DeepSeek packages web search, document parsing, and voice capabilities as open APIs
English text is being prepared.
DeepSeek has packaged the web search, PDF and Office document parsing, web link scraping, security screening, as well as OCR, ASR, and TTS capabilities from its official App into open APIs. Currently, these are only available for deepseek-flash (V4.1-Flash) and are enabled via parameter switches when calling. On pricing: web search and link scraping cost $0.75 per 1,000 calls, document parsing costs $0.225 per 1,000 pages, speech recognition costs $0.058 per hour, and off-peak input tokens cost $0.15 per million. Meanwhile, DeepSeek has open-sourced the chat-demo frontend project on GitHub under the MIT license; you can run it locally by double-clicking index.html, and it can import chat history exported from the App.
Why read this
AI score 75
The original article presents DeepSeek's newly opened capabilities—search, document parsing, voice, and more—alongside specific call pricing, making it easy for developers to estimate the cost of building their own applications.
Topics
View original postFully open-sourced!! I used Tencent WorkBuddy to run a complete Hong Kong and US stock AI investment research workflow: reviewing holdings, breaking down financial reports, and relentlessly tackling data hallucinations, with a 4-step prompt included—I suggest copying it directly. As long as you regularly read financial reports and manage holdings, you can directly use this human-AI collaboration as an AI investment research SOP workflow for future investment research reviews: the core logic is not to have AI predict tomorrow's gains or losses like a fortune teller, but to treat it as a cold auditor that thoroughly verifies the chain of evidence behind every number. Perhaps only after calculating this cost will everyone understand how important this system is: What you fear most when reading financial reports is never misjudging the major trend, but rather AI fabricating fake data, or mixing in accounting standards restated by the company in the following year, and your real-money decisions worth tens of thousands or even hundreds of thousands instantly hit a landmine; This workflow illuminates all the hidden pitfalls, and even if you lose money, you can review it clearly! I have always believed that what AI should really do in investment research is the most labor-intensive, dirty, and tedious work. I summarize it as conducting a background check on numbers, every piece of data must pass three checks: → Where it came from: which section of the annual report, and clicking it must trace back to the original text → What it is compared with: whether the accounting standards across periods are the same, and whether they can be directly subtracted → Who said it: whether it is a verified fact or the management's own explanation. Numbers whose background cannot be clearly checked are not allowed into the decision conclusions. The video contains the complete screen recording (all demonstrations use anonymized synthetic data), and the entire test is divided into four extremely clean steps: Build a holdings dashboard; if cost is not provided, resolutely do not guess. Drop in the raw holdings and price snapshot, and convert weights into Hong Kong dollars by currency. For positions without recorded cost, honestly mark the cost as missing, and prevent model hallucination interpolation. Dynamic concentration risk control. After adding buy records, automatically recalculate market value. When a single holding is diluted to 29.18% due to adding to the position, the concentration warning is automatically withdrawn, and the before and after of the rebalancing are clear at a glance. Investigate the accounting standard traps of annual report restatements. Break down business segment growth, and when the company restates standards across periods, numbers are precisely anchored to the original annual report section, and clicking it shows the restatement notes, avoiding misleading fake increments. Ultimate follow-up question, three-column physical isolation. When encountering abnormal accounts receivable, force it to break down into three columns: verified facts, the company's own explanations, and judgments still requiring manual verification, completely isolating management's empty promises from objective facts. Finally, I want to tell everyone that investment research tools are never crystal balls; what is more terrifying than losing money is losing according to no system of your own. Let's encourage each other! The original text of the 4-step prompt and experience guide are organized in the comments section; as usual, help yourselves👇
View original
全盘开源!!我用腾讯 WorkBuddy 跑通了一套港美股 AI 投研工作流:复查持仓、拆解财报、死磕数据幻觉,附 4 步提示词,建议直接照着抄。 只要你平时看财报、管持仓, 以后做投研复盘可以直接拿这套人机协作当AI投研sop工作流:核心逻辑不是去让 AI 算命猜明天涨跌,而是把它当成冷酷的审计员,把每一笔数字背后的证据链彻底查清。 可能算过这笔账大家才知道这套体系有多重要: 平时看财报最怕的从来不是看错大趋势, 而是 AI 瞎编一个假数据, 或者把公司第二年重列的会计口径混着算, 你几万甚至几十万的实盘决策就当场踩雷; 而这套工作流把所有暗坑全部照亮, 亏了钱也能清清楚楚复盘! 我一直觉得AI 在投研里真正该干的就是最费人工的脏活累活。 我把它总结为给数字查户口, 每一笔数据都必须过三关: → 哪来的:出自年报哪一节,点开必须能追溯到原文 → 跟谁比:跨期会计口径一不一样,能不能直接相减 → 谁说的:是已经核实的事实,还是管理层自己的解释 户口查不清的数字,一个都不准进决策结论。 视频里是完整录屏(演示全部使用匿名合成数据), 整套实测分为极度干净的四步: 做持仓看板,没给成本坚决不瞎猜 丢入原始持仓和价格快照,按币种折算港币权重。遇到未录入成本的头寸,老老实实标出成本缺失,杜绝模型幻觉插值。 集中度动态风控 补录买入记录后自动重新折算市值。 单只持仓因加仓被稀释到 29.18% 时, 集中度预警自动撤销,调仓前后一目了然。 排查年报重列的口径陷阱 拆解业务分部增长, 遇到公司跨期重列口径时, 数字精准锚定在原始年报章节, 点开就能看到重列附注,避免假增量误导。 终极追问,三栏物理隔离 遇到应收账款异动, 强制要求它分三栏拆解: 已验证事实、公司自己的解释、还需人工核实的判断,把管理层画饼与客观事实彻底隔绝。 最后想跟大家说,投研工具从来不是水晶球,比亏钱更可怕的是没有按照自己的体系去亏,共勉! 4 步提示词原文与体验指引整理在评论区了,老规矩大家自取👇
Body
卧槽,全体起立!!梁圣牛逼!! DeepSeek 这次是真把压箱底的看家本领全掏出来了, 直接把官方 App 里那套联网搜索、读 PDF 文档、抓网页链接、安全检查加上语音识别合成,整套打包开放成了 API,连前端界面都用 MIT 协议直接开源了。
很多刚入门的兄弟可能没反应过来这事有多炸裂, 我用大白话给你算笔账你就懂了。
以前你想自己手搓一个像样的 AI 应用,模型本身只管聊天,你要让它联网搜资料,得自己去买搜索引擎接口,
要让它读合同和 PDF,得自己写代码解析文档再硬塞进上下文,要加上语音交互,
还得再去接一套识别和朗读接口,光写这堆胶水代码就能把人折磨掉半条命。
现在 DeepSeek 直接把这堆脏活累活全揽到自己服务端里干完了。
只要调用 deepseek-flash(也就是 V4.1-Flash),你在请求里加几个参数开关,
它就能边思考边深度搜网、直接吞进 PDF 和 Office 文件、自动读取消息里的网页链接,
甚至还把输入输出的安全筛查和独立的 OCR、ASR、TTS 全给你配齐了。
说白了,以前做 AI 应用是你自己去菜市场买菜、回家洗菜、切菜再开火炒,现在 DeepSeek 直接把切菜机、烤箱和洗碗机全给你接好了,你只管点菜和端上桌。
更离谱的是这个价格,联网搜索和抓链接一千次只要 0.75 美金,解析一千页文档只要 0.225 美金,
语音识别一小时只要 0.058 美金,配合低谷期每百万输入 0.15 美金的 token 价格,个人开发者几乎是零门槛白嫖级起步。
有意思的是,他们顺手在 GitHub 开源了一个叫 chat-demo 的前端项目,不用搭复杂环境,
本地双击 index.html 就能跑,还能直接把你在 DeepSeek App 里导出的历史聊天记录导进去无缝接着聊。
我一直觉得,普通人用 AI 搞副业、做一人公司,最怕的就是把时间全耗在重复造轮子上,
虽然目前这套满血工具链只先给 Flash 开放了,高频调用也得注意控制搜索次数免得账单超标,
但顺着这个趋势看,以后手搓 AI 产品的技术门槛,真的已经被压平到跟搭积木一样了。