@librechat/agents 2.4.31 → 2.4.34

package/src/tools/search/test.ts

@@ -0,0 +1,159 @@
+/* eslint-disable no-console */
+// processWikipedia.ts
+import * as fs from 'fs';
+import * as path from 'path';
+import { processContent } from './content';
+
+/**
+ * Process a Wikipedia article (HTML and Markdown) and create a referenced version
+ */
+async function processWikipediaArticle(): Promise<void> {
+  try {
+    console.log('Starting Wikipedia article processing...');
+
+    // Define file paths - adapt these to your specific file locations
+    const htmlPath = path.resolve('./test.html');
+    const markdownPath = path.resolve('./test.md');
+    const outputPath = path.resolve('./output.md');
+
+    // Check if input files exist
+    if (!fs.existsSync(htmlPath)) {
+      throw new Error(`Wikipedia HTML file not found at ${htmlPath}`);
+    }
+
+    if (!fs.existsSync(markdownPath)) {
+      throw new Error(`Wikipedia Markdown file not found at ${markdownPath}`);
+    }
+
+    console.log('Reading Wikipedia article files...');
+    const html = fs.readFileSync(htmlPath, 'utf-8');
+    const markdown = fs.readFileSync(markdownPath, 'utf-8');
+
+    // Extract article title for logging
+    const titleMatch = /<h1[^>]*>([^<]+)<\/h1>/i.exec(html);
+    const articleTitle = titleMatch
+      ? titleMatch[1].trim()
+      : 'Wikipedia article';
+
+    console.log(`Processing "${articleTitle}"...`);
+
+    // Measure processing time
+    const startTime = process.hrtime();
+
+    // Process content
+    const result = processContent(html, markdown);
+
+    // Calculate processing time
+    const elapsed = process.hrtime(startTime);
+    const timeInMs = elapsed[0] * 1000 + elapsed[1] / 1000000;
+
+    // Generate reference appendix
+    const appendix = generateReferenceAppendix(result);
+
+    // Create complete output with the processed content and appendix
+    const completeOutput = result.markdown + appendix;
+
+    // Write to output file
+    fs.writeFileSync(outputPath, completeOutput);
+
+    // Print processing statistics
+    console.log('\nWikipedia article processing complete! ✓');
+    console.log('-'.repeat(60));
+    console.log(`Article: ${articleTitle}`);
+    console.log(`Processing time: ${timeInMs.toFixed(2)}ms`);
+    console.log('Media references replaced:');
+    console.log(`  - Links: ${result.links.length}`);
+    console.log(`  - Images: ${result.images.length}`);
+    console.log(`  - Videos: ${result.videos.length}`);
+    console.log(
+      `  - Total: ${result.links.length + result.images.length + result.videos.length}`
+    );
+    console.log(`Output saved to: ${outputPath}`);
+    console.log('-'.repeat(60));
+
+    // Print sample of the transformation
+    const sampleLines = result.markdown.split('\n').slice(0, 10).join('\n');
+    console.log('\nSample of transformed content:');
+    console.log('-'.repeat(30));
+    console.log(sampleLines);
+    console.log('-'.repeat(30));
+    console.log('... (continued in output file)');
+  } catch (error) {
+    console.error('Error processing Wikipedia article:', error);
+    process.exit(1);
+  }
+}
+
+/**
+ * Generate a comprehensive reference appendix with all media links
+ */
+function generateReferenceAppendix(result: {
+  links: Array<{ originalUrl: string; title?: string; text?: string }>;
+  images: Array<{ originalUrl: string; title?: string }>;
+  videos: Array<{ originalUrl: string; title?: string }>;
+}): string {
+  let appendix = '\n\n' + '---'.repeat(10) + '\n\n';
+  appendix += '# References\n\n';
+
+  if (result.links.length > 0) {
+    appendix += '## Links\n\n';
+    result.links.forEach((link, index) => {
+      // Clean and format text for display
+      let displayText = '';
+      if (link.text != null && link.text.trim()) {
+        // Limit length for very long link text
+        let cleanText = link.text.trim();
+        if (cleanText.length > 50) {
+          cleanText = cleanText.substring(0, 47) + '...';
+        }
+        displayText = ` - "${cleanText}"`;
+      }
+
+      appendix += `**link#${index + 1}**: ${link.originalUrl}${displayText}\n\n`;
+    });
+  }
+
+  if (result.images.length > 0) {
+    appendix += '## Images\n\n';
+    result.images.forEach((image, index) => {
+      const displayTitle =
+        image.title != null && image.title.trim()
+          ? ` - ${image.title.trim()}`
+          : '';
+      appendix += `**image#${index + 1}**: ${image.originalUrl}${displayTitle}\n\n`;
+    });
+  }
+
+  if (result.videos.length > 0) {
+    appendix += '## Videos\n\n';
+    result.videos.forEach((video, index) => {
+      const displayTitle =
+        video.title != null && video.title.trim()
+          ? ` - ${video.title.trim()}`
+          : '';
+      appendix += `**video#${index + 1}**: ${video.originalUrl}${displayTitle}\n\n`;
+    });
+  }
+
+  // Add a category breakdown to show what types of links were found
+  const totalRefs =
+    result.links.length + result.images.length + result.videos.length;
+
+  appendix += '## Summary\n\n';
+  appendix += `Total references: **${totalRefs}**\n\n`;
+  appendix += `- Links: ${result.links.length}\n`;
+  appendix += `- Images: ${result.images.length}\n`;
+  appendix += `- Videos: ${result.videos.length}\n`;
+
+  return appendix;
+}
+
+// Using async IIFE to allow for better error handling
+(async (): Promise<void> => {
+  try {
+    await processWikipediaArticle();
+  } catch (error) {
+    console.error('Unhandled error:', error);
+    process.exit(1);
+  }
+})();

package/src/tools/search/tool.ts

@@ -1,25 +1,339 @@
-/* eslint-disable no-console */
 import { z } from 'zod';
 import { tool, DynamicStructuredTool } from '@langchain/core/tools';
+import type { RunnableConfig } from '@langchain/core/runnables';
 import type * as t from './types';
+import {
+  DATE_RANGE,
+  querySchema,
+  dateSchema,
+  countrySchema,
+  imagesSchema,
+  videosSchema,
+  newsSchema,
+} from './schema';
 import { createSearchAPI, createSourceProcessor } from './search';
 import { createFirecrawlScraper } from './firecrawl';
 import { expandHighlights } from './highlights';
 import { formatResultsForLLM } from './format';
+import { createDefaultLogger } from './utils';
 import { createReranker } from './rerankers';
 import { Constants } from '@/common';
 
-const SearchToolSchema = z.object({
-  query: z
-    .string()
-    .describe(
-      'The search query string that specifies what should be searched for.'
-    ),
-});
+/**
+ * Executes parallel searches and merges the results
+ */
+async function executeParallelSearches({
+  searchAPI,
+  query,
+  date,
+  country,
+  safeSearch,
+  images,
+  videos,
+  news,
+  logger,
+}: {
+  searchAPI: ReturnType<typeof createSearchAPI>;
+  query: string;
+  date?: DATE_RANGE;
+  country?: string;
+  safeSearch: t.SearchToolConfig['safeSearch'];
+  images: boolean;
+  videos: boolean;
+  news: boolean;
+  logger: t.Logger;
+}): Promise<t.SearchResult> {
+  // Prepare all search tasks to run in parallel
+  const searchTasks: Promise<t.SearchResult>[] = [
+    // Main search
+    searchAPI.getSources({
+      query,
+      date,
+      country,
+      safeSearch,
+    }),
+  ];
 
+  if (images) {
+    searchTasks.push(
+      searchAPI
+        .getSources({
+          query,
+          date,
+          country,
+          safeSearch,
+          type: 'images',
+        })
+        .catch((error) => {
+          logger.error('Error fetching images:', error);
+          return {
+            success: false,
+            error: `Images search failed: ${error instanceof Error ? error.message : String(error)}`,
+          };
+        })
+    );
+  }
+  if (videos) {
+    searchTasks.push(
+      searchAPI
+        .getSources({
+          query,
+          date,
+          country,
+          safeSearch,
+          type: 'videos',
+        })
+        .catch((error) => {
+          logger.error('Error fetching videos:', error);
+          return {
+            success: false,
+            error: `Videos search failed: ${error instanceof Error ? error.message : String(error)}`,
+          };
+        })
+    );
+  }
+  if (news) {
+    searchTasks.push(
+      searchAPI
+        .getSources({
+          query,
+          date,
+          country,
+          safeSearch,
+          type: 'news',
+        })
+        .catch((error) => {
+          logger.error('Error fetching news:', error);
+          return {
+            success: false,
+            error: `News search failed: ${error instanceof Error ? error.message : String(error)}`,
+          };
+        })
+    );
+  }
+
+  // Run all searches in parallel
+  const results = await Promise.all(searchTasks);
+
+  // Get the main search result (first result)
+  const mainResult = results[0];
+  if (!mainResult.success) {
+    throw new Error(mainResult.error ?? 'Search failed');
+  }
+
+  // Merge additional results with the main results
+  const mergedResults = { ...mainResult.data };
+
+  // Convert existing news to topStories if present
+  if (mergedResults.news !== undefined && mergedResults.news.length > 0) {
+    const existingNewsAsTopStories = mergedResults.news
+      .filter((newsItem) => newsItem.link !== undefined && newsItem.link !== '')
+      .map((newsItem) => ({
+        title: newsItem.title ?? '',
+        link: newsItem.link ?? '',
+        source: newsItem.source ?? '',
+        date: newsItem.date ?? '',
+        imageUrl: newsItem.imageUrl ?? '',
+        processed: false,
+      }));
+    mergedResults.topStories = [
+      ...(mergedResults.topStories ?? []),
+      ...existingNewsAsTopStories,
+    ];
+    delete mergedResults.news;
+  }
+
+  results.slice(1).forEach((result) => {
+    if (result.success && result.data !== undefined) {
+      if (result.data.images !== undefined && result.data.images.length > 0) {
+        mergedResults.images = [
+          ...(mergedResults.images ?? []),
+          ...result.data.images,
+        ];
+      }
+      if (result.data.videos !== undefined && result.data.videos.length > 0) {
+        mergedResults.videos = [
+          ...(mergedResults.videos ?? []),
+          ...result.data.videos,
+        ];
+      }
+      if (result.data.news !== undefined && result.data.news.length > 0) {
+        const newsAsTopStories = result.data.news.map((newsItem) => ({
+          ...newsItem,
+          link: newsItem.link ?? '',
+        }));
+        mergedResults.topStories = [
+          ...(mergedResults.topStories ?? []),
+          ...newsAsTopStories,
+        ];
+      }
+    }
+  });
+
+  return { success: true, data: mergedResults };
+}
+
+function createSearchProcessor({
+  searchAPI,
+  safeSearch,
+  sourceProcessor,
+  onGetHighlights,
+  logger,
+}: {
+  safeSearch: t.SearchToolConfig['safeSearch'];
+  searchAPI: ReturnType<typeof createSearchAPI>;
+  sourceProcessor: ReturnType<typeof createSourceProcessor>;
+  onGetHighlights: t.SearchToolConfig['onGetHighlights'];
+  logger: t.Logger;
+}) {
+  return async function ({
+    query,
+    date,
+    country,
+    proMode = true,
+    maxSources = 5,
+    onSearchResults,
+    images = false,
+    videos = false,
+    news = false,
+  }: {
+    query: string;
+    country?: string;
+    date?: DATE_RANGE;
+    proMode?: boolean;
+    maxSources?: number;
+    onSearchResults: t.SearchToolConfig['onSearchResults'];
+    images?: boolean;
+    videos?: boolean;
+    news?: boolean;
+  }): Promise<t.SearchResultData> {
+    try {
+      // Execute parallel searches and merge results
+      const searchResult = await executeParallelSearches({
+        searchAPI,
+        query,
+        date,
+        country,
+        safeSearch,
+        images,
+        videos,
+        news,
+        logger,
+      });
+
+      onSearchResults?.(searchResult);
+
+      const processedSources = await sourceProcessor.processSources({
+        query,
+        news,
+        result: searchResult,
+        proMode,
+        onGetHighlights,
+        numElements: maxSources,
+      });
+
+      return expandHighlights(processedSources);
+    } catch (error) {
+      logger.error('Error in search:', error);
+      return {
+        organic: [],
+        topStories: [],
+        images: [],
+        videos: [],
+        news: [],
+        relatedSearches: [],
+        error: error instanceof Error ? error.message : String(error),
+      };
+    }
+  };
+}
+
+function createOnSearchResults({
+  runnableConfig,
+  onSearchResults,
+}: {
+  runnableConfig: RunnableConfig;
+  onSearchResults: t.SearchToolConfig['onSearchResults'];
+}) {
+  return function (results: t.SearchResult): void {
+    if (!onSearchResults) {
+      return;
+    }
+    onSearchResults(results, runnableConfig);
+  };
+}
+
+function createTool({
+  schema,
+  search,
+  onSearchResults: _onSearchResults,
+}: {
+  schema: t.SearchToolSchema;
+  search: ReturnType<typeof createSearchProcessor>;
+  onSearchResults: t.SearchToolConfig['onSearchResults'];
+}): DynamicStructuredTool<typeof schema> {
+  return tool<typeof schema>(
+    async (params, runnableConfig) => {
+      const { query, date, country: _c, images, videos, news } = params;
+      const country = typeof _c === 'string' && _c ? _c : undefined;
+      const searchResult = await search({
+        query,
+        date,
+        country,
+        images,
+        videos,
+        news,
+        onSearchResults: createOnSearchResults({
+          runnableConfig,
+          onSearchResults: _onSearchResults,
+        }),
+      });
+      const turn = runnableConfig.toolCall?.turn ?? 0;
+      const { output, references } = formatResultsForLLM(turn, searchResult);
+      const data: t.SearchResultData = { turn, ...searchResult, references };
+      return [output, { [Constants.WEB_SEARCH]: data }];
+    },
+    {
+      name: Constants.WEB_SEARCH,
+      description: `Real-time search. Results have required citation anchors.
+
+Note: Use ONCE per reply unless instructed otherwise.
+
+Anchors:
+- \\ue202turnXtypeY
+- X = turn idx, type = 'search' | 'news' | 'image' | 'ref', Y = item idx
+
+Special Markers:
+- \\ue203...\\ue204 — highlight start/end of cited text (for Standalone or Group citations)
+- \\ue200...\\ue201 — group block (e.g. \\ue200\\ue202turn0search1\\ue202turn0news2\\ue201)
+
+**CITE EVERY NON-OBVIOUS FACT/QUOTE:**
+Use anchor marker(s) immediately after the statement:
+- Standalone: "Pure functions produce same output. \\ue202turn0search0"
+- Standalone (multiple): "Today's News \\ue202turn0search0\\ue202turn0news0"
+- Highlight: "\\ue203Highlight text.\\ue204\\ue202turn0news1"
+- Group: "Sources. \\ue200\\ue202turn0search0\\ue202turn0news1\\ue201"
+- Group Highlight: "\\ue203Highlight for group.\\ue204 \\ue200\\ue202turn0search0\\ue202turn0news1\\ue201"
+- Image: "See photo \\ue202turn0image0."
+
+**NEVER use markdown links, [1], or footnotes. CITE ONLY with anchors provided.**
+`.trim(),
+      schema: schema,
+      responseFormat: Constants.CONTENT_AND_ARTIFACT,
+    }
+  );
+}
+
+/**
+ * Creates a search tool with a schema that dynamically includes the country field
+ * only when the searchProvider is 'serper'.
+ *
+ * @param config - The search tool configuration
+ * @returns A DynamicStructuredTool with a schema that depends on the searchProvider
+ */
 export const createSearchTool = (
   config: t.SearchToolConfig = {}
-): DynamicStructuredTool<typeof SearchToolSchema> => {
+): DynamicStructuredTool<typeof toolSchema> => {
   const {
     searchProvider = 'serper',
     serperApiKey,
@@ -29,14 +343,39 @@ export const createSearchTool = (
     topResults = 5,
     strategies = ['no_extraction'],
     filterContent = true,
+    safeSearch = 1,
     firecrawlApiKey,
     firecrawlApiUrl,
     firecrawlFormats = ['markdown', 'html'],
     jinaApiKey,
     cohereApiKey,
     onSearchResults: _onSearchResults,
+    onGetHighlights,
   } = config;
 
+  const logger = config.logger || createDefaultLogger();
+
+  const schemaObject: {
+    query: z.ZodString;
+    date: z.ZodOptional<z.ZodNativeEnum<typeof DATE_RANGE>>;
+    country?: z.ZodOptional<z.ZodString>;
+    images: z.ZodOptional<z.ZodBoolean>;
+    videos: z.ZodOptional<z.ZodBoolean>;
+    news: z.ZodOptional<z.ZodBoolean>;
+  } = {
+    query: querySchema,
+    date: dateSchema,
+    images: imagesSchema,
+    videos: videosSchema,
+    news: newsSchema,
+  };
+
+  if (searchProvider === 'serper') {
+    schemaObject.country = countrySchema;
+  }
+
+  const toolSchema = z.object(schemaObject);
+
   const searchAPI = createSearchAPI({
     searchProvider,
     serperApiKey,
@@ -54,10 +393,11 @@ export const createSearchTool = (
     rerankerType,
     jinaApiKey,
     cohereApiKey,
+    logger,
   });
 
   if (!selectedReranker) {
-    console.warn('No reranker selected. Using default ranking.');
+    logger.warn('No reranker selected. Using default ranking.');
   }
 
   const sourceProcessor = createSourceProcessor(
@@ -66,86 +406,22 @@ export const createSearchTool = (
       topResults,
       strategies,
       filterContent,
+      logger,
     },
     firecrawlScraper
   );
 
-  const search = async ({
-    query,
-    proMode = true,
-    maxSources = 5,
-    onSearchResults,
-  }: {
-    query: string;
-    proMode?: boolean;
-    maxSources?: number;
-    onSearchResults?: (sources: t.SearchResult) => void;
-  }): Promise<t.SearchResultData> => {
-    try {
-      const sources = await searchAPI.getSources(query);
-      onSearchResults?.(sources);
-
-      if (!sources.success) {
-        throw new Error(sources.error ?? 'Search failed');
-      }
-
-      const processedSources = await sourceProcessor.processSources(
-        sources,
-        maxSources,
-        query,
-        proMode
-      );
-      return expandHighlights(processedSources);
-    } catch (error) {
-      console.error('Error in search:', error);
-      return {
-        organic: [],
-        topStories: [],
-        images: [],
-        relatedSearches: [],
-        error: error instanceof Error ? error.message : String(error),
-      };
-    }
-  };
-
-  return tool<typeof SearchToolSchema>(
-    async ({ query }, runnableConfig) => {
-      const searchResult = await search({
-        query,
-        onSearchResults: _onSearchResults
-          ? (result): void => {
-            _onSearchResults(result, runnableConfig);
-          }
-          : undefined,
-      });
-      const output = formatResultsForLLM(searchResult);
-      return [output, searchResult];
-    },
-    {
-      name: Constants.WEB_SEARCH,
-      description: `
-Real-time search. Results have required unique citation anchors.
-
-Anchors:
-- \\ue202turn0searchN (web), \\ue202turn0newsN (news), \\ue202turn0imageN (image)
-
-Special Markers:
-- \\ue203...\\ue204 — mark start/end of cited span
-- \\ue200...\\ue201 — composite/group block (e.g. \\ue200cite\\ue202turn0search1\\ue202turn0news2\\ue201)
-- \\ue206 — marks grouped/summary citation areas
-
-**CITE EVERY NON-OBVIOUS FACT/QUOTE:**
-Insert the anchor marker(s) immediately after the statement:
-- "Pure functions produce same output \\ue202turn0search0."
-- Multiple: "Benefits \\ue202turn0search0\\ue202turn0news0."
-- Span: \\ue203Key: first-class functions\\ue204\\ue202turn0news1
-- Group: "Functional languages."\\ue206 or \\ue200cite\\ue202turn0search0\\ue202turn0news1\\ue201
-- Image: "See photo \\ue202turn0image0."
+  const search = createSearchProcessor({
+    searchAPI,
+    safeSearch,
+    sourceProcessor,
+    onGetHighlights,
+    logger,
+  });
 
-**NEVER use markdown links, [1], or footnotes. CITE ONLY with anchors provided.**
-`.trim(),
-      schema: SearchToolSchema,
-      responseFormat: Constants.CONTENT_AND_ARTIFACT,
-    }
-  );
+  return createTool({
+    search,
+    schema: toolSchema,
+    onSearchResults: _onSearchResults,
+  });
 };