AI 模型和平台
Perplexity 新增现场网络定位到它的决定 API
在 Google 上的首选来源中添加 Unite.AI
2026年11月4日的Perplexity 宣布表示其"决定API"可立即使用网络定位，在模型返回概率前添加可选的现场网络搜索。 其特点是，在培训后，答案可以改变的决定，包括个人的行政作用、软件包的最新发布，或活动是否仍然安排。
决定 API 接受内容为状态，返回结构化的数字解答而不是流言。 " 是 " 或 " 不 " 问题产生 " 是 " 的概率；" 选择 " 问题产生 " 每一个确定选项 " 的概率；" 分数问题 " 产生 " 分数 " 的概率。 开发者可以在分类，路由和其他自动化决定中将这些值作为阈值。 状态可以包含文本，JSON或图像，发布也允许将模型指向特定页面的URL.
模式答案前的现场搜索
在模式调用前在 Perplexity 的 API 网关中设置网格搜索模式。 网关从问题和状态文本开头每个问题构建一个查询，并行运行搜索，合并结果并删除重复的URL. 产生相同查询的问题共享搜索。 网关然后将页面标题、 URL 和文本附加到状态中作为证据。 答案仍然是基于概率的，而搜索-结果字段则让开发者检查模型所看到的。
搜查有明确的限制。 多达10次同时运行，每次搜索限时1秒，总共3秒。 不搜索图像。 查询到 Perplexity 搜索 API 可以同时包含问题文本和状态的打开部分，使发送到搜索部分整合隐私考虑的数据。
搜索失败并不一定会阻止决策响应。 如果只有一些搜索成功，模型会使用到达的结果。 如果没有成功，它会从原始状态回答并返回一个空搜索结果阵列。 因此，即使现场证据没有起到作用，答复也可以保留正常的回答结构。
准确性收益和计量差异
Perplexity报道了从2026年10月新闻和罕见的维基百科事实中抽取的328名真假声称测试同一决定模型。 每项真正的索赔都有一个不同细节的虚假对应物，没有提供证据。 准确度为56.4%，没有底线，97.3%，错误从143降至9。 中间延缓度从139毫秒增加到366毫秒，第95百分位延缓度从514毫秒增加到742毫秒，使速度成本与准确度提高值相差。
平行的Web Systems 称其PWS-决定的基准测试包含300个闭门问题：100个公开并附有参考答案，200个非公开。 Scoring使用精确的匹配，而该集是按难度而不是估计平均流量的准确性来整理的。 平行领导板报告指2026年10月29日评估的3次运行的准确性和抽样标准偏差，包括51.00%的Perplexity决定的公设精度。
Perplexity pplx-decider-v1.2 网络定位在2026年11月4日的一次运行中正确回答了100个公开提问中的86个，它描述为该公开发布领导板的公设精度最高。 其客户在这些100个问题上测得的中位纬度为370毫秒，95百分位纬度为685毫秒。 平行的潜伏度测量覆盖了整个300个问题集的三连串的完整答案，因此两组潜伏度数字有不同的范围。
定价和一体化
有根据的请求可以包括多达100个问题。 每个完成的搜索费用为0.001美元，对100个问题请求的搜索费用不超过0.10美元，但检索到的证据也作为输入符号计费，计算到256K输入限额。 标准pplx-decider-v1.2模型每百万个输入符花费0.02美元，pplx-decider-flash-v1.2每百万个输入符花费0.01美元；输出符免费。
问题分组会影响最终的象征性法案。 在标准模型上，当输入限制提问为单独分组时，证据再次被计算。 Flash独立地评价每个问题，并且每次再次向状态和证据计费。 因此，地面费用取决于搜索次数、检索文本的数量以及选定示范批次如何提出问题。
整合目前有一个打字提示。 Perplexity的Python库 0.44.0 或更晚，以及 TypeScript SDK 0.39.0 或后支持 API，但其类型尚未曝光模式和搜索-结果。 Python通过extra body通过模式，而TypeScript则在请求体中发送。
Perplexity建议的工作量包括检查账号在过去90天里是否有融资资金，观看稳定发布新版的依赖性，并在agent预订之前确认一个事件。 在这些工作流程中，将新增可检查的当前证据打下基础，从而有可能使软件达到阈值或审查。 权衡的范围超出了准确性：延迟、象征性使用、搜索费和搜索中任何缺失的证据都决定了返回的概率。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
AI Models & Platforms
Perplexity Adds Live Web Grounding to Its Decisions API
Add Unite.AI to your preferred sources on Google
Perplexity announced on November 4, 2026 that web grounding is available immediately for its Decisions API, adding optional live web searches before the model returns probabilities. The feature targets decisions whose answers can change after training, including a person’s executive role, a software package’s latest release or whether an event remains scheduled.
The Decisions API accepts content as state and returns structured numeric answers instead of prose. A yes-or-no question produces the probability of yes; a choice question produces a probability for every defined option; and a score question produces probabilities across an ordered rubric plus an expected score. Developers can use those values as thresholds in classification, routing and other automated decisions. State can contain text, JSON or images, and the launch also allows a URL that points the model toward a specific page.
Live Search Before the Model Answers
Setting mode to web_grounding activates Fast Search in Perplexity’s API gateway before the model call. The gateway builds one query per question from the question and the beginning of the state text, runs searches in parallel, merges results and removes duplicate URLs. Questions producing the same query share a search. The gateway then appends page titles, URLs and text to the state as evidence. The answers remain probability-based, while a search_results field lets developers inspect what the model saw.
The searches have explicit limits. As many as 10 run concurrently, with a one-second limit for each search and three seconds overall. Images are not searched. Queries go to the Perplexity Search API and can contain both question text and the opening portion of state, making the data sent to search part of the integration’s privacy considerations.
Search failures do not necessarily stop a decision response. If only some searches succeed, the model uses the results that arrived. If none succeeds, it answers from the original state and returns an empty search_results array. A response can therefore retain the normal answer structure even when live evidence did not contribute to it.
Accuracy Gains and Measurement Differences
Perplexity reported testing the same decision model on 328 true-or-false claims drawn from October 2026 news and rare Wikipedia facts. Each true claim had a false counterpart differing in one detail, and no evidence was supplied in state. Accuracy was 56.4% without grounding and 97.3% with it, while errors fell from 143 to nine. Median latency increased from 139 ms to 366 ms, and 95th-percentile latency from 514 ms to 742 ms, placing a measurable speed cost beside the accuracy gain.
Parallel Web Systems said its PWS-Decisions benchmark contains 300 closed-ended questions: 100 public with reference answers and 200 private. Scoring uses exact match, and the set was curated for difficulty rather than to estimate accuracy on average traffic. Parallel’s leaderboard reports mean accuracy and sample standard deviation across three runs evaluated October 29, 2026, including 51.00% public-set accuracy for Perplexity Decisions.
Perplexity said pplx-decider-v1.2 with web grounding answered 86 of the 100 public questions correctly in a November 4, 2026 run, which it described as the highest public-set accuracy against that published leaderboard. Its client measured median latency of 370 ms and 95th-percentile latency of 685 ms on those 100 questions. Parallel’s separate latency measurements cover complete answers across three runs of the full 300-question set, so the two sets of latency figures have different scopes.
Pricing and Integration
A grounded request can include up to 100 questions. Each completed search costs $0.001, putting search charges at no more than $0.10 for a 100-question request, but retrieved evidence is also billed as input tokens and counts toward the 256K input limit. The standard pplx-decider-v1.2 model costs $0.02 per million input tokens and pplx-decider-flash-v1.2 costs $0.01 per million; output tokens are free.
Question grouping affects the final token bill. On the standard model, evidence is counted again when input limits force questions into separate groups. Flash evaluates every question independently and bills the state and evidence again each time. The cost of grounding therefore depends on the number of searches, the amount of text retrieved and how the selected model batches questions.
Integration currently has a typing caveat. Perplexity’s Python library 0.44.0 or later and TypeScript SDK 0.39.0 or later support the API, but their types do not yet expose mode and search_results. Python passes the mode through extra_body, while TypeScript sends it in the request body.
Perplexity’s suggested workloads include checking whether accounts have raised funding in the past 90 days, watching dependencies for stable releases newer than a pinned version and confirming an event is still scheduled before an agent books it. In those workflows, grounding adds inspectable current evidence to a probability that software can threshold or review. The tradeoff extends beyond accuracy: latency, token usage, search fees and any missing evidence in search_results all shape what the returned probability represents.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.