LOOM 的几次演进：画布能连起来，执行链为什么还会断？

LOOM 和 AgentFlow Studio 都尝试过可视化工作流、知识库与流式对话。LOOM-PLUS 的记录明确说明，它基于 AgentFlow 骨架整合 LOOM 的部分能力。这些仓库属于一条演进路径，不能当成三套完全独立的产品成果相加。

这篇聚焦 2026 年 5 月的整合。我确定要保留的功能与检查目标，AI 编程工具承担代码实现。画布能拖动、节点能连线以后，我们还需要追到文件解析、向量入库和实际执行。

PDF 上传入口背后，还有解析器的接口

一处留下修复提交的问题来自 pdf-parse：旧调用方式与 2.x API 不一致。修复修改了两个解析位置，而不是只改上传页上的错误提示。

这类失败容易被演示环境掩盖。前端可以接收文件，接口也可能成功收到字节，后续解析仍然失败。检查需要进入实际使用的依赖版本和解析调用，不能从“文件选中了”推到“文档已经可检索”。

我没有沿用旧复盘里对所有版本范围的笼统判断。这次可以确认的是一个具体 API 适配提交；依赖管理还需要结合锁文件、干净安装和目标环境验证。

换 embedding 模型，也会改变数据契约

另一处记录涉及向量维度：原先按一种模型建表，切换模型后输出长度不同，数据库无法照旧接收。把模型名称换掉，并没有同时完成表结构与既有向量的处理。

整合记录因此把维度配置和数据库变更纳入考虑。它提醒我，模型替换会经过调用层、存储与检索，不能只检查密钥和接口地址。

这里仍有范围限制：源码里的迁移、配置和修复记录，不等于旧知识库在所有版本之间都已安全迁移。本文没有重新执行整套数据升级。

把复用写清楚，也把验证留在对应层

画布验证可以拦循环和缺配置，解析检查处理文件，检索还需要核对命中文本与引用。任何一层通过都不能包办其他层。

这轮回看让我更愿意沿着一份文档检查完整路径：上传、解析、分块、向量化、检索、回答，再观察失败时是否留下可定位信息。当前只核对历史提交与实现，没有重跑产品，也没有业务采用数据。

依据：PDF API 适配提交、固定版本的演进说明，以及 LOOM、AgentFlow 的对应源码。后续应先选定一个维护入口，再用同一批文档比较修改前后的表现。

LOOM iterations: connecting a canvas is only the start of execution

LOOM and AgentFlow Studio both explored visual workflows, knowledge bases, and streaming chat. LOOM-PLUS explicitly combines the AgentFlow skeleton with parts of LOOM. I treat these repositories as an evolution rather than three independent product outcomes.

This account focuses on the May 2026 integration. I set functional and verification requirements; AI coding tools implemented the changes. A connected canvas still needed a working path through parsing, embeddings, and execution.

Follow the file beyond upload

A recorded fix adapted two parsing locations to the pdf-parse 2.x API. Receiving bytes at an upload endpoint did not establish that the parser could process them.

The useful investigation followed the installed dependency and actual call sites. I do not repeat the older README’s broad conclusions about all version ranges: the supported claim is this specific API-adaptation change. Lockfiles, clean installation, and the target environment still matter.

A model change can alter stored data

Another recorded issue concerned embedding dimensions. A database created for one output length could not accept another model’s vectors unchanged. Replacing a model name did not migrate the schema or existing knowledge base.

The integration work therefore considered dimension configuration and database changes. This crosses the provider, persistence, and retrieval layers. Configuration and migration code do not establish that every historical knowledge base has been safely upgraded; this article reruns no such migration.

Check each layer with its own evidence

Canvas validation can catch cycles and missing configuration. Parsing checks documents; retrieval checks passages and sources. Passing one layer cannot cover the others.

The next useful exercise is to follow one document through upload, parsing, chunking, embedding, retrieval, and an answer, while retaining failures. I reviewed history and implementation without rerunning the product or establishing business adoption.

Sources include the PDF API fix, the fixed-version integration account, and corresponding LOOM and AgentFlow code. Choosing one maintenance entry would make that comparison easier to repeat.