中文精读
Dify早期的产品判断：把模型接上之后，谁来让应用持续有用
42章经单期精读 · AI工程与产品运营
嘉宾：张路宇，节目当期介绍为Dify.AI创始人｜主持：曲凯；尾声提问：制作人Celia
节目发布：2023年8月20日｜时长：74分9秒｜整理：2026年10月18日
原节目与官方介绍
接通接口以后，产品才开始接受检验
一个模型能回答问题，离一个业务能长期使用的应用有多远？2023年的这次访谈，把距离拆成了资料处理、提示调试、工具组织和上线后的持续运营。Dify希望提供的不只是调用模型的入口，也包括业务人员不断塑造应用的工作台。
曲凯在节目中披露自己参与投资Dify，也与团队存在合作关系。因此，关于团队执行力、中间层机会和竞品的评价，需要连同这个身份理解。本文保留当年的问题与方法，不将2023年的模型排名、经营数字或未来预测写成当前事实。
1. 五万应用说明接触面，不能直接说明五万次成功
节目线索：00:36–06:55。 路宇把LLMOps解释为大模型应用开发和持续运营的结合：工程师接好能力后，懂业务的运营、销售等人仍需修改知识和提示，让输出逐渐接近业务预期。传统的服务器是否可用，与回答是否真正有帮助，是两类需要同时处理的问题。
节目谈到当时约7000个GitHub星标、万余次Docker镜像拉取和四万余个云端应用，并以“五万多个应用”概括规模。镜像是用于分发部署软件的打包形式；一次拉取不能在本次整理中被当作一个独立、持续活跃的生产实例。星标也可能是收藏：有人觉得项目很酷，点一下就离开。
路宇建议结合镜像或软件包下载、搜索趋势等信号判断热度。编辑判断：这些数据可以帮助识别开发者兴趣，仍需要往后追问：应用有没有接入真实任务，多少用户持续使用，失败时谁处理，是否产生收入？数字所处的层次不同，不能互相代替。Dify当时的价值主张，恰恰是把关注从“做出来”延伸到“继续维护”。
2. 先分清缺指令、缺资料，还是默认行为不合适
节目线索：07:48–15:19。 访谈用三条路径解释应用工程化。Prompt是给模型的指令，例如纠正文法或判断情绪；检索增强是在回答前找来相关资料，再与问题一起交给模型；微调则通过训练样本调整模型的默认表现，例如让诗歌更接近指定风格。
节目用“曲凯今年做过哪些项目”举例：模型不知道私有或新近的信息，系统先检索项目资料，再组成较完整的问题。这里需要区分术语：Embedding是把内容转换成可用于比较的向量表示，检索、挑选资料和拼接上下文共同构成更完整的流程，不能把它简单等同于“把内容塞进提示”。
团队提到当时8K上下文的预算：用户问题、命中资料、历史内容与预期输出需要共同分配空间。重点不是背下这个历史容量，而是决定哪些资料值得占用空间、是否会漏掉关键条件。让另一个模型先做摘要，也需要检查压缩是否损失原意。
Few-shot，即在提示里给少量示例，可以把抽象要求变成可观察的样子。写“请采用合适风格”与给出几个合格样例，对执行者的约束不同。微调所需样本、算力等数字是当期经验，节目没有建立通用成本公式。编辑判断：选技术前先写出具体失败：答错事实、读不懂要求，还是风格不稳定；三种失败未必需要同一种修补。
3. Agent的难点在于每一步怎样可靠接起来
节目线索：15:47–23:04。 Agent在这里指利用模型、上下文和工具连续采取动作，以完成目标的系统。订机票不是生成一句答案：还要获得条件、查询结果，再按规则继续行动。访谈区分人工预先编排、完全自主和二者混合的形式，并提出把表现好的行动轨迹保存为可重复的流程。
这让探索与稳定执行有机会衔接，但一次成功的轨迹还要经受条件变化。嘉宾列举了上下文、推理质量和工具质量三类限制。一个具体案例是API自动测试：工具的接口描述太长，可能装不下或被裁剪，团队尝试更紧凑的描述方式。问题并非总在模型“大脑”，也可能在它获得的工具说明。
编辑判断：产品经理可以要求看到一条任务轨迹：每步拿到了什么资料，选择了什么工具，得到什么返回，在哪一步开始偏离。节目有关“完整项目必须一次装入上下文”或文字权重的口语解释，不宜当作普遍技术定律。更大的窗口和更短的提示，也都不能免去结果检查。
4. 提示工作台和模型中立，都要落到变更成本
节目线索：23:20–27:29、30:14–45:40。 路宇认为提示的潜力和难度都被低估。团队希望提供类似IDE的工作台——IDE是写软件时集中编辑和调试的环境——让开发者同时试多个提示变体、比较不同模型，并对照预期结果。模板的价值也需要按模型调过，而非只在菜单里显示“支持”。
应用还可能把文字生成、推理、向量检索、语音识别交给不同模型。Dify希望把模型与提示的选择从业务代码中拆开，使替换供应商不必牵动整个应用。这里的中立性有清楚的用户价值：降低更换能力的成本；是否形成长期竞争优势，则是另一项商业判断。
访谈把LangChain解释为代码工具箱与可实践的教材，把LangSmith解释为记录链条每步输入输出的调试工具，把Dify定位为包含运营工具的整体方案。路宇对LangChain深度使用时脆弱、需要重写的评价，属于竞争方在2023年的观察，不能直接成为现在的选型结论。
编辑判断：完整平台与自行组装的关键差异，是谁负责维护接口、资料、提示和调试环境。团队也计划将好用的资料处理等组件拆出，允许接入其他框架。采购时应拿一次真实变更做比较：换模型、更新一份资料或替换工具，需要谁改哪里，旧任务是否仍能完成？
5. 把岗位拆成任务，才知道今天能交付什么
节目线索：36:32–38:20、46:04–50:42、59:07–01:12:30。 私有知识并非上传一次便完成：内容更新、段落切分、权重、工具增删、调用轮次，都需要反复维护。路宇强调，不能让公司里两个懂AI的人长期包办所有业务判断。
在应用方向上，节目举到教育反馈、行政、营销和客服。作文给出改进方向，与获得可信的精确评分，是不同任务；客服理解订单和情绪，与被授权执行退款，也有不同责任。节目对这些方向持乐观态度，但没有提供独立的效果评测。
团队自己的做法是把增长工作拆成二三十项细活，再挑当期模型已较成熟的部分尝试自动化。曲凯在尾声把这个想法推进到需求定义：应用公司应认真利用已有能力，把问题限定到能交付的范围。所谓执行快，可能来自更清楚的需求和相关经验，而非省略基础工作。
未来入口少、后台有大量专业Agent，多模态、端侧小模型和AI可读资料的出版市场，都是节目中的方向性设想。把材料向量化不等于得到使用权；版权和改写许可仍是访谈承认未解决的问题。这些展望可以生成研究问题，不能替代当期交付证据。
用一个客服任务检验这期方法
下面是编辑设计的练习，不是Dify已公布的验收标准。选一项范围清楚的任务，例如根据订单和退货政策起草客服回复，先写出合格答案，再给系统一组正常、资料过期和信息缺失的案例。
| 要做的变更 | 具体观察 | 应保留的决定 |
|---|---|---|
| 更新退货政策 | 新回复是否引用新条件，旧版本能否追溯 | 谁确认资料生效 |
| 替换模型 | 同一组案例的正确性、等待和成本怎样变化 | 何时允许切换 |
| 接入退款工具 | 回复建议与实际动作是否分开记录 | 哪一步需要授权或人工接手 |
这组练习把“能跑”变成可讨论的结果，也让业务人员有具体入口参与维护。节目声称Dify能节省一到两个月开发时间，是团队当期观察；自己的项目仍需记录原型、调试与持续运营分别花了多少时间。
回听导航与关联阅读
- 00:34：LLMOps里的持续运营与参与者。
- 07:07：指令、资料与微调的选择。
- 15:48：Agent形式和可靠性问题。
- 23:20：提示工作台与预期答案。
- 38:55：工具库、调试工具与整体产品。
- 48:23：把岗位工作拆成具体应用。
- 01:08:49：曲凯对中间层机会的投资判断。
相关主题：Agent架构、最小可交付任务、开源产品指标。可对照Monica创始人肖弘谈AI助手，观察工程工具与用户产品怎样分别定义价值。
来源边界：依据完整提供的转写、官方单集介绍和既有精读重新组织，未逐句复听音频。转写大量问答误合并为主持人，本文按语境归纳并避免伪造逐句引语；姓名与产品名按官方介绍校正。所有规模、效率、竞品评价与预测均保留2023年语境，未独立验证，也不代表当前Dify能力。
English edition
Early Dify’s Product Judgment: Who Keeps an Application Useful After the Model Is Connected?
42 Zhangjing episode analysis · AI engineering and product operations
Guest: Zhang Luyu, introduced as founder of Dify.AI | Host: Qu Kai; closing questions: producer Celia
Published: August 20, 2023 | Duration: 74 minutes 9 seconds | Prepared: October 18, 2026
Original episode and official introduction
Connecting an API begins the product test
How far is a model that answers questions from an application a business can use over time? This 2023 interview breaks that distance into handling information, testing instructions, organizing tools and operating the product after launch. Dify’s ambition includes a workspace through which business colleagues can keep shaping an application.
Qu Kai discloses that he invested in Dify and has a working relationship with the team. Assessments of its execution, competitors and the opportunity for an intermediate platform should be read with that relationship in view. This analysis preserves the historical questions and methods; model rankings, business figures and predictions from 2023 are not presented as current facts.
1. Fifty thousand applications indicate exposure, not fifty thousand successes
Episode reference: 00:36–06:55. Luyu describes LLMOps as combining development with continuing operation of model applications. After engineers connect the capabilities, operations, sales and other colleagues who understand the business must still change knowledge and instructions to bring responses closer to expectations. Server availability and the usefulness of answers are separate problems that both need attention.
The interview mentions approximately 7,000 GitHub stars, more than 10,000 Docker image pulls and over 40,000 cloud applications, summarized as more than 50,000 applications. An image packages software for deployment. A pull is not established here as a distinct, continuously active production installation. Stars may also represent bookmarks: someone finds a project impressive, clicks and leaves.
Luyu suggests comparing image or package downloads and search trends. Editorial interpretation: These signals help identify developer interest, but further questions remain: has an application entered real work, how many users return, who handles failures and does it earn revenue? Different measurement layers cannot substitute for one another. Extending attention from building to maintaining is central to Dify’s proposition at the time.
2. Identify whether instructions, information or default behavior are missing
Episode reference: 07:48–15:19. The discussion introduces three engineering routes. A prompt supplies instructions, such as correcting grammar or classifying sentiment. Retrieval augmentation finds relevant material before answering and gives it to the model alongside the question. Fine-tuning uses training examples to adjust default behavior, such as producing poetry closer to a desired style.
The example asks which projects Qu Kai worked on that year. The model lacks private or recent information, so the system retrieves project material and constructs a fuller question. Terminology matters: embedding converts content into vector representations suitable for comparison. Retrieval, selection and context assembly form the larger process; embedding alone does not mean inserting material into a prompt.
An 8K context budget, discussed as a contemporary example, must accommodate the user’s question, retrieved material, conversation history and expected output. The enduring issue is which information deserves space and which essential conditions might be omitted. Having another model summarize first also requires checking whether compression loses meaning.
Few-shot prompting supplies a small set of examples. Asking for an appropriate style and showing acceptable outputs constrain execution differently. The interview’s sample and computing requirements for fine-tuning are historical experience, not a universal cost formula. Editorial interpretation: Write down the failure before choosing the repair: incorrect facts, misunderstood instructions and unstable style may call for different interventions.
3. An agent depends on reliable connections between actions
Episode reference: 15:47–23:04. An agent here uses a model, context and tools to take successive actions toward a goal. Booking a flight requires obtaining conditions, checking results and continuing under rules. The conversation distinguishes predefined orchestration, full autonomy and a mixture, proposing that successful action sequences can be saved as repeatable workflows.
This connects exploration with stable execution, although a sequence that succeeds once must still survive changing conditions. Luyu identifies context, reasoning quality and tool quality as constraints. In an API testing example, interface descriptions were too long and could be cut off, so the team tried a more compact representation. A failure may lie in the tool description the model receives.
Editorial interpretation: Ask to see a task trace: what information entered each step, which tool was selected, what it returned and where the process diverged. Conversational claims that a whole software project must enter context at once, or explanations about text weighting, should not become universal technical laws. Larger windows and shorter prompts still require checking results.
4. Prompt workspaces and model neutrality must reduce the cost of change
Episode reference: 23:20–27:29 and 30:14–45:40. Luyu argues that prompting’s potential and difficulty are both underestimated. The proposed workspace resembles an IDE, an integrated environment for editing and debugging software: developers compare prompt variants and models against expected results. Templates should be adapted to models, rather than support consisting only of a name on a menu.
An application may use different models for generation, reasoning, vector retrieval and speech recognition. Dify wants to separate model and prompt choices from business code, reducing the disruption of changing a supplier. Neutrality has a concrete user benefit in lower switching costs; whether it becomes a durable business advantage is a separate judgment.
The discussion describes LangChain as a code toolbox and practical textbook, LangSmith as a tool for tracing each step’s inputs and outputs, and Dify as an integrated product with operational tools. Luyu’s criticism that deeper LangChain applications become fragile or require rewriting is a competitor’s observation from 2023, not a current selection recommendation.
Editorial interpretation: A central difference between an integrated platform and self-assembly is who maintains interfaces, information, prompts and debugging. The team also considers offering useful components separately for use with other frameworks. Compare products through an actual change: replace a model, update a document or exchange a tool. Who must modify what, and do existing tasks still work?
5. Decompose a job to discover what can be delivered now
Episode reference: 36:32–38:20, 46:04–50:42 and 59:07–01:12:30. Private knowledge needs continuing updates, segmentation and weighting; tools and call limits also need maintenance. Two AI specialists cannot permanently supply every business judgment in a company.
Examples include educational feedback, administration, marketing and customer service. Suggesting improvements to an essay differs from producing a trustworthy precise score. Understanding an order and a customer’s mood differs from being authorized to issue a refund. The guests are optimistic about these directions but provide no independent outcome evaluation.
The team describes dividing growth work into twenty or thirty smaller activities and selecting those suited to contemporary model capabilities. In the closing discussion, Qu Kai applies this to requirements: use existing capabilities carefully and define a deliverable problem. Speed can follow from clearer understanding and relevant experience rather than skipping foundational work.
A future with few interfaces and many specialist agents, multimodal models, small local models and AI-readable publishing remains a set of directional scenarios. Vectorizing material does not grant usage rights. Copyright and permissions for rewritten work remain explicitly unresolved. These ideas generate research questions; they do not supply delivery evidence.
Test the method with one customer-service task
This is an editorial exercise, not a published Dify acceptance protocol. Choose a bounded task, such as drafting a response from an order and a return policy. Write an acceptable answer first, then test ordinary cases, outdated information and missing information.
| Change | What to observe | Decision to preserve |
|---|---|---|
| Update the return policy | Does the response use the new conditions, and can the old version be traced? | Who confirms that information is effective |
| Replace the model | How do correctness, waiting time and cost change on the same cases? | When switching is allowed |
| Connect a refund tool | Are proposed replies and executed actions recorded separately? | Where authorization or human takeover is required |
The exercise turns a runnable demonstration into assessable results and gives business colleagues a concrete maintenance role. Claims that Dify saved one or two development months are the team’s contemporary observations. A project should still measure prototyping, debugging and continuing operation separately.
Listening guide and related reading
- 00:34: Continuing operation and the people involved in LLMOps.
- 07:07: Choosing instructions, information and fine-tuning.
- 15:48: Agent forms and reliability problems.
- 23:20: Prompt workspaces and expected answers.
- 38:55: Libraries, debugging tools and integrated products.
- 48:23: Turning parts of a job into applications.
- 01:08:49: Qu Kai’s investment view of intermediate platforms.
Related themes: agent architecture, minimum deliverable tasks and open-source product metrics. Compare Monica founder Xiao Hong on AI assistants to examine how engineering tools and end-user products define value differently.
Source boundary: Reorganized from the complete supplied transcript, official episode introduction and existing close reading; audio was not checked sentence by sentence. Many exchanges were incorrectly combined under the host in transcription, so this article summarizes by context and avoids invented direct quotations. Names follow official metadata. Scale, efficiency, competitive assessments and predictions remain attributed to the 2023 discussion, are not independently verified and do not establish Dify’s current capabilities.