@librechat/agents 2.4.30 → 2.4.33

package/src/tools/search/format.ts

@@ -0,0 +1,252 @@
+import type * as t from './types';
+import { getDomainName, fileExtRegex } from './utils';
+
+function addHighlightSection(): string[] {
+  return ['\n## Highlights', ''];
+}
+
+// Helper function to format a source (organic or top story)
+function formatSource(
+  source: t.ValidSource,
+  index: number,
+  turn: number,
+  sourceType: 'search' | 'news',
+  references: t.ResultReference[]
+): string {
+  /** Array of all lines to include in the output */
+  const outputLines: string[] = [];
+
+  // Add the title
+  outputLines.push(
+    `# ${sourceType.charAt(0).toUpperCase() + sourceType.slice(1)} ${index}: ${source.title != null && source.title ? `"${source.title}"` : '(no title)'}`
+  );
+  outputLines.push(`\nAnchor: \\ue202turn${turn}${sourceType}${index}`);
+  outputLines.push(`URL: ${source.link}`);
+
+  // Add optional fields
+  if ('snippet' in source && source.snippet != null) {
+    outputLines.push(`Summary: ${source.snippet}`);
+  }
+
+  if (source.date != null) {
+    outputLines.push(`Date: ${source.date}`);
+  }
+
+  if (source.attribution != null) {
+    outputLines.push(`Source: ${source.attribution}`);
+  }
+
+  // Add highlight section or empty line
+  if ((source.highlights?.length ?? 0) > 0) {
+    outputLines.push(...addHighlightSection());
+  } else {
+    outputLines.push('');
+  }
+
+  // Process highlights if they exist
+  (source.highlights ?? [])
+    .filter((h) => h.text.trim().length > 0)
+    .forEach((h, hIndex) => {
+      outputLines.push(
+        `### Highlight ${hIndex + 1} [Relevance: ${h.score.toFixed(2)}]`
+      );
+      outputLines.push('');
+      outputLines.push('```text');
+      outputLines.push(h.text.trim());
+      outputLines.push('```');
+      outputLines.push('');
+
+      if (h.references != null && h.references.length) {
+        let hasHeader = false;
+        const refLines: string[] = [];
+
+        for (let j = 0; j < h.references.length; j++) {
+          const ref = h.references[j];
+          if (ref.reference.originalUrl.includes('mailto:')) {
+            continue;
+          }
+          references.push({
+            type: ref.type,
+            link: ref.reference.originalUrl,
+            attribution: getDomainName(ref.reference.originalUrl),
+            title: (
+              ((ref.reference.title ?? '') || ref.reference.text) ??
+              ''
+            ).split('\n')[0],
+          });
+
+          if (ref.type !== 'link') {
+            continue;
+          }
+
+          if (fileExtRegex.test(ref.reference.originalUrl)) {
+            continue;
+          }
+
+          if (!hasHeader) {
+            refLines.push('Core References:');
+            hasHeader = true;
+          }
+
+          refLines.push(
+            `- ${ref.type}#${ref.originalIndex + 1}: ${ref.reference.originalUrl}`
+          );
+          refLines.push(
+            `\t- Anchor: \\ue202turn${turn}ref${references.length - 1}`
+          );
+        }
+
+        if (hasHeader) {
+          outputLines.push(...refLines);
+          outputLines.push('');
+        }
+      }
+
+      if (hIndex < (source.highlights?.length ?? 0) - 1) {
+        outputLines.push('---');
+        outputLines.push('');
+      }
+    });
+
+  outputLines.push('');
+  return outputLines.join('\n');
+}
+
+export function formatResultsForLLM(
+  turn: number,
+  results: t.SearchResultData
+): { output: string; references: t.ResultReference[] } {
+  /** Array to collect all output lines */
+  const outputLines: string[] = [];
+
+  const addSection = (title: string): void => {
+    outputLines.push('');
+    outputLines.push(`=== ${title} ===`);
+    outputLines.push('');
+  };
+
+  const references: t.ResultReference[] = [];
+
+  // Organic (web) results
+  if (results.organic?.length != null && results.organic.length > 0) {
+    addSection(`Web Results, Turn ${turn}`);
+    for (let i = 0; i < results.organic.length; i++) {
+      const r = results.organic[i];
+      outputLines.push(formatSource(r, i, turn, 'search', references));
+      delete results.organic[i].highlights;
+    }
+  }
+
+  // Top stories (news)
+  const topStories = results.topStories ?? [];
+  if (topStories.length) {
+    addSection('News Results');
+    for (let i = 0; i < topStories.length; i++) {
+      const r = topStories[i];
+      outputLines.push(formatSource(r, i, turn, 'news', references));
+      if (results.topStories?.[i]?.highlights) {
+        delete results.topStories[i].highlights;
+      }
+    }
+  }
+
+  // // Images
+  // const images = results.images ?? [];
+  // if (images.length) {
+  //   addSection('Image Results');
+  //   const imageLines = images.map((img, i) => [
+  //     `Anchor: \ue202turn0image${i}`,
+  //     `Title: ${img.title ?? '(no title)'}`,
+  //     `Image URL: ${img.imageUrl}`,
+  //     ''
+  //   ].join('\n'));
+  //   outputLines.push(imageLines.join('\n'));
+  // }
+
+  // Knowledge Graph
+  if (results.knowledgeGraph != null) {
+    addSection('Knowledge Graph');
+    const kgLines = [
+      `**Title:** ${results.knowledgeGraph.title ?? '(no title)'}`,
+      results.knowledgeGraph.type != null
+        ? `**Type:** ${results.knowledgeGraph.type}`
+        : '',
+      results.knowledgeGraph.description != null
+        ? `**Description:** ${results.knowledgeGraph.description}`
+        : '',
+      results.knowledgeGraph.descriptionSource != null
+        ? `**Description Source:** ${results.knowledgeGraph.descriptionSource}`
+        : '',
+      results.knowledgeGraph.descriptionLink != null
+        ? `**Description Link:** ${results.knowledgeGraph.descriptionLink}`
+        : '',
+      results.knowledgeGraph.imageUrl != null
+        ? `**Image URL:** ${results.knowledgeGraph.imageUrl}`
+        : '',
+      results.knowledgeGraph.website != null
+        ? `**Website:** ${results.knowledgeGraph.website}`
+        : '',
+      results.knowledgeGraph.attributes != null
+        ? `**Attributes:**\n\`\`\`json\n${JSON.stringify(
+          results.knowledgeGraph.attributes,
+          null,
+          2
+        )}\n\`\`\``
+        : '',
+      '',
+    ].filter(Boolean);
+
+    outputLines.push(kgLines.join('\n\n'));
+  }
+
+  // Answer Box
+  if (results.answerBox != null) {
+    addSection('Answer Box');
+    const abLines = [
+      results.answerBox.title != null
+        ? `**Title:** ${results.answerBox.title}`
+        : '',
+      results.answerBox.snippet != null
+        ? `**Snippet:** ${results.answerBox.snippet}`
+        : '',
+      results.answerBox.snippetHighlighted != null
+        ? `**Snippet Highlighted:** ${results.answerBox.snippetHighlighted
+          .map((s) => `\`${s}\``)
+          .join(' ')}`
+        : '',
+      results.answerBox.link != null
+        ? `**Link:** ${results.answerBox.link}`
+        : '',
+      '',
+    ].filter(Boolean);
+
+    outputLines.push(abLines.join('\n\n'));
+  }
+
+  // People also ask
+  const peopleAlsoAsk = results.peopleAlsoAsk ?? [];
+  if (peopleAlsoAsk.length) {
+    addSection('People Also Ask');
+
+    const paaLines: string[] = [];
+    peopleAlsoAsk.forEach((p, i) => {
+      const questionLines = [
+        `### Question ${i + 1}:`,
+        `"${p.question}"`,
+        `${p.snippet != null && p.snippet ? `Snippet: ${p.snippet}` : ''}`,
+        `${p.title != null && p.title ? `Title: ${p.title}` : ''}`,
+        `${p.link != null && p.link ? `Link: ${p.link}` : ''}`,
+        '',
+      ].filter(Boolean);
+
+      paaLines.push(questionLines.join('\n\n'));
+    });
+
+    outputLines.push(paaLines.join(''));
+  }
+
+  return {
+    output: outputLines.join('\n').trim(),
+    references,
+  };
+}

package/src/tools/search/highlights.ts

@@ -0,0 +1,320 @@
+import type * as t from './types';
+
+// 2. Pre-compile all regular expressions (only do this once)
+// Group patterns by priority for early returns
+const priorityPatterns = [
+  // High priority patterns (structural)
+  [
+    { regex: /\n\n/g }, // Double newline (paragraph break)
+    { regex: /\n/g }, // Single newline
+    { regex: /={3,}\s*\n|-{3,}\s*\n/g }, // Section separators
+  ],
+  // Medium priority (semantic)
+  [
+    { regex: /[.!?][")\]]?\s/g }, // End of sentence
+    { regex: /;\s/g }, // Semicolon
+    { regex: /:\s/g }, // Colon
+  ],
+  // Low priority (any breaks)
+  [
+    { regex: /,\s/g }, // Comma
+    { regex: /\s-\s/g }, // Dash surrounded by spaces
+    { regex: /\s/g }, // Any space
+  ],
+];
+
+function findFirstMatch(text: string, regex: RegExp): number {
+  // Reset regex
+  regex.lastIndex = 0;
+
+  // For very long texts, try chunking
+  if (text.length > 10000) {
+    const chunkSize = 2000;
+    let position = 0;
+
+    while (position < text.length) {
+      const chunk = text.substring(position, position + chunkSize);
+      regex.lastIndex = 0;
+
+      const match = regex.exec(chunk);
+      if (match) {
+        return position + match.index;
+      }
+
+      // Move to next chunk with some overlap
+      position += chunkSize - 100;
+      if (position >= text.length) break;
+    }
+    return -1;
+  }
+
+  // For shorter texts, normal regex search
+  const match = regex.exec(text);
+  return match ? match.index : -1;
+}
+
+// 3. Optimized boundary finding functions
+function findLastMatch(text: string, regex: RegExp): number {
+  // Reset regex state
+  regex.lastIndex = 0;
+
+  let lastIndex = -1;
+  let lastLength = 0;
+  let match;
+
+  // For very long texts, use a different approach to avoid regex engine slowdowns
+  if (text.length > 10000) {
+    // Try dividing the text into chunks for faster processing
+    const chunkSize = 2000;
+    let startPosition = Math.max(0, text.length - chunkSize);
+
+    while (startPosition >= 0) {
+      const chunk = text.substring(startPosition, startPosition + chunkSize);
+      regex.lastIndex = 0;
+
+      let chunkLastIndex = -1;
+      let chunkLastLength = 0;
+
+      while ((match = regex.exec(chunk)) !== null) {
+        chunkLastIndex = match.index;
+        chunkLastLength = match[0].length;
+      }
+
+      if (chunkLastIndex !== -1) {
+        return startPosition + chunkLastIndex + chunkLastLength;
+      }
+
+      // Move to previous chunk with some overlap
+      startPosition = Math.max(0, startPosition - chunkSize + 100) - 1;
+      if (startPosition <= 0) break;
+    }
+    return -1;
+  }
+
+  // For shorter texts, normal regex search
+  while ((match = regex.exec(text)) !== null) {
+    lastIndex = match.index;
+    lastLength = match[0].length;
+  }
+
+  return lastIndex === -1 ? -1 : lastIndex + lastLength;
+}
+
+// 4. Find the best boundary with priority groups
+function findBestBoundary(text: string, direction = 'backward'): number {
+  if (!text || text.length === 0) return 0;
+
+  // Try each priority group
+  for (const patternGroup of priorityPatterns) {
+    for (const pattern of patternGroup) {
+      const position =
+        direction === 'backward'
+          ? findLastMatch(text, pattern.regex)
+          : findFirstMatch(text, pattern.regex);
+
+      if (position !== -1) {
+        return position;
+      }
+    }
+  }
+
+  // No match found, use character boundary
+  return direction === 'backward' ? text.length : 0;
+}
+
+/**
+ * Tracks references used in a highlight without changing their numbers
+ */
+function trackReferencesInHighlight(
+  text: string,
+  sourceResult: t.ValidSource // Source containing the original references
+): {
+  references: {
+    type: 'link' | 'image' | 'video';
+    originalIndex: number;
+    reference: t.MediaReference; // Original reference object
+  }[];
+} {
+  // Track used references
+  const references: {
+    type: 'link' | 'image' | 'video';
+    originalIndex: number;
+    reference: t.MediaReference;
+  }[] = [];
+
+  if (!text || text.length === 0 || !text.includes('#')) {
+    return { references }; // Early return
+  }
+
+  // Quick check for reference markers
+  if (
+    !text.includes('link#') &&
+    !text.includes('image#') &&
+    !text.includes('video#')
+  ) {
+    return { references };
+  }
+
+  // Get references from the source if available
+  const sourceRefs = sourceResult.references || {
+    links: [],
+    images: [],
+    videos: [],
+  };
+
+  // Find references but don't modify text
+  const refRegex = /\((link|image|video)#(\d+)(?:\s+"([^"]*)")?\)/g;
+  let match;
+
+  while ((match = refRegex.exec(text)) !== null) {
+    const [, type, indexStr] = match;
+    const originalIndex = parseInt(indexStr, 10) - 1; // Convert to 0-based
+
+    // Get the source array for this type
+    const refType = type as 'link' | 'image' | 'video';
+    const sourceArray = sourceRefs[`${refType}s`] as
+      | t.MediaReference[]
+      | undefined;
+
+    // Skip if invalid reference
+    if (
+      !sourceArray ||
+      originalIndex < 0 ||
+      originalIndex >= sourceArray.length
+    ) {
+      continue; // Skip invalid references
+    }
+
+    // Get original reference
+    const reference = sourceArray[originalIndex];
+
+    // Track if not already tracked
+    const alreadyTracked = references.some(
+      (ref) => ref.type === refType && ref.originalIndex === originalIndex
+    );
+
+    if (!alreadyTracked) {
+      references.push({
+        type: refType,
+        originalIndex,
+        reference,
+      });
+    }
+  }
+
+  return { references };
+}
+
+/**
+ * Expand highlights in search results using smart boundary detection.
+ *
+ * This implementation finds natural text boundaries like paragraphs, sentences,
+ * and phrases to provide context while maintaining readability.
+ *
+ * @param searchResults - Search results object
+ * @param mainExpandBy - Primary expansion size on each side (default: 300)
+ * @param separatorExpandBy - Additional range to look for separators (default: 150)
+ * @returns Copy of search results with expanded highlights and tracked references
+ */
+export function expandHighlights(
+  searchResults: t.SearchResultData,
+  mainExpandBy = 300,
+  separatorExpandBy = 150
+): t.SearchResultData {
+  // Avoid deep copy - only copy what we modify
+  const resultCopy = { ...searchResults };
+  if (resultCopy.organic) resultCopy.organic = [...resultCopy.organic];
+  if (resultCopy.topStories) resultCopy.topStories = [...resultCopy.topStories];
+
+  // Process the results efficiently
+  const processResultTypes = ['organic', 'topStories'] as const;
+
+  for (const resultType of processResultTypes) {
+    if (!resultCopy[resultType as 'organic' | 'topStories']) continue;
+
+    // Map results to new array with modified highlights
+    resultCopy[resultType] = resultCopy[resultType]?.map((result) => {
+      if (
+        result.content == null ||
+        result.content === '' ||
+        !result.highlights ||
+        result.highlights.length === 0
+      ) {
+        return result; // No modification needed
+      }
+
+      // Create a shallow copy with expanded highlights
+      const resultCopy = { ...result };
+      const content = result.content;
+      const highlights = [];
+      // Process each highlight
+      for (const highlight of result.highlights) {
+        const { references } = trackReferencesInHighlight(
+          highlight.text,
+          result
+        );
+
+        let startPos = content.indexOf(highlight.text);
+        let highlightLen = highlight.text.length;
+
+        if (startPos === -1) {
+          // Try with stripped whitespace
+          const strippedHighlight = highlight.text.trim();
+          startPos = content.indexOf(strippedHighlight);
+
+          if (startPos === -1) {
+            highlights.push({
+              text: highlight.text,
+              score: highlight.score,
+              references,
+            });
+            continue;
+          }
+          highlightLen = strippedHighlight.length;
+        }
+
+        // Calculate boundaries
+        const mainStart = Math.max(0, startPos - mainExpandBy);
+        const mainEnd = Math.min(
+          content.length,
+          startPos + highlightLen + mainExpandBy
+        );
+
+        const separatorStart = Math.max(0, mainStart - separatorExpandBy);
+        const separatorEnd = Math.min(
+          content.length,
+          mainEnd + separatorExpandBy
+        );
+
+        // Extract text segments
+        const headText = content.substring(separatorStart, mainStart);
+        const tailText = content.substring(mainEnd, separatorEnd);
+
+        // Find natural boundaries
+        const bestHeadBoundary = findBestBoundary(headText, 'backward');
+        const bestTailBoundary = findBestBoundary(tailText, 'forward');
+
+        // Calculate final positions
+        const finalStart = separatorStart + bestHeadBoundary;
+        const finalEnd = mainEnd + bestTailBoundary;
+
+        // Extract the expanded highlight
+        const expandedHighlightText = content
+          .substring(finalStart, finalEnd)
+          .trim();
+        highlights.push({
+          text: expandedHighlightText,
+          score: highlight.score,
+          references,
+        });
+      }
+
+      resultCopy.highlights = highlights;
+      delete resultCopy.content;
+      delete resultCopy.references;
+      return resultCopy;
+    });
+  }
+
+  return resultCopy;
+}

package/src/tools/search/index.ts

@@ -0,0 +1,2 @@
+export * from './tool';
+export type * from './types';