feat(api): increase search depth for notebook-scoped chat queries

Notebook @mentions in normal chat used limit=3/collection and cap=8 results,
while dedicated notebook chats use limit=30 and cap=40. This boosts notebook-
scoped searches to limit=10, cap=20, reranker 20→12, and 8000 char context
budget — all gated on notebookCollectionIds so general chat is unaffected.

Author: Movm

apps/api/agents/langgraph/ChatGraph/nodes/rerankNode.ts

@@ -52,13 +52,18 @@ export async function rerankNode(state: ChatGraphState): Promise<Partial<ChatGra
   const startTime = Date.now();
   const { searchResults, searchQuery, aiWorkerPool, hasTemporal } = state;
 
+  // Notebook-scoped searches get higher limits for deeper recall
+  const isNotebookScoped = (state.notebookCollectionIds?.length ?? 0) > 0;
+  const inputLimit = isNotebookScoped ? 20 : RERANK_INPUT_LIMIT;
+  const outputLimit = isNotebookScoped ? 12 : RERANK_OUTPUT_LIMIT;
+
   // Skip reranking if too few results
   if (searchResults.length <= 3) {
     log.info(`[Rerank] Skipping — only ${searchResults.length} results`);
     return { rerankTimeMs: Date.now() - startTime };
   }
 
-  const candidates = searchResults.slice(0, RERANK_INPUT_LIMIT);
+  const candidates = searchResults.slice(0, inputLimit);
 
   log.info(
     `[Rerank] Reranking ${candidates.length} results for query: "${searchQuery?.slice(0, 50)}..."`
@@ -115,7 +120,7 @@ ${passageList}`;
     // B3: Apply MMR diversity reranking as second pass
     // Keep top 2 results unchanged, apply diversity for positions 3+
     const diverse = filtered.length > 3 ? applyMMR(filtered, 0.7, 2) : filtered;
-    const reranked = diverse.slice(0, RERANK_OUTPUT_LIMIT);
+    const reranked = diverse.slice(0, outputLimit);
 
     log.info(
       `[Rerank] Complete: ${candidates.length} → ${reranked.length} results (diversity applied) in ${rerankTimeMs}ms`

apps/api/agents/langgraph/ChatGraph/nodes/respondNode.ts

@@ -203,7 +203,10 @@ async function formatSearchContext(state: ChatGraphState): Promise<string> {
   }
 
   // Default: budget-based truncation
-  const topResults = state.searchResults.slice(0, MAX_SEARCH_RESULTS);
+  // Notebook-scoped searches get more results and higher budget for deeper answers
+  const isNotebookScoped = (state.notebookCollectionIds?.length ?? 0) > 0;
+  const maxResults = isNotebookScoped ? 12 : MAX_SEARCH_RESULTS;
+  const topResults = state.searchResults.slice(0, maxResults);
 
   // Document chat gets the highest budget for focused Q&A
   const isDocumentChat = state.documentChatIds?.length > 0;
@@ -213,9 +216,11 @@ async function formatSearchContext(state: ChatGraphState): Promise<string> {
   const isMultiSource = (state.searchSources?.length || 0) > 1;
   const budget = isDocumentChat
     ? SEARCH_CONTEXT_BUDGET_DOCUMENTCHAT
-    : hasCrawledContent || isMultiSource
-      ? SEARCH_CONTEXT_BUDGET_CRAWLED
-      : SEARCH_CONTEXT_BUDGET;
+    : isNotebookScoped
+      ? SEARCH_CONTEXT_BUDGET_DOCUMENTCHAT
+      : hasCrawledContent || isMultiSource
+        ? SEARCH_CONTEXT_BUDGET_CRAWLED
+        : SEARCH_CONTEXT_BUDGET;
 
   // Crawled results get 2x weight in budget allocation
   const weightedRelevance = topResults.map((r) => {

apps/api/agents/langgraph/ChatGraph/nodes/searchNode.ts

@@ -480,18 +480,22 @@ export async function searchNode(state: ChatGraphState): Promise<Partial<ChatGra
         }
         // Deduplicate in case of overlap
         const uniqueCollections = [...new Set(collectionsToSearch)];
+        const isNotebookScoped =
+          state.notebookCollectionIds && state.notebookCollectionIds.length > 0;
 
         const query = searchQuery || '';
 
         // Search all sub-queries (if decomposed) across all collections
+        // Notebook-scoped searches get deeper recall (10 vs 3 per collection)
         const subQueries = state.subQueries?.length ? state.subQueries : [query];
+        const perCollectionLimit = isNotebookScoped ? 10 : 3;
 
         const searchPromises = uniqueCollections.flatMap((collection) =>
           subQueries.map((sq) =>
             executeDirectSearch({
               query: sq,
               collection,
-              limit: 3,
+              limit: perCollectionLimit,
               filters: detectedFilters || undefined,
             }).catch((err: any) => {
               log.warn(
@@ -531,8 +535,10 @@ export async function searchNode(state: ChatGraphState): Promise<Partial<ChatGra
         }
 
         // Sort by relevance and take top results
+        // Notebook-scoped searches keep more candidates for reranking
         allResults.sort((a, b) => (b.relevance || 0) - (a.relevance || 0));
-        results = allResults.slice(0, 8);
+        const resultsCap = isNotebookScoped ? 20 : 8;
+        results = allResults.slice(0, resultsCap);
         citations = buildCitations(results);
 
         // Track which collections were searched for observability