From 5d7b33e80094463e6fef893e33480e425e58b81c Mon Sep 17 00:00:00 2001 From: huanghuoguoguo <1051233107@qq.com> Date: Sun, 19 Apr 2026 15:53:28 +0800 Subject: [PATCH] docs: add rerank model documentation - Add rerank model section to zh/en model configuration pages with supported providers table and Bailian extra_args guide - Update zh/en features pages with Cohere/Jina/Voyage providers and RAG rerank support - Add optional rerank configuration section to zh/en knowledge base pages --- en/insight/features.mdx | 5 +++- en/usage/knowledge/readme.mdx | 11 +++++++++ en/usage/models/readme.mdx | 39 +++++++++++++++++++++++++++++++ zh/insight/features.mdx | 5 +++- zh/usage/knowledge/readme.mdx | 11 +++++++++ zh/usage/models/readme.mdx | 44 +++++++++++++++++++++++++++++++++-- 6 files changed, 111 insertions(+), 4 deletions(-) diff --git a/en/insight/features.mdx b/en/insight/features.mdx index 0193a8172..5409fc62d 100644 --- a/en/insight/features.mdx +++ b/en/insight/features.mdx @@ -55,13 +55,16 @@ For platform-specific feature support, please check [this document](/en/insight/ | [ModelScope](https://modelscope.cn/docs/model-service/API-Inference/intro) | ✅ | LLM gateway (MaaS) | | [TBox (Ant TreasureBox)](https://tbox.cn/) | ✅ | LLMOps platform | | [Coze](https://coze.cn) | ✅ | LLMOps platform | +| [Cohere](https://cohere.com/) | ✅ | Rerank models | +| [Jina](https://jina.ai/) | ✅ | Rerank models | +| [Voyage AI](https://voyageai.com/) | ✅ | Rerank models | ## LLM Application Paradigm Implementation | Application Paradigm | Status | Remarks | | --- | --- | --- | | Agent | ✅ | | -| RAG | ✅ | Using Chroma as vector database | +| RAG | ✅ | Using Chroma as vector database, supports Rerank re-scoring | | MCP | ✅ | Support Stdio and HTTP protocol | ## System Extensibility diff --git a/en/usage/knowledge/readme.mdx b/en/usage/knowledge/readme.mdx index 3722e94fb..1f8a9e4b2 100644 --- a/en/usage/knowledge/readme.mdx +++ b/en/usage/knowledge/readme.mdx @@ -29,6 +29,17 @@ Go to your pipeline configuration, under the "AI" tab, choose `Local Agent` as t use_kb +### Configure Reranking (Optional) + +After selecting a knowledge base, you can configure a **Rerank Model** below to improve retrieval quality: + +- **Rerank Model**: Select a configured rerank model (if none available, add one in [Model Configuration](/en/usage/models/readme) first) +- **Rerank Top-K**: Number of most relevant results to keep after reranking, default is 5 + + +Rerank models are optional. When configured, candidate documents from knowledge base retrieval are re-scored by the rerank model, keeping only the top-K most relevant results before passing them to the LLM, significantly reducing noise from irrelevant content. + + LangBot's built-in knowledge base can be used only when the runner is set to `Local Agent`. For other runners, refer to their documentation. diff --git a/en/usage/models/readme.mdx b/en/usage/models/readme.mdx index 5d7557403..b0662c463 100644 --- a/en/usage/models/readme.mdx +++ b/en/usage/models/readme.mdx @@ -45,3 +45,42 @@ The system has integrated the official embedding model provided by seekdb, no pa - On the "Embedding Model" page, select "seekdb-built-in"; - Click "Save" to use it immediately; - Then select this model in your knowledge base to take effect. + +### Rerank Model + +Rerank models re-score RAG retrieval results to significantly improve the precision of knowledge base answers. When a knowledge base returns multiple candidate documents, the rerank model uses a cross-encoder to precisely score each document's relevance to the query, selecting only the most relevant content. + +Enter `Model Name` and select `Model Provider`, then submit. After that, select this model in your pipeline's "AI" configuration. + + +Rerank models are optional. Without one, knowledge base results are used in their original retrieval order. With a rerank model configured, results are re-scored and only the top-K most relevant results are kept. + + +**Supported providers and models:** + +| Provider | Recommended Models | Notes | +| --- | --- | --- | +| [SiliconFlow](https://siliconflow.cn/) | `BAAI/bge-reranker-v2-m3`, `Qwen/Qwen3-Reranker-8B` | Recommended for China | +| [Jina](https://jina.ai/) | `jina-reranker-v2-base-multilingual` | Multilingual | +| [Cohere](https://cohere.com/) | `rerank-v3.5` | Industry leading | +| [Voyage AI](https://voyageai.com/) | `rerank-2.5` | | +| [GiteeAI](https://ai.gitee.com/) | `bge-reranker-v2-m3`, `Qwen3-Reranker-8B` | | +| [Aliyun Bailian](https://bailian.console.aliyun.com/) | `qwen3-rerank`, `gte-rerank` | Requires extra args, see below | +| [OpenRouter](https://openrouter.ai/) | `cohere/rerank-v3.5`, `cohere/rerank-4-pro` | Universal gateway | +| [302.AI](https://share.302.ai/SuTG99) | `jina-reranker-m0`, `qwen3-rerank` | | +| OpenAI Compatible | Any service supporting `/rerank` | Universal | + +### Using Rerank Models with Aliyun Bailian + +Aliyun Bailian's rerank API endpoint path differs from its LLM endpoint. You need to specify `rerank_url` in the model's **extra arguments**: + +1. When adding a model, select the `Aliyun Bailian` provider +2. Enter the model name (e.g., `qwen3-rerank`) +3. Add an extra argument: + - Key: `rerank_url` + - Type: `string` + - Value: `https://dashscope.aliyuncs.com/compatible-api/v1/reranks` + + +Other providers (SiliconFlow, Jina, Cohere, etc.) do not require extra arguments — they use the standard `/rerank` endpoint. + diff --git a/zh/insight/features.mdx b/zh/insight/features.mdx index 1940fafef..8e06a861f 100644 --- a/zh/insight/features.mdx +++ b/zh/insight/features.mdx @@ -55,13 +55,16 @@ LangBot 是一个开源的大语言模型原生即时通信机器人开发平台 | [ModelScope](https://modelscope.cn/docs/model-service/API-Inference/intro) | ✅ | 大模型聚合平台 | | [TBox（蚂蚁百宝箱）](https://tbox.cn/) | ✅ | LLMOps 平台 | | [Coze](https://coze.cn) | ✅ | LLMOps 平台 | +| [Cohere](https://cohere.com/) | ✅ | Rerank 模型 | +| [Jina](https://jina.ai/) | ✅ | Rerank 模型 | +| [Voyage AI](https://voyageai.com/) | ✅ | Rerank 模型 | ## LLM 应用范式实现 | 应用范式 | 状态 | 备注 | | --- | --- | --- | | Agent | ✅ | | -| RAG | ✅ | 使用 Chroma 作为向量数据库 | +| RAG | ✅ | 使用 Chroma 作为向量数据库，支持 Rerank 重排序 | | MCP | ✅ | 支持 Stdio 和 HTTP 两种方式 | ## 系统扩展性 diff --git a/zh/usage/knowledge/readme.mdx b/zh/usage/knowledge/readme.mdx index e4b69d7da..ba9cd5467 100644 --- a/zh/usage/knowledge/readme.mdx +++ b/zh/usage/knowledge/readme.mdx @@ -29,6 +29,17 @@ LangBot 原生支持 RAG（Retrieval-Augmented Generation，即检索增强生 use_kb +### 配置重排序（可选） + +选择知识库后，您可以在下方配置**重排序模型**来提升检索质量： + +- **重排序模型**：选择已配置的 Rerank 模型（如未配置，请先到[模型配置](/zh/usage/models/readme)中添加） +- **重排序保留数量（Top-K）**：重排序后保留的最相关结果数量，默认为 5 + + +重排序模型是可选的。配置后，知识库检索返回的候选文档会经过重排序模型重新打分，只保留最相关的 Top-K 条，再交给大模型回答，显著减少无关信息干扰。 + + 仅当运行器为`内置 Agent`时，才可以使用 LangBot 原生知识库，若要在使用其他运行器时使用知识库，请参考所使用的运行器对应产品的文档。 diff --git a/zh/usage/models/readme.mdx b/zh/usage/models/readme.mdx index bd0622f2c..3f3943c2d 100644 --- a/zh/usage/models/readme.mdx +++ b/zh/usage/models/readme.mdx @@ -23,7 +23,8 @@ LangBot Models 是 LangBot 官方提供的模型服务。当您使用 LangBot Sp ├── API Keys: [sk-xxx] ├── LLM 模型: gpt-4o ├── LLM 模型: gpt-4o-mini -└── 向量模型: text-embedding-3-small +├── 向量模型: text-embedding-3-small +└── 重排序模型: rerank-v3.5 ``` @@ -57,6 +58,45 @@ LangBot Models 是 LangBot 官方提供的模型服务。当您使用 LangBot Sp ### 使用 seekdb 内置嵌入模型（零配置） 系统已集成 seekdb 提供的官方嵌入模型，无需填写任何参数。 -- 在“嵌入模型”页面选择「seekdb-内置」； +- 在”嵌入模型”页面选择「seekdb-内置」； - 点击「保存」即可立即使用； - 后续在知识库中选择该模型即可生效。 + +### 重排序模型 (Rerank Model) + +重排序模型用于对 RAG 检索结果进行重新评分排序，显著提升知识库回答的精准度。当知识库返回多个候选文档时，重排序模型会使用 cross-encoder 对每个文档与查询的相关性进行精确打分，筛选出最相关的内容。 + +填入 `模型名称`、选择 `模型供应商`，然后提交即可。之后请在流水线的”AI 能力”配置中选择此模型。 + + +重排序模型是可选的。不配置时，知识库检索结果将按向量相似度排序直接使用。配置后，检索结果会经过重排序模型重新打分，只保留最相关的 Top-K 条结果。 + + +**支持的供应商和模型：** + +| 供应商 | 推荐模型 | 备注 | +| --- | --- | --- | +| [SiliconFlow](https://siliconflow.cn/) | `BAAI/bge-reranker-v2-m3`, `Qwen/Qwen3-Reranker-8B` | 国内推荐 | +| [Jina](https://jina.ai/) | `jina-reranker-v2-base-multilingual` | 多语言支持 | +| [Cohere](https://cohere.com/) | `rerank-v3.5` | 全球领先 | +| [Voyage AI](https://voyageai.com/) | `rerank-2.5` | | +| [GiteeAI](https://ai.gitee.com/) | `bge-reranker-v2-m3`, `Qwen3-Reranker-8B` | | +| [阿里云百炼](https://bailian.console.aliyun.com/) | `qwen3-rerank`, `gte-rerank` | 需要额外参数，见下方 | +| [OpenRouter](https://openrouter.ai/) | `cohere/rerank-v3.5`, `cohere/rerank-4-pro` | 通用网关 | +| [302.AI](https://share.302.ai/SuTG99) | `jina-reranker-m0`, `qwen3-rerank` | | +| OpenAI 兼容 | 任意支持 `/rerank` 的服务 | 通用 | + +### 阿里云百炼使用重排序模型 + +阿里云百炼的重排序 API 端点路径与其 LLM 端点不同，需要在模型的**额外参数**中指定 `rerank_url`： + +1. 添加模型时，选择 `阿里云百炼` 供应商 +2. 填入模型名称（如 `qwen3-rerank`） +3. 在额外参数中添加一个参数： + - Key: `rerank_url` + - Type: `string` + - Value: `https://dashscope.aliyuncs.com/compatible-api/v1/reranks` + + +其他供应商（SiliconFlow、Jina、Cohere 等）无需填写额外参数，使用标准的 `/rerank` 端点即可。 +