open-webui/Fu-Jie_openwebui-extensions
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: Introduce Wisdom Synthesizer pipeline and documentation, replacing the MoE Prompt Refiner.

Browse Source
This commit is contained in:
fujie
2026-03-23 00:33:20 +08:00
parent 2fae50a39f
commit e6f22b0f82
11 changed files with 375 additions and 467 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
docs/plugins/pipelines/index.md
View File

@@ -17,15 +17,12 @@ Pipelines extend beyond simple transformations to implement:
<div class="grid cards" markdown>
- :material-view-module:{ .lg .middle } **MoE Prompt Refiner**
- :material-view-module:{ .lg .middle } **Wisdom Synthesizer**
[:octicons-tag-24: v0.1.0](https://github.com/Fu-Jie/open-webui-pipeline-wisdom-synthesizer){ .bubble }
---
An external pipeline filter that refactors aggregate requests with collective wisdom to output structured expert reports.
Refines prompts for Mixture of Experts (MoE) summary requests to generate high-quality comprehensive reports.
**Version:** 1.0.0
[:octicons-arrow-right-24: Documentation](moe-prompt-refiner.md)
[:octicons-arrow-right-24: Documentation](wisdom-synthesizer.md)
</div>

11
docs/plugins/pipelines/index.zh.md
View File

@@ -17,15 +17,12 @@ Pipelines 不仅是简单转换，还可以实现：
<div class="grid cards" markdown>
- :material-view-module:{ .lg .middle } **MoE Prompt Refiner**
- :material-view-module:{ .lg .middle } **Wisdom Synthesizer**
[:octicons-tag-24: v0.1.0](https://github.com/Fu-Jie/open-webui-pipeline-wisdom-synthesizer){ .bubble }
---
智能拦截并重构多模型汇总请求，发挥集体智慧（Collective Wisdom），将常规汇总熔炼为专家级对比报告。
为 Mixture of Experts（MoE）汇总请求优化提示词，生成高质量综合报告。
**版本：** 1.0.0
[:octicons-arrow-right-24: 查看文档](moe-prompt-refiner.md)
[:octicons-arrow-right-24: 查看文档](wisdom-synthesizer.md)
</div>

109
docs/plugins/pipelines/moe-prompt-refiner.md
View File

@@ -1,109 +0,0 @@
# MoE Prompt Refiner
<span class="category-badge pipeline">Pipeline</span>
<span class="version-badge">v1.0.0</span>
Refines prompts for Mixture of Experts (MoE) summary requests to generate high-quality comprehensive reports.
---
## Overview
The MoE Prompt Refiner is an advanced pipeline that optimizes prompts before sending them to multiple expert models, then synthesizes the responses into comprehensive, high-quality reports.
## Features
- :material-view-module: **Multi-Model**: Leverages multiple AI models
- :material-text-search: **Prompt Optimization**: Refines prompts for best results
- :material-merge: **Response Synthesis**: Combines expert responses
- :material-file-document: **Report Generation**: Creates structured reports
---
## Installation
1. Download the pipeline file: [`moe_prompt_refiner.py`](https://github.com/Fu-Jie/openwebui-extensions/tree/main/plugins/pipelines)
2. Upload to OpenWebUI: **Admin Panel** → **Settings** → **Functions**
3. Configure expert models and settings
4. Enable the pipeline
---
## How It Works
```mermaid
graph TD
A[User Prompt] --> B[Prompt Refiner]
B --> C[Expert Model 1]
B --> D[Expert Model 2]
B --> E[Expert Model N]
C --> F[Response Synthesizer]
D --> F
E --> F
F --> G[Comprehensive Report]
```
---
## Configuration
| Option | Type | Default | Description |
|--------|------|---------|-------------|
| `expert_models` | list | `[]` | List of models to consult |
| `synthesis_model` | string | `"auto"` | Model for synthesizing responses |
| `report_format` | string | `"markdown"` | Output format |
---
## Use Cases
- **Research Reports**: Gather insights from multiple AI perspectives
- **Comprehensive Analysis**: Multi-faceted problem analysis
- **Decision Support**: Balanced recommendations from diverse models
- **Content Creation**: Rich, multi-perspective content
---
## Example
**Input Prompt:**
```
Analyze the pros and cons of microservices architecture
```
**Output Report:**
```markdown
# Microservices Architecture Analysis
## Executive Summary
Based on analysis from multiple expert perspectives...
## Advantages
1. **Scalability** (Expert A)...
2. **Technology Flexibility** (Expert B)...
## Disadvantages
1. **Complexity** (Expert A)...
2. **Distributed System Challenges** (Expert C)...
## Recommendations
Synthesized recommendations based on expert consensus...
```
---
## Requirements
!!! note "Prerequisites"
- OpenWebUI v0.3.0 or later
- Access to multiple LLM models
- Sufficient API quotas for multi-model queries
!!! warning "Resource Usage"
This pipeline makes multiple API calls per request. Monitor your usage and costs.
---
## Source Code
[:fontawesome-brands-github: View on GitHub](https://github.com/Fu-Jie/openwebui-extensions/tree/main/plugins/pipelines){ .md-button }

109
docs/plugins/pipelines/moe-prompt-refiner.zh.md
View File

@@ -1,109 +0,0 @@
# MoE Prompt Refiner
<span class="category-badge pipeline">Pipeline</span>
<span class="version-badge">v1.0.0</span>
为 Mixture of Experts（MoE）汇总请求优化提示词，生成高质量的综合报告。
---
## 概览
MoE Prompt Refiner 是一个高级 Pipeline，会在将请求发送给多个专家模型前先优化提示词，然后综合各模型回复，输出结构化的高质量报告。
## 功能特性
- :material-view-module: **多模型**：同时利用多个 AI 模型
- :material-text-search: **提示词优化**：在发送前优化 prompt 获得更好结果
- :material-merge: **结果合成**：整合专家回复
- :material-file-document: **报告生成**：输出结构化报告
---
## 安装
1. 下载 Pipeline 文件：[`moe_prompt_refiner.py`](https://github.com/Fu-Jie/openwebui-extensions/tree/main/plugins/pipelines)
2. 上传到 OpenWebUI：**Admin Panel** → **Settings** → **Functions**
3. 配置专家模型及相关参数
4. 启用该 Pipeline
---
## 工作流程
```mermaid
graph TD
A[User Prompt] --> B[Prompt Refiner]
B --> C[Expert Model 1]
B --> D[Expert Model 2]
B --> E[Expert Model N]
C --> F[Response Synthesizer]
D --> F
E --> F
F --> G[Comprehensive Report]
```
---
## 配置项
| 选项 | 类型 | 默认值 | 说明 |
|--------|------|---------|-------------|
| `expert_models` | list | `[]` | 需要咨询的模型列表 |
| `synthesis_model` | string | `"auto"` | 用于综合回复的模型 |
| `report_format` | string | `"markdown"` | 输出格式 |
---
## 适用场景
- **研究报告**：从多个 AI 视角收集洞见
- **综合分析**：多角度问题拆解
- **决策支持**：获得多模型的平衡建议
- **内容创作**：生成多视角的丰富内容
---
## 示例
**输入 Prompt：**
```
Analyze the pros and cons of microservices architecture
```
**输出报告：**
```markdown
# Microservices Architecture Analysis
## Executive Summary
Based on analysis from multiple expert perspectives...
## Advantages
1. **Scalability** (Expert A)...
2. **Technology Flexibility** (Expert B)...
## Disadvantages
1. **Complexity** (Expert A)...
2. **Distributed System Challenges** (Expert C)...
## Recommendations
Synthesized recommendations based on expert consensus...
```
---
## 运行要求
!!! note "前置条件"
- OpenWebUI v0.3.0 及以上
- 可以访问多个 LLM 模型
- 有足够的 API 配额支撑多模型请求
!!! warning "资源消耗"
此 Pipeline 每次请求会进行多次 API 调用，请关注用量与成本。
---
## 源码
[:fontawesome-brands-github: 在 GitHub 查看](https://github.com/Fu-Jie/openwebui-extensions/tree/main/plugins/pipelines){ .md-button }

73
docs/plugins/pipelines/wisdom-synthesizer.md Normal file
View File

@@ -0,0 +1,73 @@
# Wisdom Synthesizer (Collective Wisdom Synthesizer)
An external pipeline filter (Pipeline/Filter) for **Open WebUI** that intercepts multi-model aggregate requests to leverage collective wisdom, reshaping **basic and linear aggregate outputs** into structured, high-contrast **expert analysis reports**.
![Effect Demonstration](wisdom_synthesizer.gif)
---
## 🚀 Key Features
* **Smart Interception**: Automatically catches Open WebUI's “Summarize various models' responses” requests.
* **Dynamic Parsing**: Strips away generic formatting and precisely extracts the **original user query** and **each model's individual response**.
* **Wisdom Fusion**: Directs the summary model to act as a “Chief Analyst”, enforcing a critical evaluation workflow instead of generic merging.
* **Standardized Output Structure**: Guarantees output layout includes:
* **【Core Consensus】**: Aggregated common ground across models.
* **【Key Divergences】**: Comparative breakdown of different perspectives/approaches.
* **【Unique Insights】**: Spotlighting innovative points found in a single model.
* **【Synthesis & Recommendation】**: An action-oriented, blended strategy set.
---
## 📦 Installation & Usage (Pipelines Mode)
> [!IMPORTANT]
> **Prerequisite**:
> This plugin relies on the official **[open-webui/pipelines](https://github.com/open-webui/pipelines)** framework. Please ensure your Open WebUI backend is already connected to an active `pipelines` runner environment beforehand.
This plugin runs as a single-file pipeline filter component and supports importing with just a single click:
### 🚀 One-Click Import via URL (Recommended 🌟)
1. Log into your Open WebUI board, go to **Admin settings** -> **Pipelines** tab.
2. Click **“Add Pipeline”** and paste the **GitHub Raw link** of `wisdom_synthesizer.py` into the address bar.
3. Save configurations to load automatically.
Below is the visual operational guide for getting it loaded:
![Installation Guide](install.gif)
---
## ⚙️ Valves Configuration
Configuration items inside safe Valves toggles:
| Parameter | Default | Description |
| :--- | :--- | :--- |
| `pipelines` | `["*"]` | Target model IDs to apply this Filter to *（Keep default for global）* |
| `priority` | `0` | Filter pipeline execution order priority (lower numbers execute first). |
| `model_id` | `None` | (Optional) Force the summarize job to run on a dedicated high-spec summary model. |
| `trigger_prefix` | `You have been provided...` | Pre-set phrase to trigger interception. Usually requires no changes. |
| `query_start_marker` | `'the latest user query: "'` | Anchor used to locate the start of the original query. |
| `query_end_marker` | `'"\n\nYour task is to'` | Anchor used to locate the end of the original query. |
| `response_start_marker` | `"Responses from models: "` | Anchor used to locate where the model responses begin. |
> [!TIP]
> **Configuration Tip**:
> The default `["*"]` allows the filter to securely adapt to any aggregator models chosen on the fly. In most scenarios, **keeping this default configuration** is highly recommended.
---
## 🤝 Related Projects
If you're building inside the Open WebUI ecosystem, you might find my other plugins sets helpful:
* 🚀 **[openwebui-extensions](https://github.com/Fu-Jie/openwebui-extensions)** —— A comprehensive collection of Actions, Pipes, and Tools to supercharge your workspace.
* 🪄 **[open-webui-prompt-plus](https://github.com/Fu-Jie/open-webui-prompt-plus)** —— Enhances Prompt engineering with AI-powered generators, Spotlight-style searches, and interactive forms.
---
## 📄 License
[MIT License](LICENSE)

73
docs/plugins/pipelines/wisdom-synthesizer.zh.md Normal file
View File

@@ -0,0 +1,73 @@
# Wisdom Synthesizer (集体智慧合成器)
专为 **Open WebUI** 设计的外置管道过滤器（Pipeline/Filter），旨在通过智能拦截并重构多模型汇总请求，发挥集体智慧（Collective Wisdom），将原本较为**基础和扁平的常规汇总**熔炼为结构清晰、具备多维对比度的**专家级综合分析报告**。
![效果演示](wisdom_synthesizer.gif)
---
## 🚀 核心功能
* **智能拦截**：自动捕获 Open WebUI 的“汇总多模型响应”请求（通过内置前缀触发）。
* **动态解析**：剥离多余格式，精准提取**原始用户问题**与**各模型的独立回答**。
* **智慧融合**：摒弃基础的模型合并，强制总结模型扮演“首席分析师”，发挥集体智慧审视全局。
* **规范输出**：将汇总响应熔炼为以下结构：
* **【核心共识】**: 提炼模型间的相同点。
* **【关键分歧】**: 对比不同视角的碰撞。
* **【独特洞察】**: 发现单一模型闪光点。
* **【综合建议】**: 最终形成有弹性的熔铸方案。
---
## 📦 安装与使用 (Pipelines 模式)
> [!IMPORTANT]
> **前提条件**：
> 本插件依赖于 Open WebUI 官方的 **[open-webui/pipelines](https://github.com/open-webui/pipelines)** 框架插件系统。请确保你的 Open WebUI 后端已经架设好或已连接底层的 `pipelines` 服务端环境。
本插件为单文件管道过滤组件，支持在面板中一键拉取安装：
### 🚀 通过 URL 一键导入 (推荐 🌟)
1. 登录你的 Open WebUI 后台，进入 **管理员设置** -> **Pipelines** 选项卡。
2. 点击 **“添加 Pipeline”**，并在地址栏中复制贴入此仓库中 `wisdom_synthesizer.py` 的 **GitHub Raw 链接**。
3. 点击 **保存** 即可成功加载。
以下是操作动态演示：
![安装操作图](install.gif)
---
## ⚙️ Valves 管道配置
进入管道配置项，可动态调整以下参数：
| 参数 | 默认值 | 说明 |
| :--- | :--- | :--- |
| `pipelines` | `["*"]` | 应用此 Filter 的目标模型 ID *（如果要全局生效保持默认）* |
| `priority` | `0` | 过滤器管道执行优先级（数字越小，越优先执行） |
| `model_id` | `None` | (可选) 强制将汇总任务流向你指定的某个专用高性能总结模型 |
| `trigger_prefix` | `You have been provided...` | 用于触发拦截的提示词起始句柄前缀。一般无需修改 |
| `query_start_marker` | `'the latest user query: "'` | 解析原始查询的起始标记锚点 |
| `query_end_marker` | `'"\n\nYour task is to'` | 解析原始查询的结束标记锚点 |
| `response_start_marker` | `"Responses from models: "` | 解析各个模型独立响应的起始锚点标志 |
> [!TIP]
> **配置建议**：
> 默认值 `["*"]` 可在所有选定的汇总模型上自适应生效。在绝大多数情况下，你**仅需保持此默认参数**便可保障全局自适应拦截。
---
## 🤝 友情链接 (Related Projects)
如果你对 Open WebUI 的扩展生态感兴趣，欢迎关注我的其它开源方案：
* 🚀 **[openwebui-extensions](https://github.com/Fu-Jie/openwebui-extensions)** —— 包含各种增强 Actions、Pipes、Tools 等一篮子开源插件合集，助你解锁更多黑魔法。
* 🪄 **[open-webui-prompt-plus](https://github.com/Fu-Jie/open-webui-prompt-plus)** —— 包含 AI 驱动的提示词生成器、Spotlight 搜索框及交互变量表单，极速拉满提示词工程。
---
## 📄 开源许可
[MIT License](LICENSE)