diff --git a/web/static/i18n/en-US.json b/web/static/i18n/en-US.json
index 9d742e5f..5dde56eb 100644
--- a/web/static/i18n/en-US.json
+++ b/web/static/i18n/en-US.json
@@ -2178,11 +2178,27 @@
"subIndexFilter": "Sub-index filter (optional)",
"subIndexFilterPlaceholder": "e.g. prod, must match an indexing sub_indexes tag",
"subIndexFilterHint": "Empty = no filter. When set, only rows whose sub_indexes contain this tag (legacy rows with empty sub_indexes still match).",
+ "ragPipelineHeader": "RAG pipeline (MultiQuery + Rerank)",
+ "ragPipelineHint": "MultiQuery and rerank are always on: LLM query rewrite → vector prefetch & fusion → HTTP rerank → dedupe & budget truncate.",
+ "multiQueryMaxQueries": "MultiQuery rewrite variant limit",
+ "multiQueryMaxQueriesPlaceholder": "4",
+ "multiQueryMaxQueriesHint": "Max LLM-generated retrieval variants (including paraphrases of the original query). Recommended 3–4, max 8.",
+ "rerankProvider": "Rerank provider",
+ "rerankProviderAuto": "Auto (infer from Base URL)",
+ "rerankProviderCohere": "Cohere-compatible API",
+ "rerankProviderHint": "DashScope uses gte-rerank; other compatible endpoints use /v1/rerank. Leave empty to infer from Base URL below.",
+ "rerankModel": "Rerank model (optional)",
+ "rerankModelPlaceholder": "Empty: DashScope→gte-rerank, Cohere→rerank-multilingual-v3.0",
+ "rerankBaseUrl": "Rerank Base URL (optional)",
+ "rerankBaseUrlPlaceholder": "Leave empty to reuse embedding / OpenAI base_url",
+ "rerankApiKey": "Rerank API Key (optional)",
+ "rerankApiKeyPlaceholder": "Leave empty to reuse embedding / OpenAI api_key",
+ "rerankApiKeyHint": "On rerank failure, results fall back to fusion order; search still works.",
"postRetrieveHeader": "Post-retrieval (dedupe / budget)",
"postRetrieveDedupeAuto": "Results are always deduped by normalized text (whitespace-collapsed bodies). No setting required.",
"prefetchTopK": "Prefetch candidates (vector stage)",
- "prefetchTopKPlaceholder": "0",
- "prefetchTopKHint": "0 = same as Top-K; larger values fetch more vector hits before dedupe/truncate (max 200).",
+ "prefetchTopKPlaceholder": "20",
+ "prefetchTopKHint": "Vector candidates per MultiQuery variant; 0 uses built-in max(top_k×4, 20) (max 200).",
"maxContextChars": "Max returned characters (Unicode)",
"maxContextCharsPlaceholder": "0",
"maxContextCharsHint": "0 = unlimited; keeps whole chunks in rank order until the budget is exceeded.",
diff --git a/web/static/i18n/zh-CN.json b/web/static/i18n/zh-CN.json
index 08a52d71..96ce1f5d 100644
--- a/web/static/i18n/zh-CN.json
+++ b/web/static/i18n/zh-CN.json
@@ -2166,11 +2166,27 @@
"subIndexFilter": "子索引过滤（可选）",
"subIndexFilterPlaceholder": "如 prod，与索引 sub_indexes 一致",
"subIndexFilterHint": "留空不过滤；填写后仅检索向量行 sub_indexes 中含该标签的结果（未打标旧行仍保留）。",
+ "ragPipelineHeader": "RAG 管线（MultiQuery + Rerank）",
+ "ragPipelineHint": "MultiQuery 与精排始终启用：LLM 改写多路检索 → 向量预取与融合 → HTTP 精排 → 去重与预算截断。",
+ "multiQueryMaxQueries": "MultiQuery 改写变体上限",
+ "multiQueryMaxQueriesPlaceholder": "4",
+ "multiQueryMaxQueriesHint": "LLM 生成的检索变体数量上限（含原问语义覆盖）；建议 3~4，最大 8。",
+ "rerankProvider": "精排提供商",
+ "rerankProviderAuto": "自动（按 Base URL 推断）",
+ "rerankProviderCohere": "Cohere 兼容 API",
+ "rerankProviderHint": "DashScope 使用 gte-rerank；其他兼容端点走 /v1/rerank。留空时按下方 Base URL 自动推断。",
+ "rerankModel": "精排模型（可选）",
+ "rerankModelPlaceholder": "留空：DashScope→gte-rerank，Cohere→rerank-multilingual-v3.0",
+ "rerankBaseUrl": "精排 Base URL（可选）",
+ "rerankBaseUrlPlaceholder": "留空则复用嵌入 / OpenAI 的 base_url",
+ "rerankApiKey": "精排 API Key（可选）",
+ "rerankApiKeyPlaceholder": "留空则复用嵌入 / OpenAI 的 api_key",
+ "rerankApiKeyHint": "精排失败时自动降级为融合排序，检索仍可用。",
"postRetrieveHeader": "检索后处理（去重 / 预算）",
"postRetrieveDedupeAuto": "检索结果会自动按正文规范化去重（合并仅空白不同的重复片段），无需配置。",
"prefetchTopK": "预取候选数（向量阶段）",
- "prefetchTopKPlaceholder": "0",
- "prefetchTopKHint": "0 表示与 Top-K 相同；大于 Top-K 时先多取候选再经去重/截断回到 Top-K（上限 200）。",
+ "prefetchTopKPlaceholder": "20",
+ "prefetchTopKHint": "每条 MultiQuery 变体的向量候选数；0 表示内置 max(top_k×4, 20)（上限 200）。",
"maxContextChars": "返回内容最大字符数（Unicode）",
"maxContextCharsPlaceholder": "0",
"maxContextCharsHint": "0 表示不限制；按检索顺序整段保留 chunk，超出则丢弃后续。",
diff --git a/web/static/js/settings.js b/web/static/js/settings.js
index b5d875df..654cbed8 100644
--- a/web/static/js/settings.js
+++ b/web/static/js/settings.js
@@ -389,10 +389,35 @@ async function loadConfig(loadTools = true) {
subIdxFilterInput.value = knowledge.retrieval?.sub_index_filter || '';
}
+ const mq = knowledge.retrieval?.multi_query || {};
+ const mqMaxInput = document.getElementById('knowledge-multi-query-max-queries');
+ if (mqMaxInput) {
+ const mqVal = parseInt(mq.max_queries, 10);
+ mqMaxInput.value = (!isNaN(mqVal) && mqVal > 0) ? mqVal : 4;
+ }
+ const rr = knowledge.retrieval?.rerank || {};
+ const rerankProviderSelect = document.getElementById('knowledge-rerank-provider');
+ if (rerankProviderSelect) {
+ const p = (rr.provider || '').toLowerCase();
+ rerankProviderSelect.value = (p === 'dashscope' || p === 'cohere') ? p : '';
+ }
+ const rerankModelInput = document.getElementById('knowledge-rerank-model');
+ if (rerankModelInput) {
+ rerankModelInput.value = rr.model || '';
+ }
+ const rerankBaseUrlInput = document.getElementById('knowledge-rerank-base-url');
+ if (rerankBaseUrlInput) {
+ rerankBaseUrlInput.value = rr.base_url || '';
+ }
+ const rerankApiKeyInput = document.getElementById('knowledge-rerank-api-key');
+ if (rerankApiKeyInput) {
+ rerankApiKeyInput.value = rr.api_key || '';
+ }
+
const post = knowledge.retrieval?.post_retrieve || {};
const prefetchInput = document.getElementById('knowledge-post-retrieve-prefetch-top-k');
if (prefetchInput) {
- prefetchInput.value = post.prefetch_top_k ?? 0;
+ prefetchInput.value = post.prefetch_top_k ?? 20;
}
const maxCharsInput = document.getElementById('knowledge-post-retrieve-max-chars');
if (maxCharsInput) {
@@ -1273,8 +1298,25 @@ async function applySettings() {
return isNaN(val) ? 0.7 : val;
})(),
sub_index_filter: document.getElementById('knowledge-retrieval-sub-index-filter')?.value?.trim() || '',
+ multi_query: {
+ max_queries: (() => {
+ const v = parseInt(document.getElementById('knowledge-multi-query-max-queries')?.value, 10);
+ if (isNaN(v) || v <= 0) return 4;
+ return Math.min(8, v);
+ })()
+ },
+ rerank: {
+ provider: document.getElementById('knowledge-rerank-provider')?.value?.trim() || '',
+ model: document.getElementById('knowledge-rerank-model')?.value?.trim() || '',
+ base_url: document.getElementById('knowledge-rerank-base-url')?.value?.trim() || '',
+ api_key: document.getElementById('knowledge-rerank-api-key')?.value?.trim() || ''
+ },
post_retrieve: {
- prefetch_top_k: parseInt(document.getElementById('knowledge-post-retrieve-prefetch-top-k')?.value, 10) || 0,
+ prefetch_top_k: (() => {
+ const raw = document.getElementById('knowledge-post-retrieve-prefetch-top-k')?.value;
+ const v = parseInt(raw, 10);
+ return isNaN(v) ? 20 : Math.max(0, v);
+ })(),
max_context_chars: parseInt(document.getElementById('knowledge-post-retrieve-max-chars')?.value, 10) || 0,
max_context_tokens: parseInt(document.getElementById('knowledge-post-retrieve-max-tokens')?.value, 10) || 0
}
diff --git a/web/templates/index.html b/web/templates/index.html
index 6820bdd6..9ee25e43 100644
--- a/web/templates/index.html
+++ b/web/templates/index.html
@@ -3009,14 +3009,46 @@
留空表示不过滤；非空时仅检索 sub_indexes 含该标签的向量行（未打标旧数据仍会命中）。
+
+ MultiQuery 与精排始终启用：LLM 改写多路检索 → 向量预取与融合 → HTTP 精排 → 去重与预算截断。
+
+
+
+ LLM 生成的检索变体数量上限（含原问语义覆盖）；建议 3~4，最大 8。
+
+
+
+
+ DashScope 使用 gte-rerank；其他兼容端点走 /v1/rerank。留空时按下方 Base URL 自动推断。
+
+
+
+
+
+
+
+
+
+
+
+
+ 精排失败时自动降级为融合排序，检索仍可用。
+
+
检索结果会自动按正文规范化去重（合并仅空白不同的重复片段），无需配置。
-
- 0 表示与 Top-K 相同；大于 Top-K 时先多取候选再经去重/截断回到 Top-K（上限 200）。
+
+ 每条 MultiQuery 变体的向量候选数；0 表示内置 max(top_k×4, 20)（上限 200）。