Design FTO AI Agent ")])]),e("div",te,[e("div",se,[e("ul",null,[e("li",{id:"benchTabTit1",class:d({"bench-active-li":s(t)===1}),onClick:a[1]||(a[1]=o=>c(1))}," PatentBench-Design FTO ",2),e("li",{id:"benchTabTit2",class:d({"bench-active-li":s(t)===2}),onClick:a[2]||(a[2]=o=>c(2))}," Methodology ",2),e("li",{id:"benchTabTit3",class:d({"bench-active-li":s(t)===3}),onClick:a[3]||(a[3]=o=>c(3))}," Use Cases ",2)])])]),h(e("div",ie,ce,512),[[p,s(t)===1]]),h(e("div",oe,le,512),[[p,s(t)===2]]),h(e("div",pe,he,512),[[p,s(t)===3]]),e("div",fe,[e("div",null,[me,e("button",{type:"button",onClick:a[4]||(a[4]=o=>l(""))}," Try Design FTO AI Agent ")])])]))}});const ge=I(ve,[["__scopeId","data-v-4fffec73"]]),ue={key:2,class:"p-10 text-center"},Te=b({__name:"[slug]",setup(n){const t=S(),r=C(()=>String(t.params.slug||""));return u(()=>({title:r.value==="design-fto"?"Patsnap PatentBench for Design FTO Search":"Patsnap PatentBench for Novelty Search"})),(v,l)=>{const c=W,i=ge;return s(r)==="novelty-search"?(m(),k(c,{key:0})):s(r)==="design-fto"?(m(),k(i,{key:1})):(m(),y("div",ue,"Benchmark page not found."))}}});export{Te as default};
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.