larkup 0.1.41 → 0.1.42

package/CHANGELOG.md

@@ -1,5 +1,17 @@
 # larkup
 
+## 0.1.42
+
+### Patch Changes
+
+- Improve smart RAG and web fallback, PDF visual retrieval, and resilient video indexing.
+- Updated dependencies
+  - @larkup/core@0.2.6
+  - @larkup/marketplace@0.1.12
+  - @larkup/scraper@0.1.30
+  - @larkup/tool-doc-editor@0.2.7
+  - @larkup/tool-video-audio@0.2.4
+
 ## 0.1.41
 
 ### Patch Changes

package/app/api/chat/route.ts

@@ -16,6 +16,11 @@ import type { CustomModelConfig } from '@larkup/core/types';
 import { DEFAULT_SYSTEM_PROMPT } from '@larkup/core/types';
 
 import { getChatTools } from './tools';
+import {
+  requiresCurrentWebSearch,
+  requiresKnowledgeBaseSearch,
+  retrievalToolsForStep,
+} from '@/lib/retrieval-routing';
 
 export const maxDuration = 60;
 
@@ -66,7 +71,7 @@ const SMART_RETRIEVAL_POLICY = `
 
 TOOL RULES:
 1. searchKnowledgeBase → Use FIRST when the question is about the user's indexed data (documents, files, images, videos, diagrams, anything they uploaded).
-2. webSearch → Use ONLY after searchKnowledgeBase returned nothing relevant AND the question needs current/public facts.
+2. webSearch → Use for clearly current/public facts. If a prior local search was not relevant or failed, use it as the one permitted fallback. If web search fails, use the knowledge base as the one permitted fallback.
 3. Other tools → Use when the question specifically requires them.
 4. General knowledge → Answer directly for common knowledge unrelated to the user's data.
 5. If nothing helps, say so honestly.
@@ -75,6 +80,8 @@ DO:
 - Keep answers short and direct. 1–3 sentences unless the task is complex.
 - Search at most ONCE per tool per turn. Never repeat the same search.
 - For questions about the user's own facts or preferences (for example, "what is my favourite fruit?"), searchKnowledgeBase is required before answering.
+- If a knowledge-base result contains PDF images and the question asks what is shown, named, counted, or connected in that visual, call analyzeImageDeeply before answering. Call presentMedia when the user asks to see the image.
+- For an indexed video outcome question, inspect endingContext before saying that a winner or final score is unavailable. Use only explicitly supported evidence.
 - Use presentMedia to show images/videos/audio from indexed content.
 - For follow-ups like "show me that", reuse the mediaAssetId from earlier results — do not search again.
 
@@ -103,37 +110,6 @@ function latestUserText(messages: UIMessage[]): string {
   return '';
 }
 
-/** Only force KB search when the user explicitly refers to their own data.
- * Matches "my document", "our files", "the uploaded image" etc. but NOT
- * generic uses like "I want to know", "can you tell me", "I think". */
-function requiresKnowledgeBaseSearch(text: string): boolean {
-  // Skip generic conversational patterns
-  if (/^(hi|hello|hey|thanks|thank you|ok|sure|yes|no|please|help)\b/i.test(text.trim())) {
-    return false;
-  }
-  // Match possessive references to user data
-  return (
-    /\b(my|our)\s+(favo(?:u)?rite|preference|choice|answer|name|result|score|winner)\b/i.test(
-      text,
-    ) ||
-    /\b(?:what|which|who|where|when|how)\b[\s\S]{0,80}\b(my|our)\b/i.test(text) ||
-    /\b(my|our)\s+(document|file|data|image|picture|diagram|video|audio|upload|corpus|knowledge|database|db|pdf|report|spreadsheet|presentation)/i.test(
-      text,
-    ) ||
-    /\b(the|this|that)\s+(document|file|diagram|image|picture|upload|pdf|report)/i.test(text) ||
-    /\b(uploaded?|indexed|scraped|knowledge base|corpus)\b/i.test(text) ||
-    /\bshow\s+me\b/i.test(text)
-  );
-}
-
-/** Questions whose answer is inherently time-sensitive should not rely on the
- * model's training snapshot when Web Search is available. */
-function requiresCurrentWebSearch(text: string): boolean {
-  return /\b(who won|winner|final score|match result|latest|current|today|yesterday|breaking news|weather|stock price|exchange rate|election result|schedule)\b/i.test(
-    text,
-  );
-}
-
 export async function POST(req: Request) {
   const {
     messages,
@@ -401,7 +377,12 @@ ${fieldLines}`;
       : '') +
     tabularContext +
     docContext;
-  const tools = await getChatTools({ serverId, docSessionId, config });
+  const tools = await getChatTools({
+    serverId,
+    docSessionId,
+    config,
+    origin: new URL(req.url).origin,
+  });
   const userText = latestUserText(messagesToProcess);
   const forceKnowledgeBaseSearch =
     requiresKnowledgeBaseSearch(userText) && Boolean(tools.searchKnowledgeBase);
@@ -420,38 +401,16 @@ ${fieldLines}`;
     system: systemPrompt,
     messages: await convertToModelMessages(safeMessages, { tools }),
     maxOutputTokens: 4096,
-    // Three steps: retrieve → optionally present media → answer.
+    // Three steps: retrieve → optionally inspect/present media → answer.
     stopWhen: stepCountIs(3),
     toolChoice: 'auto',
-    prepareStep: ({ stepNumber }) => {
-      // Step 0: force KB search if needed
-      if (forceKnowledgeBaseSearch && stepNumber === 0) {
-        return { toolChoice: { type: 'tool', toolName: 'searchKnowledgeBase' } };
-      }
-      // For questions about the user's data, let the model evaluate the KB
-      // result before falling back to the web. Forcing both tools used to make
-      // enabled web search override a perfectly relevant local answer.
-      if (forceWebSearch && !forceKnowledgeBaseSearch && stepNumber === 0) {
-        return { toolChoice: { type: 'tool', toolName: 'webSearch' } };
-      }
-      // After a forced tool, and after the optional KB → web fallback, remove
-      // search tools. This guarantees one concise search surface rather than
-      // a chain of near-identical web queries.
-      if (stepNumber >= 2) {
-        return {
-          activeTools: Object.keys(tools).filter(
-            (name) => name !== 'webSearch' && name !== 'searchKnowledgeBase',
-          ),
-        };
-      }
-      if (forceKnowledgeBaseSearch && stepNumber >= 1) {
-        return { activeTools: Object.keys(tools).filter((name) => name !== 'searchKnowledgeBase') };
-      }
-      if (forceWebSearch && stepNumber >= 1) {
-        return { activeTools: Object.keys(tools).filter((name) => name !== 'webSearch') };
-      }
-      return undefined;
-    },
+    prepareStep: ({ stepNumber }) =>
+      retrievalToolsForStep({
+        stepNumber,
+        forceKnowledgeBaseSearch,
+        forceWebSearch: forceWebSearch && !forceKnowledgeBaseSearch,
+        toolNames: Object.keys(tools),
+      }),
     onFinish: async ({ usage, response }) => {
       const { trackUsageEvent, estimateCost } = await import('@larkup/core/analytics-store');
       const u = usage as any;

package/app/api/chat/tools.ts

@@ -166,15 +166,21 @@ async function formatKnowledgeHits(query: string, rawHits: any[], topK: number)
   const trailingDocuments = timestampedDocuments.filter(
     (document) => Number(document.metadata?.endSecs ?? 0) >= mediaDuration - trailingHorizonSecs,
   );
+  // Result announcements can occur shortly before the literal final minutes
+  // (for example during a trophy or score recap). Preserve the latest explicit
+  // outcome evidence from the whole indexed timeline, not only the trailing
+  // window returned by a semantic match.
+  const timelineOutcomeDocuments = isOutcomeQuery(query)
+    ? timestampedDocuments.filter((document) => isOutcomeEvidence(document.content)).slice(-3)
+    : [];
   const signaledEndingDocuments = isOutcomeQuery(query)
     ? trailingDocuments.filter((document) => isOutcomeEvidence(document.content)).slice(-2)
     : [];
   const timestampedEnding = [
     ...new Map(
-      [...signaledEndingDocuments, ...trailingDocuments.slice(-4)].map((document) => [
-        document.id,
-        document,
-      ]),
+      [...timelineOutcomeDocuments, ...signaledEndingDocuments, ...trailingDocuments.slice(-4)].map(
+        (document) => [document.id, document],
+      ),
     ).values(),
   ].sort(
     (left, right) => Number(left.metadata?.endSecs ?? 0) - Number(right.metadata?.endSecs ?? 0),
@@ -287,7 +293,18 @@ async function formatKnowledgeHits(query: string, rawHits: any[], topK: number)
         // document, rather than creating media-library assets. Expose those
         // URLs explicitly so the presentation tool can validate and render
         // them in chat.
-        images: hit.metadata?.images,
+        images:
+          hit.metadata?.images ??
+          (hit.metadata?.isImage && hit.metadata?.imageUrl
+            ? [
+                {
+                  imageUrl: hit.metadata.imageUrl,
+                  pageNumber: hit.metadata.pageNumber,
+                  index: hit.metadata.index,
+                  description: hit.metadata.description,
+                },
+              ]
+            : undefined),
         metadata: hit.metadata,
         timelineContext,
         endingContext: hitIndex === 0 ? endingContext : undefined,
@@ -312,13 +329,15 @@ export async function getChatTools(context: {
   serverId?: string;
   docSessionId?: string;
   config?: any;
+  /** Public origin of the incoming chat request, used for local media URLs. */
+  origin?: string;
 }) {
   const { serverId, docSessionId, config } = context;
 
   const builtInTools: Record<string, any> = {
     searchKnowledgeBase: tool({
       description:
-        "HIGHEST PRIORITY TOOL — Search the user's private RAG knowledge base. ALWAYS use this FIRST before webSearch or any other tool when the question may relate to the user's indexed data. This includes questions about documents, files, images, diagrams, videos, or anything the user may have uploaded. Do not use for general knowledge, casual conversation, or writing tasks. Search once with a focused query. For media, inspect the returned timeline and endingContext, and reuse exact mediaAssetId/startSecs/endSecs values with presentMedia when a preview is useful. Never guess timestamps.",
+        "HIGHEST PRIORITY TOOL — Search the user's private RAG knowledge base. ALWAYS use this FIRST before webSearch or any other tool when the question may relate to the user's indexed data. This includes questions about documents, files, images, diagrams, videos, or anything the user may have uploaded. Do not use for general knowledge, casual conversation, or writing tasks. Search once with a focused query. If it returns no relevant evidence or an error and the question asks for public/current information, webSearch is available once as a fallback. For media, inspect the returned timeline and endingContext, and reuse exact mediaAssetId/startSecs/endSecs values with presentMedia when a preview is useful. Never guess timestamps.",
       inputSchema: z.object({
         query: z.string().describe('The search query for the knowledge base.'),
       }),
@@ -620,7 +639,10 @@ export async function getChatTools(context: {
           // Fetch the image to base64
           // Handle both absolute URLs and relative local uploads
           const fetchUrl = imageUrl.startsWith('/')
-            ? `http://127.0.0.1:${process.env.PORT || 3000}${imageUrl}`
+            ? new URL(
+                imageUrl,
+                context.origin ?? `http://127.0.0.1:${process.env.PORT || 3000}`,
+              ).toString()
             : imageUrl;
           const res = await fetch(fetchUrl);
           if (!res.ok) throw new Error(`Failed to fetch image: ${res.statusText}`);
@@ -631,7 +653,10 @@ export async function getChatTools(context: {
 
           // Post to our describe-image endpoint which handles model setup
           const descRes = await fetch(
-            `http://127.0.0.1:${process.env.PORT || 3000}/api/describe-image`,
+            new URL(
+              '/api/describe-image',
+              context.origin ?? `http://127.0.0.1:${process.env.PORT || 3000}`,
+            ).toString(),
             {
               method: 'POST',
               headers: { 'Content-Type': 'application/json' },
@@ -936,11 +961,35 @@ export async function getChatTools(context: {
 
     webSearch: tool({
       description:
-        "Search the public internet for current information, news, or facts. IMPORTANT: Only use this AFTER searchKnowledgeBase has been tried first and returned no relevant results. Never use webSearch as a first resort — the user's indexed knowledge base always takes priority. Use webSearch for questions about current events, public facts, or topics clearly outside the user's indexed data.",
+        'Search the public internet for current information, news, or facts. Use it first for clearly current/public requests, or once after an irrelevant local search. If it fails or returns no useful results, searchKnowledgeBase is available once as a fallback. Never repeat either search tool in the same turn.',
       inputSchema: z.object({
         query: z.string().describe('The search query.'),
       }),
       execute: async ({ query }) => {
+        const knowledgeBaseFallback = async (error: string) => {
+          try {
+            const knowledge = await queryKnowledgeBase(query, 5, serverId ?? null);
+            return {
+              error,
+              results: [],
+              fallback: {
+                source: 'knowledge-base',
+                attemptedBecause: 'Public web search failed or returned no results.',
+                ...knowledge,
+              },
+            };
+          } catch (fallbackError: any) {
+            return {
+              error,
+              results: [],
+              fallback: {
+                source: 'knowledge-base',
+                error: fallbackError?.message || 'Knowledge-base fallback failed.',
+                hits: [],
+              },
+            };
+          }
+        };
         const provider = config?.webSearchProvider || 'tavily';
         const req = new Request('http://localhost/api/search', {
           method: 'POST',
@@ -980,11 +1029,14 @@ export async function getChatTools(context: {
 
           if (res.status >= 400) {
             const err = await res.json();
-            return { error: err.error || 'Search failed', results: [] };
+            return knowledgeBaseFallback(err.error || 'Search failed');
           }
 
           const data = await res.json();
           const items = data.items || data.results || [];
+          if (items.length === 0) {
+            return knowledgeBaseFallback('Public search returned no results.');
+          }
           return {
             results: items.slice(0, 10).map((i: any) => ({
               title: i.title,
@@ -993,7 +1045,7 @@ export async function getChatTools(context: {
             })),
           };
         } catch (e: any) {
-          return { error: e.message || 'Search failed', results: [] };
+          return knowledgeBaseFallback(e.message || 'Search failed');
         }
       },
     }),
@@ -1016,6 +1068,9 @@ export async function getChatTools(context: {
         // Keeping it available prevents an enabled web-search configuration
         // from silently removing the user's local knowledge base.
         'searchKnowledgeBase',
+        // PDF/image hits can require one visual pass after retrieval. This is
+        // available only to a follow-up tool step, never auto-run on chat.
+        'analyzeImageDeeply',
         'presentMedia',
         'fillDocumentForm',
         'editDocument',

package/app/api/media/process/route.ts

@@ -646,7 +646,9 @@ async function processMediaWithTool(
       const scenePromise = (async () => {
         const extractedFrames = result.frames as { path: string; timestampSecs: number }[];
         const analysisWindowSecs = visualAnalysisWindow(result.meta.durationSecs);
-        const frameGroups = groupFramesByWindow(extractedFrames, analysisWindowSecs, 6);
+        // Chronological batches preserve visual context while keeping a
+        // long recording within a practical vision-request budget.
+        const frameGroups = groupFramesByWindow(extractedFrames, analysisWindowSecs, 12);
         if (frameGroups.length === 0) {
           await reportStage('vision', {
             status: 'skipped',
@@ -662,10 +664,37 @@ async function processMediaWithTool(
           message: `Understanding 0 of ${frameGroups.length} visual sequences...`,
         });
         let analyzed = 0;
+        let skipped = 0;
+        let visionThrottled = false;
+        let lastVisionRequestAt = 0;
+        const visionProvider = globalConfig.chatProvider || globalConfig.embeddingProvider;
+        // Providers commonly rate-limit visual generation much more tightly
+        // than text generation. Serializing and pacing Google requests avoids
+        // turning a temporary 429 into a failed video index.
+        const visionConcurrency = visionProvider === 'google' ? 1 : limits.apiConcurrency;
+        const visionMinIntervalMs = visionProvider === 'google' ? 4_250 : 0;
+        const recordProgress = () => {
+          analyzed++;
+          const progressStep = Math.max(1, Math.ceil(frameGroups.length / 100));
+          if (analyzed % progressStep === 0 || analyzed === frameGroups.length) {
+            void reportStage('vision', {
+              status: 'running',
+              current: analyzed,
+              total: frameGroups.length,
+              unit: 'sequences',
+              message: `Understanding visual sequence ${analyzed} of ${frameGroups.length}...`,
+            }).catch(() => {});
+          }
+        };
         const descriptions = await mapConcurrent(
           frameGroups,
-          limits.apiConcurrency,
+          visionConcurrency,
           async (frames: { path: string; timestampSecs: number }[]) => {
+            if (visionThrottled) {
+              skipped++;
+              recordProgress();
+              return null;
+            }
             const base64Images = await Promise.all(
               frames.map(async (frame) => (await fs.readFile(frame.path)).toString('base64')),
             );
@@ -675,14 +704,20 @@ async function processMediaWithTool(
               Math.max(startSecs + 1, frames.at(-1)!.timestampSecs + frameIntervalSecs),
             );
             const timestamps = frames.map((frame) => formatTime(frame.timestampSecs)).join(', ');
-            const descRes = await fetchWithRetry(
-              scopedApiUrl('/api/describe-image', reqUrl, serverId),
-              {
-                method: 'POST',
-                headers: { 'Content-Type': 'application/json' },
-                body: JSON.stringify({
-                  base64Images,
-                  prompt: `These are chronological frames from "${asset.fileName}" at ${timestamps}.
+            if (visionMinIntervalMs > 0) {
+              const remaining = visionMinIntervalMs - (Date.now() - lastVisionRequestAt);
+              if (remaining > 0) await new Promise((resolve) => setTimeout(resolve, remaining));
+            }
+            lastVisionRequestAt = Date.now();
+            try {
+              const descRes = await fetchWithRetry(
+                scopedApiUrl('/api/describe-image', reqUrl, serverId),
+                {
+                  method: 'POST',
+                  headers: { 'Content-Type': 'application/json' },
+                  body: JSON.stringify({
+                    base64Images,
+                    prompt: `These are chronological frames from "${asset.fileName}" at ${timestamps}.
 Write one evidence note for this time range. Track people, actions, interactions, state changes,
 important objects, and exact meaningful on-screen text. Perform careful OCR, including Arabic and
 other right-to-left text without reversing labels. For scoreboards or result animations, compare
@@ -690,38 +725,44 @@ frames in timestamp order and report the latest visible score as the current/fin
 each number with the correct nearby team/person label. Record a winner or result only when visually
 supported. Omit boilerplate about facts that are not shown. Be factual, information-dense, and
 concise.`,
-                }),
-              },
-            );
-            analyzed++;
-            const progressStep = Math.max(1, Math.ceil(frameGroups.length / 100));
-            if (analyzed % progressStep === 0 || analyzed === frameGroups.length) {
-              void reportStage('vision', {
-                status: 'running',
-                current: analyzed,
-                total: frameGroups.length,
-                unit: 'sequences',
-                message: `Understanding visual sequence ${analyzed} of ${frameGroups.length}...`,
-              }).catch(() => {});
-            }
-            if (!descRes.ok) {
-              let detail = '';
-              try {
-                const errorBody = (await descRes.json()) as { error?: string };
-                detail = errorBody.error?.trim() || '';
-              } catch {
-                // Keep the status-only fallback for non-JSON failures.
+                  }),
+                },
+              );
+              if (!descRes.ok) {
+                let detail = '';
+                try {
+                  const errorBody = (await descRes.json()) as { error?: string };
+                  detail = errorBody.error?.trim() || '';
+                } catch {
+                  // Keep the status-only fallback for non-JSON failures.
+                }
+                if (/quota|rate limit|resource_exhausted|too many requests/i.test(detail)) {
+                  visionThrottled = true;
+                }
+                skipped++;
+                console.warn(
+                  `Visual sequence ${formatTime(startSecs)}–${formatTime(endSecs)} was skipped: ${
+                    detail || `HTTP ${descRes.status}`
+                  }`,
+                );
+                return null;
               }
-              throw new Error(
-                `Visual sequence analysis failed (${descRes.status})${
-                  detail ? `: ${detail}` : '.'
-                }`,
+              const { description } = (await descRes.json()) as { description?: string };
+              return description
+                ? { text: description, startSecs, endSecs: Math.max(startSecs + 1, endSecs) }
+                : null;
+            } catch (error) {
+              skipped++;
+              console.warn(
+                `Visual sequence ${formatTime(startSecs)}–${formatTime(
+                  endSecs,
+                )} failed; preserving the remaining transcript evidence.`,
+                error,
               );
+              return null;
+            } finally {
+              recordProgress();
             }
-            const { description } = (await descRes.json()) as { description?: string };
-            return description
-              ? { text: description, startSecs, endSecs: Math.max(startSecs + 1, endSecs) }
-              : null;
           },
         );
         await reportStage('vision', {
@@ -729,7 +770,12 @@ concise.`,
           current: frameGroups.length,
           total: frameGroups.length,
           unit: 'sequences',
-          message: `Understood ${frameGroups.length} visual sequences with OCR.`,
+          message:
+            skipped > 0
+              ? `Understood ${
+                  frameGroups.length - skipped
+                } visual sequences; ${skipped} were skipped while the vision provider was unavailable. Speech evidence remains searchable.`
+              : `Understood ${frameGroups.length} visual sequences with OCR.`,
         });
         return descriptions;
       })();
@@ -1002,6 +1048,8 @@ concise.`,
 function visualAnalysisWindow(durationSecs: number): number {
   if (durationSecs >= 4 * 60 * 60) return 15 * 60;
   if (durationSecs > 60 * 60) return 5 * 60;
+  if (durationSecs > 30 * 60) return 3 * 60;
+  if (durationSecs > 10 * 60) return 2 * 60;
   return 60;
 }
 

package/components/chat/chat-workspace.tsx

@@ -107,7 +107,7 @@ function AttachmentCard({
       <div className="size-7 shrink-0 rounded-md overflow-hidden bg-background border border-border/90 flex items-center justify-center p-1">
         <img src={iconPath} alt={file.name} className="w-full h-full object-contain" />
       </div>
-      <div className="flex flex-col justify-center max-w-[120px]">
+      <div className="flex flex-col justify-center max-w-30">
         <span className="text-xs font-medium truncate text-foreground">{file.name}</span>
         <span className="text-[10px] text-muted-foreground">
           {(file.size / 1024).toFixed(0)} KB
@@ -684,7 +684,7 @@ function ChatWorkspaceInner() {
               <button
                 type="button"
                 onClick={() => setShowModelSelect((o) => !o)}
-                className="flex w-[300px] justify-between items-center gap-1.5 rounded-lg border border-border bg-white dark:bg-card px-3 py-1.5 text-xs text-foreground transition hover:bg-muted"
+                className="flex w-75 justify-between items-center gap-1.5 rounded-lg border border-border bg-white dark:bg-card px-3 py-1.5 text-xs text-foreground transition hover:bg-muted"
               >
                 <span className="flex items-center gap-2 truncate">
                   {selectedModelData && (
@@ -709,7 +709,7 @@ function ChatWorkspaceInner() {
               {showModelSelect && (
                 <>
                   <div className="fixed inset-0 z-40" onClick={() => setShowModelSelect(false)} />
-                  <div className="absolute right-0 top-full z-50 mt-1 w-[300px] rounded-lg border border-border bg-card ">
+                  <div className="absolute right-0 top-full z-50 mt-1 w-75 rounded-lg border border-border bg-card ">
                     {/* Search input */}
                     <div className="flex items-center gap-2 border-b border-border px-3 py-2">
                       <Search className="size-3.5 shrink-0 text-muted-foreground" />
@@ -815,7 +815,7 @@ function ChatWorkspaceInner() {
                           type="button"
                           className="flex h-8 w-8 items-center justify-center rounded-lg text-muted-foreground transition hover:bg-secondary hover:text-foreground"
                         >
-                          <History className="h-[16px] w-[16px]" />
+                          <History className="h-4 w-4" />
                         </button>
                       }
                     />
@@ -823,7 +823,7 @@ function ChatWorkspaceInner() {
                 />
                 <TooltipContent>Chat History</TooltipContent>
               </Tooltip>
-              <SheetContent side="left" className="w-[300px] sm:w-[350px] p-0 flex flex-col">
+              <SheetContent side="left" className="w-75 sm:w-87.5 p-0 flex flex-col">
                 <SheetHeader className="p-4 border-b">
                   <SheetTitle className="text-sm font-semibold">Chat History</SheetTitle>
                 </SheetHeader>
@@ -890,7 +890,7 @@ function ChatWorkspaceInner() {
                     onClick={newChat}
                     className="flex h-8 w-8 items-center justify-center rounded-lg text-muted-foreground transition hover:bg-secondary hover:text-foreground"
                   >
-                    <SquarePen className="h-[16px] w-[16px]" />
+                    <SquarePen className="h-4 w-4" />
                   </button>
                 }
               />
@@ -1016,7 +1016,7 @@ function ChatWorkspaceInner() {
                       type="button"
                       className="flex h-9 w-9 shrink-0 items-center justify-center rounded-xl text-muted-foreground transition hover:bg-secondary hover:text-foreground"
                     >
-                      <Plus className="h-[18px] w-[18px]" />
+                      <Plus className="h-4.5 w-4.5" />
                     </button>
                   }
                 />
@@ -1143,7 +1143,7 @@ function ChatWorkspaceInner() {
                   rows={1}
                   disabled={!ready}
                   placeholder="How can I help you…"
-                  className="max-h-48 min-h-[44px] w-full resize-none bg-transparent px-2 py-3 text-[15px] leading-relaxed text-foreground outline-none placeholder:text-muted-foreground [&::-webkit-scrollbar]:hidden disabled:opacity-50"
+                  className="max-h-48 min-h-11 w-full resize-none bg-transparent px-2 py-3 text-[15px] leading-relaxed text-foreground outline-none placeholder:text-muted-foreground [&::-webkit-scrollbar]:hidden disabled:opacity-50"
                 />
               </div>
 
@@ -1153,9 +1153,9 @@ function ChatWorkspaceInner() {
                 className="flex h-9 w-9 shrink-0 items-center justify-center rounded-xl bg-primary text-primary-foreground transition hover:opacity-90 disabled:opacity-40"
               >
                 {isBusy || isDocLoading ? (
-                  <Loader2 className="h-[18px] w-[18px] animate-spin" />
+                  <Loader2 className="h-4.5 w-4.5 animate-spin" />
                 ) : (
-                  <ArrowUp className="h-[18px] w-[18px]" />
+                  <ArrowUp className="h-4.5 w-4.5" />
                 )}
               </button>
             </div>

package/components/chat/message-item.tsx

@@ -538,13 +538,9 @@ export function MessageItem({
     const editor = useDocEditor();
     updateFromToolResult = editor.updateFromToolResult;
     openCanvas = editor.openCanvas;
-  } catch {
-    // Not wrapped in DocEditorProvider in some contexts
-  }
+  } catch {}
 
   const anyMessage = message as any;
-
-  // Stabilize the parts array to prevent re-renders from creating new array refs each time
   const parts: any[] = useMemo(() => {
     const p: any[] = message.parts ? [...message.parts] : [];
     if (anyMessage.toolInvocations && Array.isArray(anyMessage.toolInvocations)) {
@@ -563,11 +559,8 @@ export function MessageItem({
     return p;
   }, [message.parts, anyMessage.toolInvocations]);
 
-  // Track which tool call IDs have already been applied to prevent re-application
   const appliedToolCallsRef = useRef<Set<string>>(new Set());
 
-  // Effect to apply document edits to the Canvas context
-  // Uses a stable fingerprint to avoid re-triggering on every render
   const docToolFingerprint = useMemo(() => {
     if (!isLast) return '';
     return parts

package/components/data/media-panel.tsx

@@ -731,8 +731,7 @@ function MediaContent({
       if (!res.ok) throw new Error(data.error ?? 'Media import failed');
       const ids = (data.assets as MediaAsset[]).map((asset) => asset.id);
       toast.success(`${mediaType === 'video' ? 'Video' : 'Audio'} import started`, {
-        description:
-          'Queued for download and indexing. Live progress will appear once the source starts responding; you can safely leave in the meantime.',
+        description: 'Queued for download and indexing. You can safely leave in the meantime.',
         duration: 8_000,
       });
       await processAssets(ids);

package/components/data/upload-panel.tsx

@@ -219,7 +219,10 @@ export function UploadPanel({ onAdded }: { onAdded: () => void }) {
             format: 'plain',
             rawContent: text,
             fileObject: file,
-            indexImages: !text || text.trim().length === 0,
+            // PDF diagrams are often embedded alongside ordinary text. Keep
+            // visual indexing on by default, with the existing per-file toggle
+            // available for users who only want text extraction.
+            indexImages: true,
           });
         } else {
           const content = await file.text();
@@ -572,7 +575,7 @@ export function UploadPanel({ onAdded }: { onAdded: () => void }) {
               </button>
             </div>
           </div>
-          <ul className="space-y-1.5 max-h-[350px] overflow-y-auto pr-1">
+          <ul className="space-y-1.5 max-h-87.5 overflow-y-auto pr-1">
             {staged.map((f) => (
               <li
                 key={f.id}
@@ -782,7 +785,7 @@ export function UploadPanel({ onAdded }: { onAdded: () => void }) {
                       <p className="text-[11px] text-muted-foreground mb-2 -mt-1">
                         Selected columns will be combined to form the searchable content.
                       </p>
-                      <div className="max-h-[120px] overflow-y-auto space-y-2 pr-2">
+                      <div className="max-h-30 overflow-y-auto space-y-2 pr-2">
                         {editingFile.keys?.map((k) => {
                           const isSelected = editingFile.contentKeys?.includes(k);
                           return (
@@ -871,7 +874,7 @@ export function UploadPanel({ onAdded }: { onAdded: () => void }) {
                       <p className="text-[11px] text-muted-foreground mb-2 -mt-1">
                         These columns will be stored as searchable metadata.
                       </p>
-                      <div className="max-h-[120px] overflow-y-auto space-y-2 pr-2">
+                      <div className="max-h-30 overflow-y-auto space-y-2 pr-2">
                         {editingFile.keys?.map((k) => {
                           const isSelected = editingFile.metadataKeys?.includes(k);
                           const disabled = k === editingFile.titleKey;