慣性聚合 高效追讀感興趣之博客、新聞、科技資訊
閱原文 以慣性聚合開啟

推薦訂閱源

SecWiki News
SecWiki News
S
Secure Thoughts
N
News and Events Feed by Topic
NISL@THU
NISL@THU
WordPress大学
WordPress大学
H
Hacker News: Front Page
cs.CL updates on arXiv.org
cs.CL updates on arXiv.org
GbyAI
GbyAI
Scott Helme
Scott Helme
Hacker News: Ask HN
Hacker News: Ask HN
S
Security @ Cisco Blogs
J
Java Code Geeks
T
The Blog of Author Tim Ferriss
Attack and Defense Labs
Attack and Defense Labs
The Register - Security
The Register - Security
Y
Y Combinator Blog
Latest news
Latest news
小众软件
小众软件
Know Your Adversary
Know Your Adversary
P
Proofpoint News Feed
P
Palo Alto Networks Blog
C
Cyber Attacks, Cyber Crime and Cyber Security
H
Help Net Security
C
Comments on: Blog
The GitHub Blog
The GitHub Blog
T
Tailwind CSS Blog
博客园 - 聂微东
阮一峰的网络日志
阮一峰的网络日志
K
Kaspersky official blog
MongoDB | Blog
MongoDB | Blog
宝玉的分享
宝玉的分享
Google DeepMind News
Google DeepMind News
C
CERT Recently Published Vulnerability Notes
V
Visual Studio Blog
M
MIT News - Artificial intelligence
F
Full Disclosure
T
Tor Project blog
F
Fortinet All Blogs
B
Blog RSS Feed
博客园 - 三生石上(FineUI控件)
A
Arctic Wolf
量子位
Last Week in AI
Last Week in AI
www.infosecurity-magazine.com
www.infosecurity-magazine.com
博客园_首页
T
The Exploit Database - CXSecurity.com
P
Proofpoint News Feed
酷 壳 – CoolShell
酷 壳 – CoolShell
The Hacker News
The Hacker News
G
Google Developers Blog

DEV Community

Centralized procurement D365: global address book + vendors Why I Can't Stop Thinking About Google's New A2A Protocol Perovskite cell scaps simulation analysis ¿Qué significan esas letras del CVSS? Guía para entenderlo de una vez scrcpy Integration in a Tauri App — Android Screen Mirroring on Mac Shopify theme editor: design tokens merchants can edit Dataverse security restructure: lessons applied too late Floatkit is live now!!! SimGemma: Democratizing STEM Education with Offline-First AI Simulations What to monitor in an AI agent before you launch (and after) The precedence rule deserves a name Diffusion Language Models Are Here: Deep Dive into NVIDIA's Nemotron-Labs DLM Architecture [Boost] I Still Remember the Day Our Server Stall Almost Killed the Product Launch AI Agents Need More Than Fact-Checking Evaluation & Benchmark Results 5 things `flutter_gemma` doesn't tell you about shipping Gemma 4 on Android How I Indexed 2,000 Claude Code Skills (And What the Install Data Says About AI Coding in 2026) Architecting Instant Micro-Loans: Data Pipelines and KYC Automation Bulk Rename Files from the Command Line with Python Virtual SOC Analyst This project was an absolute blast to build for the Hermes Agent Challenge. If you found the architecture layout or the local automation breakdown helpful, please drop a ❤️ or a 🦄 on the post! Let me know if you want me to write a follow-up guide specifi How I built a fully offline AI assistant on Android with Gemma 4 E2B How I Got Users to Willingly Wait 1 Minute for an API Call (Without Over-Engineering) What Training Exists for Security Professionals Learning AI and Data Science? Easier Bets to Get Early Customer Validation and VC Attention django-deploy-probes — deployment probe endpoints for Django AI Won’t Replace Developers. Weak Thinking Will. Building Micro Agents as Production-Grade Microservices Why Open-Weight Models Like Gemma 4 Are the Future of Secure Backend Architecture I lost 3 enterprise clients in one night because of a GitHub repo. So I built a tool to make sure it never happens again. Building a Local AI SOC Analyst on an M1 MacBook Pro Carelo: A Modern Dual-Pane File Manager for Linux AI API Pricing in 2026: What You Actually Pay for GPT-5.5, Claude Opus, Gemini, and 20+ Models I Built a Free Offline-First Event Operations Platform at 13. Here's Why the Architecture Is Different. I Built an AI Tools Directory. These 10 Lessons Hurt the Most. The "Disappearing Zero": Handling Numeric Inputs in React Native Forms I Finished My Local AI Coding Agent After 5 Months — Eve Agent V2 Unleashed published Neuropsychology: What Brain Damage Reveals About the Mind Shipping Gemma 4 speech recognition in a Windows .NET desktop app: a 5-variant model-selection tour Engineers Don’t Fail Technical Interviews Because They’re Bad at Tech — They Fail Because They Ignore Communication The 20% of ML theory that earns its keep in production WeiQi - (Go) game based productivity tool Diário de dev #1: o que 15 minutos desbloqueou 远程安装及部署应用 · 用户配合指南 The Complete Guide to API Design in 2026: REST, GraphQL, and tRPC in Production 🐍 Flask Python Structured Logging — What Most Miss in Production CSS in 2026: Container Queries, Cascade Layers, and the End of Utility-Class Bloat TypeScript 5.5 — The Features That Actually Matter for Production Code Database Migration Strategies That Actually Work in Production
忠信之门:众团队多所忽略之代理层
SapotaCorp · 2026-05-24 · via DEV Community

去岁,一B2B SaaS之众,得客怒函。客之属吏,询公司AI助者,其计是否含SSO。助者对曰然。客之司事者,费二日之力,欲成其制,诉诸支持,乃知助者之误。盖SSO乃在企级之计,而客所持者,乃专业之计也。

助者索检文牍,未得确证何级含单点登录之事,乃据似可信者,依训练之数据,作流畅之答。用户无从知其乃幻生之语也。

此非"更优之模"。若LLM更巨,则同此不足之境,必更自信地妄生幻象。此之解,实乃本应自始即存之层:一忠信之关,用以验察代理之应,是否确根植于所取之境,方送诸用户。

此乃生产AI智能体最高杠杆干预之一。众团队多弃之,盖因其失效模式未显,直至客诉方知。

忠实度实测何物

忠实度一问而已:智能体之应,其言是否得乎其检索之境所证?

若使者在知识库中查之,得见"Pro层级含基本功能X、Y、Z。Enterprise层级则除X、Y、Z外,复有进阶功能A、B、C,兼有SSO",则应答言"尔之Pro计划含SSO"者,非其本真。所索之境,实不支此言。

此异于"应答是否正确"之问。正确需凭实据,忠实唯取所索之文。无需人介入,可自验之。

匠者:自应中析原子之断,核之与所取之境,计其分。若分不及阈,则应中非信,不可发。

门之实然运作何如

其理明矣:

  1. 代理据所索之境生成应答;
  2. 别以LLM(判者)析应答之原子主张;
  3. 每一主张,判者察所索之境是否佐之;
  4. 忠实之分数,乃所佐主张之比也。
  5. 若分数不及阈限(吾等默认为0.85),则应拒之。
  6. 或重试以改之境,或返"此非有据可答"

如Ragas之框架,直能为之。亦可用一LLM之呼,以结构化之提示,自建之。判者之模,不必为生产之模。吾辈常以GPT-4o-mini或Claude Haiku为判者,以节成本;其准确已足为此任。

何以此能及模型之大小

大模之幻,非减也,反增其信。设境不足,GPT-4o所出之谬,文辞愈工,结构愈整,声威愈重,较GPT-3.5尤甚。

忠信之关,作用于模型之上层。其不察模型之自信何如,惟察应答之主张,能否溯源于所取之境。

团队稽核,忠信之关得察顾客尝报为误之应答约四成。此多者,易模型而增价,亦不能察也。

其阈之问

定信实之阈于何处,乃产之决,非技之决也

  • 九五以上:甚严。用于律法之谏、医道之讯、财策之荐、规约之遵。其费者,增"吾不能答"之应耳,此乃高风险域之宜价也
  • 自0.85至0.95:此乃B2B SaaS之常规生产设置。能捕捉多数自信之幻觉,而不拒合法之应答,纵其有微末未获支持之辞藻。
  • 自0.70至0.85:此则更为宽纵。宜用于用户可自验之内部工具,或初阶产品,盖因拒应过多,则损用户体验。
  • 低于0.70:此实已失效。不荐用于面向客户。

吾所与共事之团队,业于 B2B SaaS。初立阈值为 0.88,察拒率(约应答之六成),一周后觉拒率过峻,有损用户体验,遂调阈为 0.85。

门坏何为

当应答未通过忠实度检验时,代理有三项选择:

重试,增以文脉。使者在败绩之问再索。或初索未足,再巡乃显阙文。一重可,再重止。过此则勿复。

返曰:"此非敢决也。"诚陈其限。显一实弊,可由众解(文牍不足,诘问不明),胜于自信而误。

转托于人。使者呈问于司事,附以所索之文。宜于接客之制,盖"不知"非可终局也。

诸生产之队,皆运其三。初试(价廉,常可解),次用诚然“吾不知”(低局可容),大事或复问,则升之。

吾所运于队者

原制乃客服之吏,兼用RAG于文牍。吾增:

  1. 每应皆验其诚,以GPT-4o-mini为判之模。
  2. 產生回應之閾值定於0.85。
  3. 若初回應未合檢核,則增強檢索重試一次。
  4. 兩次檢核皆失敗之回應,則以誠實之退讓應之(「吾等文獻中未載此項特定資訊。君欲否命人手處理?」)。
  5. 每次失敗之忠實檢核皆記錄之,俾團隊得審察趨勢,增進文獻涵蓋。

客报之误答,首月减六成。诚信之阍,未增正误,惟止系统自信而贻误于客。诚实之"吾不知"初引忧惧(客将怏怏乎?),然终受嘉纳。客择"吾不知"胜于误答,纵欲速答也。

不期之益,乃失检之录。今团队得列每问,文牍不能确答者。此遂为文牍积案。半载之内,客报之患,较之关前基线,已减八成,其半由关之设,其半由文牍之进,关之所显也。

门不足时

忠信之关,可防一弊:言而无据,境迁而失。然不察二事:

  • 境之谬也。若 RAG 之流,引文有误,则应答必从谬本。须审之。
  • 境之陈也。忠信于六月前之文,今已陈腐。须版本之序,时鲜之察。
  • 言辞微妙而谬误。虽事由境生,然推论之间,失其真谛。需更审慎之评,或需人力复审之。

此门虽为生产之要,然非其全备。乃最高杠杆之单干预,然非唯一干预之策。

桃子之荐

凡司事之属,掌实问者(如客问、内知、合规,误则损):

  • 于应答之途设诚信之关
  • 用廉判之模(GPT-4o-mini、Haiku),以节费
  • 初定阈于0.85,依拒率调之
  • 施重试一遭、诚退之策
  • 每次失败皆录以备档之进

基础设施之费,约计每应答一元之万分之一。客所报之误,首月通常减四至六成

此非生产之B2B代理者可择,乃化演示为产品之层也

若尔代理自信而误

若君之团队有客报AI助人答非所问,而"吾将易以更优之模"未能解之,则所缺之层,殆必为忠实之验也。

萨波塔提供一周之约,于尔既有之代理添忠诚之验,校准阈限以尔往昔之报,并运重试与回退之理作为可行之 PR。吾等已为之于客服代理、内蕴知识库及合规之器。

经由人工智能工程之页相接。 並舉數例,以見吾代理所給非是之答。診斷之對話,常能顯信實之隙與記錄之罅,此門將助以揭示之。