fix: context-aware web search query when PDF files are uploaded

app/api/web-search/route.ts

@@ -3,10 +3,13 @@
  *
  * POST /api/web-search
  * Simple JSON request/response using Tavily search.
+ * When pdfText is provided, enhances the query using LLM to extract
+ * meaningful keywords from the document.
  */
 
 import { searchWithTavily, formatSearchResultsAsContext } from '@/lib/web-search/tavily';
 import { resolveWebSearchApiKey } from '@/lib/server/provider-config';
+import { enhanceSearchQuery } from '@/lib/web-search/query-enhancer';
 import { createLogger } from '@/lib/logger';
 import { apiError, apiSuccess } from '@/lib/server/api-response';
 
@@ -15,9 +18,22 @@ const log = createLogger('WebSearch');
 export async function POST(req: Request) {
   try {
     const body = await req.json();
-    const { query, apiKey: clientApiKey } = body as {
+    const {
+      query,
+      apiKey: clientApiKey,
+      pdfText,
+      modelConfig,
+    } = body as {
       query?: string;
       apiKey?: string;
+      pdfText?: string;
+      modelConfig?: {
+        modelString?: string;
+        apiKey?: string;
+        baseUrl?: string;
+        providerType?: string;
+        requiresApiKey?: boolean;
+      };
     };
 
     if (!query || !query.trim()) {
@@ -33,7 +49,10 @@ export async function POST(req: Request) {
       );
     }
 
-    const result = await searchWithTavily({ query: query.trim(), apiKey });
+    // Enhance query with PDF context if available
+    const effectiveQuery = await enhanceSearchQuery(query.trim(), pdfText, modelConfig);
+
+    const result = await searchWithTavily({ query: effectiveQuery, apiKey });
     const context = formatSearchResultsAsContext(result);
 
     return apiSuccess({

app/generation-preview/page.tsx

@@ -307,12 +307,23 @@ function GenerationPreviewContent() {
         const wsSettings = useSettingsStore.getState();
         const wsApiKey =
           wsSettings.webSearchProvidersConfig?.[wsSettings.webSearchProviderId]?.apiKey;
+        const wsModelConfig = getCurrentModelConfig();
         const res = await fetch('/api/web-search', {
           method: 'POST',
           headers: { 'Content-Type': 'application/json' },
           body: JSON.stringify({
             query: currentSession.requirements.requirement,
             apiKey: wsApiKey || undefined,
+            pdfText: currentSession.pdfText || undefined,
+            modelConfig: currentSession.pdfText
+              ? {
+                  modelString: wsModelConfig.modelString,
+                  apiKey: wsModelConfig.apiKey,
+                  baseUrl: wsModelConfig.baseUrl,
+                  providerType: wsModelConfig.providerType,
+                  requiresApiKey: wsModelConfig.requiresApiKey,
+                }
+              : undefined,
           }),
           signal,
         });

lib/server/classroom-generation.ts

@@ -20,6 +20,7 @@ import { parseModelString } from '@/lib/ai/providers';
 import { resolveApiKey, resolveWebSearchApiKey } from '@/lib/server/provider-config';
 import { resolveModel } from '@/lib/server/resolve-model';
 import { searchWithTavily, formatSearchResultsAsContext } from '@/lib/web-search/tavily';
+import { enhanceSearchQuery } from '@/lib/web-search/query-enhancer';
 import { persistClassroom } from '@/lib/server/classroom-storage';
 import {
   generateMediaForClassroom,
@@ -241,7 +242,8 @@ export async function generateClassroom(
     if (tavilyKey) {
       try {
         log.info('Running web search for requirement context...');
-        const searchResult = await searchWithTavily({ query: requirement, apiKey: tavilyKey });
+        const searchQuery = await enhanceSearchQuery(requirement, pdfText);
+        const searchResult = await searchWithTavily({ query: searchQuery, apiKey: tavilyKey });
         researchContext = formatSearchResultsAsContext(searchResult);
         if (researchContext) {
           log.info(`Web search returned ${searchResult.sources.length} sources`);

lib/web-search/query-enhancer.ts

@@ -0,0 +1,86 @@
+/**
+ * Search Query Enhancer
+ *
+ * When a PDF is uploaded but the user requirement is too vague
+ * (e.g. "讲讲这篇论文"), uses LLM to extract meaningful keywords
+ * from the document text for a better web search query.
+ */
+
+import { callLLM } from '@/lib/ai/llm';
+import { resolveModel, type ResolvedModel } from '@/lib/server/resolve-model';
+import { createLogger } from '@/lib/logger';
+
+const log = createLogger('QueryEnhancer');
+
+/**
+ * Enhance a search query using PDF text context.
+ *
+ * If pdfText is provided, uses a lightweight LLM call to extract
+ * the document's title, key topics, and important terms, then
+ * constructs a concise search query.
+ *
+ * Returns the original query if no PDF is available or if
+ * enhancement fails.
+ *
+ * @param modelConfig - Optional model config for the LLM call.
+ *   If not provided, uses server-configured default model.
+ */
+export async function enhanceSearchQuery(
+  query: string,
+  pdfText: string | undefined,
+  modelConfig?: {
+    modelString?: string;
+    apiKey?: string;
+    baseUrl?: string;
+    providerType?: string;
+    requiresApiKey?: boolean;
+  },
+): Promise<string> {
+  if (!pdfText) return query;
+
+  try {
+    let resolved: ResolvedModel;
+    try {
+      resolved = resolveModel(modelConfig || {});
+    } catch {
+      log.warn('No model available for query enhancement, using original query');
+      return query;
+    }
+
+    const snippet = pdfText.slice(0, 2000);
+
+    const result = await callLLM(
+      {
+        model: resolved.model,
+        messages: [
+          {
+            role: 'system',
+            content:
+              'You extract search keywords from documents. Return ONLY a concise search query string (under 200 characters), no explanation.',
+          },
+          {
+            role: 'user',
+            content: `The user uploaded a document and said: "${query}"
+
+Extract the document's title, key topics, and important terms from the text below, then combine them into a concise web search query (under 200 characters) that would find relevant supplementary materials.
+
+Document text (first 2000 chars):
+${snippet}`,
+          },
+        ],
+        maxOutputTokens: resolved.modelInfo?.outputWindow,
+      },
+      'enhance-search-query',
+    );
+
+    const enhanced = result.text.trim();
+    if (enhanced && enhanced.length > 5 && enhanced.length <= 400) {
+      log.info(`Enhanced search query: "${query}" → "${enhanced}"`);
+      return enhanced;
+    }
+  } catch (e) {
+    log.warn('Search query enhancement failed, using original query:', e);
+  }
+
+  return query;
+}