feat: cross-document synthesis — query across all notebooks

CHANGELOG.md

@@ -2,6 +2,13 @@
 
 All notable changes to Notebook LM will be documented in this file.
 
+## [0.2.0.0] - 2026-04-15
+
+### Added
+- Cross-document synthesis: click "All" in the chat header to query across every notebook at once. Ask "what do these papers disagree about?" and get a sourced answer that names which notebook and document each claim came from.
+- Source panel now shows notebook labels when in cross-notebook mode, so you can trace every answer back to its origin.
+- LLM prompt explicitly instructs the model to compare, contrast, and name disagreements across sources from different notebooks.
+
 ## [0.1.1.0] - 2026-04-15
 
 ### Added

VERSION

@@ -1 +1 @@
-0.1.0.0
+0.2.0.0

apps/desktop/src/components/chat/ChatView.tsx

@@ -24,6 +24,8 @@ export function ChatView({ pendingSuggest, onSuggestConsumed }: { pendingSuggest
   const notebooks = useAppStore((s) => s.notebooks);
   const toggleSourcePanel = useAppStore((s) => s.toggleSourcePanel);
   const status = useAppStore((s) => s.status);
+  const crossNotebookMode = useAppStore((s) => s.crossNotebookMode);
+  const setCrossNotebookMode = useAppStore((s) => s.setCrossNotebookMode);
 
   const activeNotebook = notebooks.find((nb) => nb.notebook_id === activeNotebookId) ?? null;
 
@@ -113,14 +115,31 @@ export function ChatView({ pendingSuggest, onSuggestConsumed }: { pendingSuggest
     <div className="chat-view">
       <div className="chat-header">
         <div className="chat-header-title">
-          <h2>{activeNotebook ? activeNotebook.title : 'Notebook LM'}</h2>
-          {activeNotebook && (
+          <h2>{crossNotebookMode ? 'All Notebooks' : (activeNotebook ? activeNotebook.title : 'Notebook LM')}</h2>
+          {crossNotebookMode ? (
+            <span className="chat-header-meta">
+              {notebooks.length} {notebooks.length === 1 ? 'notebook' : 'notebooks'}
+            </span>
+          ) : activeNotebook ? (
             <span className="chat-header-meta">
               {activeNotebook.source_count} {activeNotebook.source_count === 1 ? 'doc' : 'docs'}
             </span>
-          )}
+          ) : null}
         </div>
         <div className="chat-header-actions">
+          {notebooks.length > 1 && (
+            <button
+              type="button"
+              className={`chat-header-btn ${crossNotebookMode ? 'chat-header-btn-active' : ''}`}
+              onClick={() => {
+                setCrossNotebookMode(!crossNotebookMode);
+                clearChat();
+              }}
+              title={crossNotebookMode ? 'Switch to single notebook' : 'Query across all notebooks'}
+            >
+              {crossNotebookMode ? 'Single' : 'All'}
+            </button>
+          )}
           {messages.length > 0 && (
             <button type="button" className="chat-header-btn" onClick={clearChat}>
               New chat

apps/desktop/src/components/chat/chat.css

@@ -78,6 +78,17 @@
   transform: scale(0.97);
 }
 
+.chat-header-btn-active {
+  background: var(--color-accent);
+  color: #0c0a09;
+  border-color: var(--color-accent);
+}
+
+.chat-header-btn-active:hover {
+  background: var(--color-accent-hover);
+  color: #0c0a09;
+}
+
 /* ---- Message area ---- */
 
 .chat-messages {

apps/desktop/src/components/layout/SourcePanel.tsx

@@ -9,6 +9,8 @@ function relevanceColor(score: number): string {
 
 export function SourcePanel() {
   const activeSources = useAppStore((s) => s.activeSources);
+  const crossNotebookMode = useAppStore((s) => s.crossNotebookMode);
+  const notebooks = useAppStore((s) => s.notebooks);
 
   if (activeSources.length === 0) return null;
 
@@ -20,23 +22,31 @@ export function SourcePanel() {
       </div>
 
       <div className="source-panel-list">
-        {activeSources.map((source, i) => (
-          <div key={`${source.source_path}-${i}`} className="source-card">
-            <span className="source-card-name">{source.document_name}</span>
-            {source.relevance_score != null && (
-              <div className="source-relevance-track">
-                <div
-                  className="source-relevance-bar"
-                  style={{
-                    width: `${source.relevance_score}%`,
-                    backgroundColor: relevanceColor(source.relevance_score),
-                  }}
-                />
-              </div>
-            )}
-            <p className="source-card-preview">{source.preview}</p>
-          </div>
-        ))}
+        {activeSources.map((source, i) => {
+          const nbId = (source as Record<string, unknown>).notebook_id as string | undefined;
+          const nbName = nbId ? notebooks.find((nb) => nb.notebook_id === nbId)?.title : null;
+
+          return (
+            <div key={`${source.source_path}-${i}`} className="source-card">
+              {crossNotebookMode && nbName && (
+                <span className="source-card-notebook">{nbName}</span>
+              )}
+              <span className="source-card-name">{source.document_name}</span>
+              {source.relevance_score != null && (
+                <div className="source-relevance-track">
+                  <div
+                    className="source-relevance-bar"
+                    style={{
+                      width: `${source.relevance_score}%`,
+                      backgroundColor: relevanceColor(source.relevance_score),
+                    }}
+                  />
+                </div>
+              )}
+              <p className="source-card-preview">{source.preview}</p>
+            </div>
+          );
+        })}
       </div>
     </aside>
   );

apps/desktop/src/components/layout/layout.css

@@ -441,6 +441,16 @@
   box-shadow: var(--shadow-md);
 }
 
+.source-card-notebook {
+  display: block;
+  font-size: 10px;
+  font-weight: 500;
+  color: var(--color-accent);
+  text-transform: uppercase;
+  letter-spacing: 0.05em;
+  margin-bottom: 2px;
+}
+
 .source-card-name {
   font-size: var(--text-sm);
   font-weight: 550;

apps/desktop/src/hooks/useChat.ts

@@ -30,10 +30,13 @@ export function useChat() {
       store.setIsStreaming(true);
       store.setActiveSources([]);
 
+      // Cross-notebook mode: send all notebook IDs instead of just one
+      const crossMode = store.crossNotebookMode;
       const body = {
         prompt,
         history,
-        notebook_id: store.activeNotebookId,
+        notebook_id: crossMode ? null : store.activeNotebookId,
+        notebook_ids: crossMode ? store.notebooks.map((nb) => nb.notebook_id) : null,
         conversation_id: store.activeConversationId,
       };
 

apps/desktop/src/store/app-store.ts

@@ -19,6 +19,9 @@ interface AppState {
   conversations: Conversation[];
   activeConversationId: string | null;
 
+  // Cross-notebook
+  crossNotebookMode: boolean;
+
   // Documents
   documents: DocumentInfo[];
 
@@ -45,6 +48,9 @@ interface AppState {
   setConversations: (convs: Conversation[]) => void;
   setActiveConversationId: (id: string | null) => void;
 
+  // Actions — cross-notebook
+  setCrossNotebookMode: (val: boolean) => void;
+
   // Actions — documents
   setDocuments: (docs: DocumentInfo[]) => void;
 
@@ -64,6 +70,7 @@ export const useAppStore = create<AppState>((set) => ({
   isStreaming: false,
   conversations: [],
   activeConversationId: null,
+  crossNotebookMode: false,
   documents: [],
   sourcePanelOpen: true,
   activeSources: [],
@@ -95,6 +102,9 @@ export const useAppStore = create<AppState>((set) => ({
   setConversations: (convs) => set({ conversations: convs }),
   setActiveConversationId: (id) => set({ activeConversationId: id }),
 
+  // Cross-notebook
+  setCrossNotebookMode: (val) => set({ crossNotebookMode: val }),
+
   // Documents
   setDocuments: (docs) => set({ documents: docs }),
 

apps/desktop/src/types.ts

@@ -7,6 +7,7 @@ export interface ChatRequest {
   prompt: string;
   history?: ChatMessage[];
   notebook_id?: string | null;
+  notebook_ids?: string[] | null;
   conversation_id?: string | null;
 }
 

backend/notebooklm_backend/models/chat.py

@@ -14,6 +14,7 @@ class ChatRequest(BaseModel):
     prompt: str
     history: List[ChatMessage] | None = None
     notebook_id: str | None = Field(None, description="Optional notebook ID for RAG-enabled chat")
+    notebook_ids: List[str] | None = Field(None, description="Optional list of notebook IDs for cross-notebook synthesis")
     conversation_id: str | None = Field(None, description="Optional conversation ID for persistence")
 
 

backend/notebooklm_backend/routes/chat.py

@@ -75,6 +75,7 @@ async def event_generator():
                 prompt=payload.prompt,
                 history=history,
                 notebook_id=payload.notebook_id,
+                notebook_ids=payload.notebook_ids,
             ):
                 if event.get("type") == "meta":
                     sources_data = event.get("sources")

backend/notebooklm_backend/services/chat.py

@@ -89,6 +89,7 @@ async def stream_reply(
         prompt: str,
         history: Iterable[ChatMessage] | None = None,
         notebook_id: str | None = None,
+        notebook_ids: list[str] | None = None,
     ) -> AsyncIterator[dict[str, Any]]:
         """
         Stream tokens using SSE-friendly event payloads:
@@ -101,7 +102,42 @@ async def stream_reply(
         sources: list[dict[str, Any]] = []
         rag_context: RAGContext | None = None
 
-        if notebook_id and self._rag_service:
+        # Cross-notebook synthesis mode
+        if notebook_ids and len(notebook_ids) > 1 and self._rag_service:
+            try:
+                full_question = prompt
+                if history:
+                    recent_context = "\n".join([f"{msg.role}: {msg.content}" for msg in list(history)[-3:]])
+                    full_question = f"Previous conversation:\n{recent_context}\n\nCurrent question: {prompt}"
+                context = await self._rag_service.prepare_prompt_cross_notebook(
+                    notebook_ids=notebook_ids,
+                    question=full_question,
+                    top_k=20,
+                )
+                metrics.update(context.metrics)
+                prompt_text = context.prompt or None
+                sources = self._format_sources(context.sources)
+                rag_context = context
+                if not prompt_text:
+                    yield {
+                        "type": "meta",
+                        "provider": self.provider,
+                        "sources": sources,
+                        "metrics": metrics or None,
+                    }
+                    yield {
+                        "type": "done",
+                        "reply": "No relevant documents found across the selected notebooks.",
+                        "metrics": metrics or None,
+                    }
+                    return
+            except Exception as e:
+                import logging
+                logger = logging.getLogger(__name__)
+                logger.error(f"Cross-notebook RAG failed: {e}", exc_info=True)
+                prompt_text = None
+                metrics = {}
+        elif notebook_id and self._rag_service:
             try:
                 full_question = prompt
                 if history:
@@ -202,14 +238,15 @@ def _format_sources(self, sources: List[SourceAttribution]) -> list[dict[str, An
             if src.distance is not None:
                 # L2 distance normalization: 0 -> 100%, unbounded -> approaches 0%
                 relevance = round(100 / (1 + src.distance))
-            formatted.append(
-                {
-                    "source_path": src.source_path,
-                    "preview": preview,
-                    "distance": src.distance,
-                    "relevance_score": relevance,
-                }
-            )
+            entry: dict[str, Any] = {
+                "source_path": src.source_path,
+                "preview": preview,
+                "distance": src.distance,
+                "relevance_score": relevance,
+            }
+            if src.notebook_id:
+                entry["notebook_id"] = src.notebook_id
+            formatted.append(entry)
         return formatted
 
     def _record_metrics(

backend/notebooklm_backend/services/rag.py

@@ -15,6 +15,7 @@ class SourceAttribution:
     source_path: str
     content: str
     distance: float | None
+    notebook_id: str | None = None
 
 
 @dataclass
@@ -201,3 +202,82 @@ async def query(self, notebook_id: str, question: str, top_k: int = 5) -> RAGRes
             metrics["total_ms"] = metrics.get("total_ms", 0.0) + metrics["llm_ms"]
 
         return RAGResponse(answer=answer, sources=context.sources, metrics=metrics)
+
+    async def prepare_prompt_cross_notebook(
+        self,
+        notebook_ids: list[str],
+        question: str,
+        top_k: int = 20,
+    ) -> RAGContext:
+        """Prepare a prompt that retrieves and synthesizes across multiple notebooks."""
+        total_start = time.perf_counter()
+        metrics: dict[str, float] = {}
+
+        retrieval_start = time.perf_counter()
+        query_results = self.vector_store.query_across_notebooks(
+            notebook_ids=notebook_ids,
+            query=question,
+            top_k=top_k,
+        )
+        metrics["retrieval_ms"] = (time.perf_counter() - retrieval_start) * 1000
+
+        documents = query_results.get("documents", [[]])[0]
+        metadatas = query_results.get("metadatas", [[]])[0]
+        distances = query_results.get("distances", [[]])[0]
+
+        if not documents:
+            return RAGContext(prompt="", sources=[], metrics=metrics)
+
+        # Group chunks by notebook + source file
+        source_groups: dict[str, list[tuple[int, str, str]]] = {}
+        for idx, (doc, metadata) in enumerate(zip(documents, metadatas)):
+            nb_id = metadata.get("notebook_id", "unknown") if isinstance(metadata, dict) else "unknown"
+            source_path = metadata.get("source_path", "unknown") if isinstance(metadata, dict) else "unknown"
+            key = f"{nb_id}::{source_path}"
+            if key not in source_groups:
+                source_groups[key] = []
+            source_groups[key].append((idx, doc, nb_id))
+
+        # Build prompt with notebook labels
+        prompt_parts = []
+        for key, chunks in source_groups.items():
+            nb_id, source_path = key.split("::", 1)
+            source_name = Path(source_path).name if source_path != "unknown" else "Document"
+            prompt_parts.append(f"From notebook '{nb_id}', document '{source_name}':")
+            for idx, doc, _ in chunks:
+                prompt_parts.append(f"  [Source {idx+1}]: {doc}")
+            prompt_parts.append("")
+
+        prompt_context = "\n".join(prompt_parts)
+        notebook_count = len(set(
+            m.get("notebook_id", "?") for m in metadatas if isinstance(m, dict)
+        ))
+
+        prompt = (
+            "You are answering a question by synthesizing information ACROSS MULTIPLE NOTEBOOKS.\n"
+            f"You have access to {notebook_count} notebooks with excerpts from various documents.\n"
+            "Rules:\n"
+            "- Compare and contrast information from different notebooks and documents.\n"
+            "- When sources from different notebooks disagree, name the disagreement explicitly.\n"
+            "- When citing, reference which notebook and document the information came from.\n"
+            "- If the answer is not present in the excerpts, "
+            "reply: 'I could not find this across the provided notebooks.'\n\n"
+            f"Excerpts grouped by notebook and document:\n{prompt_context}\n"
+            f"Question: {question}\n\n"
+            "Answer (synthesize across all sources):"
+        )
+
+        sources = [
+            SourceAttribution(
+                source_path=metadata.get("source_path", "unknown") if isinstance(metadata, dict) else "unknown",
+                content=document,
+                distance=distances[idx] if idx < len(distances) else None,
+                notebook_id=metadata.get("notebook_id") if isinstance(metadata, dict) else None,
+            )
+            for idx, (document, metadata) in enumerate(zip(documents, metadatas))
+        ]
+
+        metrics["prep_ms"] = (time.perf_counter() - total_start) * 1000
+        metrics["total_ms"] = metrics["prep_ms"]
+        metrics["notebooks_queried"] = float(len(notebook_ids))
+        return RAGContext(prompt=prompt, sources=sources, metrics=metrics)