feat: Exa.ai web search as supplementary context in the answer pipeline

app/api/sandbox/route.ts

@@ -4,6 +4,7 @@ import { codeBlock } from 'common-tags';
 import { textModel } from '@/lib/ai';
 import { URL_CONTEXT_FALLBACK_CONFIDENCE } from '@/lib/autopilot';
 import { generateEmbedding, serializeEmbedding } from '@/lib/embeddings';
+import { searchWeb } from '@/lib/exa-search';
 import { parseLLMJSON } from '@/lib/parse-llm-json';
 import { createServerClient } from '@/lib/supabase/server';
 
@@ -61,11 +62,13 @@ async function runGroundedAnswerAgent({
   subject,
   question,
   retrievedContext,
+  webContext,
   tonePolicy,
 }: {
   subject: string;
   question: string;
   retrievedContext: string;
+  webContext?: string;
   tonePolicy?: string | null;
 }) {
   const systemPrompt = codeBlock`
@@ -102,6 +105,8 @@ async function runGroundedAnswerAgent({
       Customer question:
       ${question}
 
+      ${webContext ? `Web search results:\n${webContext}\n` : ''}
+
       Retrieved knowledge base sections:
       ${retrievedContext}
 
@@ -147,6 +152,18 @@ export async function POST(request: Request) {
   const tonePolicy = org?.tone_policy ?? null;
   const questionText = `${subject}\n${question}`;
 
+  /* ---------------------- WEB SEARCH + VECTOR RETRIEVAL ------------------ */
+
+  const embeddingPromise =
+    datasources && datasources.length > 0
+      ? generateEmbedding(questionText)
+      : Promise.resolve([] as number[]);
+
+  const [webContext, embeddingResult] = await Promise.all([
+    searchWeb(questionText),
+    embeddingPromise,
+  ]);
+
   /* -------------------------- VECTOR RETRIEVAL --------------------------- */
 
   const matchedSections: Array<{
@@ -158,17 +175,14 @@ export async function POST(request: Request) {
     similarity: number;
   }> = [];
 
-  if (datasources && datasources.length > 0) {
-    const embedding = await generateEmbedding(questionText);
-    if (embedding.length > 0) {
-      const { data } = await supabase.rpc('match_sections', {
-        embedding: serializeEmbedding(embedding),
-        match_threshold: 0.1,
-        p_organization_id: orgId,
-        match_count: 8,
-      });
-      if (data) matchedSections.push(...data);
-    }
+  if (embeddingResult.length > 0) {
+    const { data } = await supabase.rpc('match_sections', {
+      embedding: serializeEmbedding(embeddingResult),
+      match_threshold: 0.1,
+      p_organization_id: orgId,
+      match_count: 8,
+    });
+    if (data) matchedSections.push(...data);
   }
 
   // Expand context by fetching immediate neighbors (position ± 1) of each
@@ -232,7 +246,7 @@ export async function POST(request: Request) {
     datasourceUrl: datasources?.find((d) => d.id === s.datasource_id)?.url,
   }));
 
-  if (!retrievedContext.trim()) {
+  if (!retrievedContext.trim() && !webContext.trim()) {
     return new Response(
       JSON.stringify({
         html: '<p>No relevant information found in the knowledge base for this question.</p>',
@@ -252,6 +266,7 @@ export async function POST(request: Request) {
     subject,
     question,
     retrievedContext,
+    webContext,
     tonePolicy,
   });
 

app/api/webhooks/reply/route.ts

@@ -7,6 +7,7 @@ import { textModel } from '@/lib/ai';
 import { URL_CONTEXT_FALLBACK_CONFIDENCE } from '@/lib/autopilot';
 import { cleanBody } from '@/lib/cleanBody';
 import { generateEmbedding, serializeEmbedding } from '@/lib/embeddings';
+import { searchWeb } from '@/lib/exa-search';
 import { parseLLMJSON } from '@/lib/parse-llm-json';
 import { createServiceClient } from '@/lib/supabase/service';
 
@@ -225,13 +226,15 @@ async function runGroundedAnswerAgent({
   question,
   retrievedContext,
   apiContext,
+  webContext,
   conversationHistory,
   tonePolicy,
 }: {
   subject: string;
   question: string;
   retrievedContext: string;
   apiContext?: string;
+  webContext?: string;
   conversationHistory?: string;
   tonePolicy?: string | null;
 }) {
@@ -273,6 +276,8 @@ async function runGroundedAnswerAgent({
 
       ${apiContext ? `Live API data:\n${apiContext}\n` : ''}
 
+      ${webContext ? `Web search results:\n${webContext}\n` : ''}
+
       Retrieved knowledge base sections:
       ${retrievedContext}
 
@@ -347,11 +352,14 @@ export async function POST(request: Request) {
 
   /* --------------------------- MCP TOOL GATHERING ------------------------- */
 
-  const apiContext = await gatherContextViaMcp(
-    thread?.subject ?? '',
-    record.cleaned_body,
-    mcpServers ?? []
-  );
+  const [apiContext, webContext] = await Promise.all([
+    gatherContextViaMcp(
+      thread?.subject ?? '',
+      record.cleaned_body,
+      mcpServers ?? []
+    ),
+    searchWeb(questionText),
+  ]);
 
   /* -------------------------- VECTOR RETRIEVAL --------------------------- */
 
@@ -376,6 +384,7 @@ export async function POST(request: Request) {
     question: record.cleaned_body,
     retrievedContext,
     apiContext,
+    webContext,
     conversationHistory,
     tonePolicy: org?.tone_policy,
   });

app/org/[slug]/dashboard/page.tsx

@@ -58,6 +58,7 @@ export default async function DashboardPage({ params }: Props) {
       initialTonePolicy={org.tone_policy ?? null}
       initialMcpServers={mcpServers ?? []}
       workflowsCount={workflows?.length ?? 0}
+      webSearchEnabled={!!process.env.EXA_API_KEY}
     />
   );
 }

app/org/[slug]/page.tsx

@@ -58,6 +58,7 @@ export default async function OrgPage({ params }: Props) {
       initialTonePolicy={org.tone_policy ?? null}
       initialMcpServers={mcpServers ?? []}
       workflowsCount={workflows?.length ?? 0}
+      webSearchEnabled={!!process.env.EXA_API_KEY}
     />
   );
 }

components/organization/WelcomeDashboard.tsx

@@ -54,6 +54,7 @@ interface Props {
   initialTonePolicy: string | null;
   initialMcpServers: Tables<'mcp_server'>[];
   workflowsCount: number;
+  webSearchEnabled: boolean;
 }
 
 // ─── Stat chip ────────────────────────────────────────────────────────────────
@@ -211,6 +212,7 @@ export function WelcomeDashboard({
   initialTonePolicy,
   initialMcpServers,
   workflowsCount,
+  webSearchEnabled,
 }: Props) {
   const { copied, copyToClipboard } = useCopyToClipboard();
   const [, setAddDataSource] = useAddDataSource();
@@ -318,6 +320,15 @@ export function WelcomeDashboard({
           <StatChip value={sources.length} label="Data sources" />
           <StatChip value={mcpServers.length} label="MCP servers" />
           <StatChip value={workflowsCount} label="Workflows" />
+          <div className="group relative flex items-center gap-3 border border-[#FF4500]/20 bg-muted/50 px-4 py-3 transition-all duration-200 hover:border-[#FF4500]/60 hover:bg-[#FF4500]/5 overflow-hidden">
+            <span className="absolute left-0 top-0 h-full w-0.5 bg-[#FF4500] scale-y-0 origin-bottom transition-transform duration-200 group-hover:scale-y-100" />
+            <span className={cn(
+              'font-mono text-xs uppercase tracking-[0.2em]',
+              webSearchEnabled ? 'text-emerald-500' : 'text-muted-foreground'
+            )}>
+              🌐 Web Search {webSearchEnabled ? 'On' : 'Off'}
+            </span>
+          </div>
         </div>
       </div>
 

env.example

@@ -4,3 +4,4 @@ SUPABASE_SERVICE_KEY=
 NEXT_PUBLIC_BASE_URL=http://localhost:3000
 RESEND_API_KEY=
 GOOGLE_GENERATIVE_AI_API_KEY=
+EXA_API_KEY=

lib/exa-search.ts

@@ -0,0 +1,48 @@
+import Exa from 'exa-js';
+
+/** Maximum number of Exa search results to include as context. */
+const EXA_MAX_RESULTS = 3;
+
+/** Maximum characters to include from each result's text. */
+const EXA_MAX_TEXT_CHARS = 1500;
+
+function formatSearchResult(
+  r: { title?: string | null; url?: string | null; text?: string | null },
+  index: number
+): string {
+  const title = r.title ? `[${index + 1}] ${r.title}` : `[${index + 1}]`;
+  const url = r.url ? `Source: ${r.url}` : '';
+  const text = r.text?.trim() ?? '';
+  return [title, url, text].filter(Boolean).join('\n');
+}
+
+/**
+ * Perform an Exa neural web search and return a plain-text summary of the
+ * top results suitable for inclusion in the grounded answer agent prompt.
+ *
+ * Returns an empty string when EXA_API_KEY is not set or the search fails.
+ */
+export async function searchWeb(query: string): Promise<string> {
+  const apiKey = process.env.EXA_API_KEY;
+  if (!apiKey) return '';
+
+  try {
+    const exa = new Exa(apiKey);
+
+    const result = await exa.searchAndContents(query, {
+      numResults: EXA_MAX_RESULTS,
+      type: 'neural',
+      useAutoprompt: true,
+      text: { maxCharacters: EXA_MAX_TEXT_CHARS },
+    });
+
+    if (!result.results?.length) return '';
+
+    return result.results
+      .map((r, i) => formatSearchResult(r, i))
+      .join('\n\n---\n\n');
+  } catch (err) {
+    console.error('[exa-search] Web search failed:', err);
+    return '';
+  }
+}

package.json

@@ -36,6 +36,7 @@
     "clsx": "^2.1.1",
     "common-tags": "^1.8.2",
     "date-fns": "^3.6.0",
+    "exa-js": "^2.7.0",
     "geist": "^1.7.0",
     "isomorphic-dompurify": "3.0.0-rc.2",
     "jotai": "^2.18.0",