fix: improve json parsing resilience and disable gemini thinking

- add jsonrepair fallback for truncated LLM output
- disable gemini built-in thinking mode (thinkingBudget: 0)
- increase token limits for errorAnalyzer, queryRewriter, serpCluster
- switch production default to gemini-2.5-flash-lite
- fix normalizeHostName to handle wildcard patterns

config.json

@@ -38,14 +38,19 @@
         },
         "evaluator": {
           "temperature": 0.6,
-          "maxTokens": 200
+          "maxTokens": 1000
+        },
+        "errorAnalyzer": {
+          "maxTokens": 4000
         },
-        "errorAnalyzer": {},
         "queryRewriter": {
-          "temperature": 0.1
+          "temperature": 0.1,
+          "maxTokens": 4000
         },
         "researchPlanner": {},
-        "serpCluster": {},
+        "serpCluster": {
+          "maxTokens": 4000
+        },
         "agent": {
           "temperature": 0.7
         },

jina-ai/config.json

@@ -34,31 +34,35 @@
   "models": {
     "gemini": {
       "default": {
-        "model": "gemini-2.5-flash",
+        "model": "gemini-2.5-flash-lite",
         "temperature": 0.6,
         "maxTokens": 8000
       },
       "tools": {
         "coder": {
-          "maxTokens": 2000,
+          "maxTokens": 2000
-          "model": "gemini-2.5-flash-lite"
         },
         "researchPlanner": {},
         "evaluator": {
           "maxTokens": 2000
         },
-        "serpCluster": {},
+        "serpCluster": {
+          "maxTokens": 4000
+        },
         "errorAnalyzer": {
-          "maxTokens": 1000
+          "maxTokens": 4000
         },
         "queryRewriter": {
-          "maxTokens": 2000
+          "maxTokens": 4000
+        },
+        "agent": {
+          "model": "gemini-2.5-flash"
+        },
+        "agentBeastMode": {
+          "model": "gemini-2.5-flash"
         },
-        "agent": {},
-        "agentBeastMode": {},
         "fallback": {
-          "maxTokens": 8000,
+          "maxTokens": 8000
-          "model": "gemini-2.5-flash-lite"
         },
         "finalizer": {},
         "reducer": {

package-lock.json

@@ -22,6 +22,7 @@
         "express-validator": "^7.2.1",
         "hjson": "^3.2.2",
         "jsdom": "^26.0.0",
+        "jsonrepair": "^3.13.1",
         "node-fetch": "^3.3.2",
         "sharp": "^0.34.2",
         "undici": "^7.3.0",
@@ -6230,6 +6231,15 @@
         "url": "https://github.com/chalk/chalk?sponsor=1"
       }
     },
+    "node_modules/jsonrepair": {
+      "version": "3.13.1",
+      "resolved": "https://registry.npmjs.org/jsonrepair/-/jsonrepair-3.13.1.tgz",
+      "integrity": "sha512-WJeiE0jGfxYmtLwBTEk8+y/mYcaleyLXWaqp5bJu0/ZTSeG0KQq/wWQ8pmnkKenEdN6pdnn6QtcoSUkbqDHWNw==",
+      "license": "ISC",
+      "bin": {
+        "jsonrepair": "bin/cli.js"
+      }
+    },
     "node_modules/jwa": {
       "version": "2.0.0",
       "resolved": "https://registry.npmjs.org/jwa/-/jwa-2.0.0.tgz",

package.json

@@ -40,6 +40,7 @@
     "express-validator": "^7.2.1",
     "hjson": "^3.2.2",
     "jsdom": "^26.0.0",
+    "jsonrepair": "^3.13.1",
     "node-fetch": "^3.3.2",
     "sharp": "^0.34.2",
     "undici": "^7.3.0",

src/utils/safe-generator.ts

@@ -8,9 +8,15 @@ import {
 } from "ai";
 import { TokenTracker } from "./token-tracker";
 import { getModel, ToolName, getToolConfig } from "../config";
-import Hjson from 'hjson'; // Import Hjson library
+import Hjson from 'hjson';
 import { logError, logDebug, logWarning } from '../logging';
 
+// Dynamic import for ESM module
+const getJsonRepair = async () => {
+  const { jsonrepair } = await import('jsonrepair');
+  return jsonrepair;
+};
+
 interface GenerateObjectResult<T> {
   object: T;
   usage: LanguageModelUsage;
@@ -154,6 +160,13 @@ export class ObjectGeneratorSafe {
         messages,
         maxTokens: getToolConfig(model).maxTokens,
         temperature: getToolConfig(model).temperature,
+        providerOptions: {
+          google: {
+            thinkingConfig: {
+              thinkingBudget: 0  // Disable Gemini's built-in thinking to avoid conflict with our schema's think field
+            }
+          }
+        }
       });
 
       this.tokenTracker.trackUsage(model, result.usage);
@@ -200,6 +213,13 @@ export class ObjectGeneratorSafe {
               schema: distilledSchema,
               prompt: `Following the given JSON schema, extract the field from below: \n\n ${failedOutput}`,
               temperature: getToolConfig('fallback').temperature,
+              providerOptions: {
+                google: {
+                  thinkingConfig: {
+                    thinkingBudget: 0
+                  }
+                }
+              }
             });
 
             this.tokenTracker.trackUsage('fallback', fallbackResult.usage); // Track against fallback model
@@ -224,29 +244,43 @@ export class ObjectGeneratorSafe {
   private async handleGenerateObjectError<T>(error: unknown): Promise<GenerateObjectResult<T>> {
     if (NoObjectGeneratedError.isInstance(error)) {
       logWarning('Object not generated according to schema, fallback to manual parsing', { error });
+      const rawText = (error as any).text;
+
+      // 1. First try standard JSON parsing
       try {
-        // First try standard JSON parsing
+        const partialResponse = JSON.parse(rawText);
-        const partialResponse = JSON.parse((error as any).text);
         logDebug('JSON parse success!');
         return {
           object: partialResponse as T,
           usage: (error as any).usage
         };
-      } catch (parseError) {
+      } catch (jsonError) {
-        // Use Hjson to parse the error response for more lenient parsing
+        // 2. Try jsonrepair to fix truncated/malformed JSON
         try {
-          const hjsonResponse = Hjson.parse((error as any).text);
+          const jsonrepair = await getJsonRepair();
+          const repairedJson = jsonrepair(rawText);
+          const repairedResponse = JSON.parse(repairedJson);
+          logDebug('jsonrepair parse success!');
+          return {
+            object: repairedResponse as T,
+            usage: (error as any).usage
+          };
+        } catch (repairError) {
+          // 3. Try Hjson for lenient parsing (trailing commas, comments, etc.)
+          try {
+            const hjsonResponse = Hjson.parse(rawText);
             logDebug('Hjson parse success!');
             return {
               object: hjsonResponse as T,
               usage: (error as any).usage
             };
           } catch (hjsonError) {
-          logError('Both JSON and Hjson parsing failed:', { error: hjsonError });
+            logError('All JSON parsing attempts failed (JSON, jsonrepair, Hjson):', { error: hjsonError });
             throw error;
           }
         }
       }
+    }
     throw error;
   }
 }

src/utils/url-tools.ts

@@ -194,6 +194,18 @@ const extractUrlParts = (urlStr: string) => {
 };
 
 export const normalizeHostName = (hostStr: string) => {
+  // Handle wildcard patterns like *.medium.com
+  if (hostStr.startsWith('*.')) {
+    hostStr = hostStr.slice(2);
+  }
+
+  // If it doesn't look like a URL (no protocol), just clean up the hostname directly
+  if (!hostStr.includes('://')) {
+    const cleaned = hostStr.startsWith('www.') ? hostStr.slice(4) : hostStr;
+    return cleaned.toLowerCase();
+  }
+
+  // Try to parse as URL
   const extract = extractUrlParts(hostStr);
   const host = extract.hostname;
   if (!host) {