Fix fact checker service failures and refresh Parallel APIs

typescript-recipes/parallel-fact-checker-cerebras/README.md

@@ -96,7 +96,7 @@ RULES:
     });
 
     // Stream facts as they're extracted
-    for await (const chunk of factsResult.textStream) {
+    for await (const chunk of modelText(factsResult)) {
       // Parse complete FACT: lines and emit them immediately
       // ...
     }
@@ -123,12 +123,14 @@ async function verifyFact(
 ): Promise<void> {
   try {
     // Search for evidence using Parallel
-    const searchResult = await parallel.beta.search({
+    const searchResult = await parallel.search({
       objective: `Find reliable sources to verify or refute this claim: "${fact.text}"`,
       search_queries: [fact.text],
-      processor: "base",
-      max_results: 5,
-      max_chars_per_result: 2000,
+      mode: "basic",
+      advanced_settings: {
+        max_results: 5,
+        excerpt_settings: { max_chars_per_result: 2000 },
+      },
     });
 
     // Have LLM analyze the evidence
@@ -148,7 +150,7 @@ EXPLANATION: [Brief 1-2 sentence explanation]`,
 Evidence from web search:
 ${JSON.stringify(searchResult.results?.slice(0, 3).map(r => ({
   title: r.title,
-  excerpt: r.excerpts?.slice(0, 500)
+  excerpt: r.excerpts?.join("\n").slice(0, 2000)
 })), null, 2)}`,
       maxOutputTokens: 500,
     });
@@ -176,11 +178,10 @@ await Promise.all(
 When a user provides a URL, we extract the content using Parallel's Extract API:
 
 ```typescript
-const extractResult = await parallel.beta.extract({
+const extractResult = await parallel.extract({
   urls: [extractUrl],
   objective: "Extract the article text and key claims from this webpage",
-  excerpts: true,
-  full_content: false,
+  advanced_settings: { full_content: false },
 });
 
 const rawContent = extractResult.results[0].full_content ||
@@ -214,6 +215,8 @@ sendSSE(controller, encoder, { type: "warning", message: "Content was truncated
 sendSSE(controller, encoder, { type: "error", error: "Demo is experiencing high traffic. Please try again later." });
 ```
 
+The `modelText` helper consumes the AI SDK's `fullStream` and throws on error events. Reading `textStream` alone silently filters those events out, which can make failed inference look like content with no claims. Only an actual empty response should be reported as having no verifiable claims.
+
 The frontend handles these events to update the UI in real-time, showing facts as they're extracted and verdicts as they arrive.
 
 ## Setup
@@ -297,6 +300,16 @@ Extract content from URL and fact-check it. Content is truncated to 5,000 charac
 | `error` | Error occurred |
 | `complete` | Processing finished |
 
+## Validation and operations
+
+Run `npm test` for regression tests and `npm run typecheck` for TypeScript validation. Test both pasted text and URL input with working keys before deployment.
+
+The public demo runs as the `cerebras-fact-checker` Cloudflare Worker. Worker invocation logs are enabled in `wrangler.jsonc`. In Cloudflare, use Workers & Pages to inspect this Worker's Metrics and Observability views. Request counts include page loads and API requests; filter by `/check` and `/extract` where possible. A streamed application error can still return HTTP 200, so HTTP success counts alone do not measure successful fact checks.
+
+There is no browser analytics or conversion-event instrumentation in this recipe. Cloudflare traffic and logs can help diagnose usage and failures, but they do not provide a visitor-to-successful-check funnel. Parallel org usage counts API calls, which can include several searches per fact-checking session.
+
+If the deployed demo fails while local tests work, inspect the deployed Worker's inference errors and the credentials in the correct environment. Do not copy a personal key into the public demo. Provision the demo's own service credentials, then test both input paths again.
+
 ## Demo Limitations
 
 This demo has several limitations: