I can't affort the full model lol. Use jsonrepair module to fix agent malformed JSON instead.

agent/agent.ts

@@ -11,18 +11,13 @@ import { loopEndConditional } from "./conditionals/loop_end";
 import { sort } from "./nodes/sort";
 import { triggerEventSetup } from "./nodes/triggerEventSetup";
 import { createEnsembleNode } from "./nodes/ensembleNode";
-import { selfEvalSetup } from "./nodes/selfEvalSetup";
 
 const triggerEventToolNode = createToolNode(triggerEventToolsByName);
-const peToolNode = createToolNode(triggerEventToolsByName);
 
 const normalisationModel = createModelNode([], "normalization.txt");
 const triggerEventModel = createModelNode(triggerEventToolsByName, "trigger.txt");
-const evaluationModel = createModelNode([], "eval.txt");
-const peModel = createModelNode(triggerEventToolsByName, "posteval.txt");
 
-const triggerEventToolConditional = createToolConditional("triggerEventToolNode", selfEvalSetup.name);
-const peToolConditional = createToolConditional("peToolNode", verificationSetup.name);
+const triggerEventToolConditional = createToolConditional("triggerEventToolNode", verificationSetup.name);
 
 const roNode = createEnsembleNode("ROBERTA", "roberta");
 const flNode = createEnsembleNode("FLAN", "flan");
@@ -38,12 +33,6 @@ const agent = new StateGraph(MessagesState)
   .addNode("triggerEventToolNode", triggerEventToolNode)
   .addNode("triggerEventModel", triggerEventModel)
 
-  .addNode(selfEvalSetup.name, selfEvalSetup)
-  .addNode("evaluationModel", evaluationModel)
-  
-  .addNode("peToolNode", peToolNode)
-  .addNode("peModel", peModel)
-
   .addNode(verificationSetup.name, verificationSetup)
 
   .addNode("roNode", roNode)
@@ -60,16 +49,9 @@ const agent = new StateGraph(MessagesState)
   .addEdge(triggerEventSetup.name, "triggerEventModel")
   
   // @ts-expect-error
-  .addConditionalEdges("triggerEventModel", triggerEventToolConditional, ["triggerEventToolNode", selfEvalSetup.name])
+  .addConditionalEdges("triggerEventModel", triggerEventToolConditional, ["triggerEventToolNode", verificationSetup.name])
   .addEdge("triggerEventToolNode", "triggerEventModel")
   
-  .addEdge(selfEvalSetup.name, "evaluationModel")
-  .addEdge("evaluationModel", "peModel")
-
-  // @ts-expect-error
-  .addConditionalEdges("peModel", peToolConditional, ["peToolNode", verificationSetup.name])
-  .addEdge("peToolNode", "peModel")
-  
   .addEdge(verificationSetup.name, "roNode")
   .addEdge(verificationSetup.name, "flNode")
   .addEdge(verificationSetup.name, "lrNode")

agent/nodes/model.ts

@@ -9,7 +9,7 @@ export function createModelNode(tools: any, promptPath: string): GraphNode<typeo
         const sysPrompt = await hydratePrompt(promptPath, state);
 
         const model = new ChatOpenAI({
-            model: "gpt-5-mini"
+            model: "gpt-5.4-mini"
         });
         const modelWithTools = model.bindTools(Object.values(tools));
 

agent/nodes/selfEvalSetup.ts

@@ -1,21 +0,0 @@
-import { GraphNode } from "@langchain/langgraph";
-import { MessagesState, ProposedTriggerEventArray } from "../state";
-import { logger } from "../utils/logger";
-import { queryScraper } from "../tools/webSearch";
-import { rankAndDisplayData } from "../tools/triggerEventTools";
-
-export const selfEvalSetup: GraphNode<typeof MessagesState> = async (state) => {
-  let genResponse = state.messages.at(-1)?.content.toString() ?? "";
-  const parsed = ProposedTriggerEventArray.parse(JSON.parse(genResponse));
-
-  for (let i = 0; i < parsed.length; i++) {
-    const search = parsed[i].SearchQuery
-    const data = await queryScraper(search);
-    const output = await rankAndDisplayData(data, search);
-
-    parsed[i].context = output;
-  }
-
-  return { evalTriggerEvent: parsed };
-
-};

agent/nodes/verificationSetup.ts

@@ -1,8 +1,7 @@
 import { GraphNode } from "@langchain/langgraph";
 import { MessagesState, ProposedTriggerEventArray } from "../state";
 import { logger } from "../utils/logger";
-import { queryScraper } from "../tools/webSearch";
-import { rankAndDisplayData } from "../tools/triggerEventTools";
+import { jsonrepair } from 'jsonrepair'
 
 export const verificationSetup: GraphNode<typeof MessagesState> = async (state) => {
   //this is kinda doing two things, but having two nodes for it seems overkill
@@ -11,8 +10,20 @@ export const verificationSetup: GraphNode<typeof MessagesState> = async (state)
     logger.warn("No trigger events in memory, parsing")
 
     let genResponse = state.messages.at(-1)?.content.toString() ?? "";
-    const parsed = ProposedTriggerEventArray.parse(JSON.parse(genResponse));
 
+    const repaired = jsonrepair(genResponse);
+
+    const parsed = ProposedTriggerEventArray.parse(JSON.parse(repaired));
+
+    for (let i = 0; i < parsed.length; i++) {
+      const search = parsed[i].SearchQuery
+      // const data = await queryScraper(search);
+      // const output = await rankAndDisplayData(data, search);
+
+      // parsed[i].context = output;
+      parsed[i].context = "NONE"
+    }
+    
     return { proposedTriggerEvent: parsed, proposedTriggerEventIndex: 0 };
   }
   else {

agent/package-lock.json

@@ -20,6 +20,7 @@
         "dotenv": "^17.2.3",
         "exponential-backoff": "^3.1.3",
         "fs": "^0.0.1-security",
+        "jsonrepair": "^3.13.3",
         "langchain": "^1.2.14",
         "selenium-webdriver": "^4.40.0",
         "tldts": "^7.0.23",
@@ -2075,6 +2076,15 @@
       "integrity": "sha512-ZClg6AaYvamvYEE82d3Iyd3vSSIjQ+odgjaTzRuO3s7toCdFKczob2i0zCh7JE8kWn17yvAWhUVxvqGwUalsRA==",
       "license": "ISC"
     },
+    "node_modules/jsonrepair": {
+      "version": "3.13.3",
+      "resolved": "https://registry.npmjs.org/jsonrepair/-/jsonrepair-3.13.3.tgz",
+      "integrity": "sha512-BTznj0owIt2CBAH/LTo7+1I5pMvl1e1033LRl/HUowlZmJOIhzC0zbX5bxMngLkfT4WnzPP26QnW5wMr2g9tsQ==",
+      "license": "ISC",
+      "bin": {
+        "jsonrepair": "bin/cli.js"
+      }
+    },
     "node_modules/jszip": {
       "version": "3.10.1",
       "resolved": "https://registry.npmjs.org/jszip/-/jszip-3.10.1.tgz",

agent/package.json

@@ -24,6 +24,7 @@
     "dotenv": "^17.2.3",
     "exponential-backoff": "^3.1.3",
     "fs": "^0.0.1-security",
+    "jsonrepair": "^3.13.3",
     "langchain": "^1.2.14",
     "selenium-webdriver": "^4.40.0",
     "tldts": "^7.0.23",

agent/prompts/eval.txt

@@ -1,30 +0,0 @@
-You are an impartial and meticulous evaluator assessing LLM’s response based on key quality dimensions
-of honesty and usefulness. Your goal is to provide structured feedback that can be used to improve the
-response.
-Evaluation task: please follow these steps carefully:
-1. Analyze the response based on the three dimensions below.
-2. Provide justifications first: write a brief explanation justifying your assessment for each dimension.
-3. Assign scores after justification: assign a score from 1 (poor) to 10 (excellent) for each dimension
-based on your justification.
-4. Synthesize: provide a brief overall impression and the single most important suggestion for improve-
-ment, keeping in mind that explanation/honesty is the top priority, then followed by guidance.
-Critique dimensions (evaluate in this order):
-(1) Speficicity and usefullness: Can the proposed event be used to create a dataset of concrete events mapped to later
-disinformation.
-(2) Existance: Using the context provided, can the user be certain that the proposed trigger event actually happened
-(3) Causality: Is there a possible link from the proposed trigger event to the disinformaiton at hand
-Overall impression & key improvement suggestion: Briefly summarize the overall quality and state the
-most critical change needed to improve the response.
-
-Disinformation query: 
-###NTITLE###
-Disinformation date:
-###CDATE###
-
-LLM’s response to evaluate:
-###LM###
-
-Provided context:
-###VESEARCHES###
-
-Let's think it through step by step

agent/prompts/hydratePrompt.ts

@@ -15,10 +15,6 @@ export async function hydratePrompt(path: string, state: any) : Promise<string>
         raw = raw.replace("###LM###", state.messages.at(-1).content);
     }
 
-    if (raw.indexOf("###L2M###") != -1) {
-        raw = raw.replace("###L2M###", state.messages.at(-2).content);
-    }
-
     if (raw.indexOf("###NTITLE###") != -1) {
         raw = raw.replace("###NTITLE###", state.normalizedClaim);
     }
@@ -37,12 +33,5 @@ export async function hydratePrompt(path: string, state: any) : Promise<string>
         raw = raw.replace("###TESEARCH###", output)
     }
 
-    if (raw.indexOf("###VESEARCHES###") != -1) {
-        const output = state.evalTriggerEvent
-            .map(e => e.context)
-            .join("\n")
-        raw = raw.replace("###VESEARCHES###", output)
-    }
-
     return raw;
 }

agent/prompts/posteval.txt

@@ -1,40 +0,0 @@
-You are an expert editor tasked with making targeted improvements to an existing LLM’s response based
-on a specific critique with the primary goal of enhancing its score according to evaluation standards while
-preserving its strengths.
-Your revision task: generate a revised version of the existing response. Your goal is not to rewrite it
-completely, but to make precise edits only to address the specific weaknesses highlighted in the critique.
-Instructions for editing:
-- Identify specific flaws: carefully read the critique and pinpoint the exact issues raised (e.g., unclear
-explanation, vagueness, inappropriate responses, the key suggestion).
-- Perform minimal targeted edits: modify only the necessary sentences or paragraphs within the existing
-response to directly fix these identified flaws.
-- Strongly preserve strengths: crucially keep all other parts of the existing response intact. Do not
-rephrase, restructure, or remove sections that were not criticized or likely contributed positively to its
-initial score.
-- Ensure coherence: verify that your targeted edits integrate smoothly and do not introduce contradictions
-or awkward phrasing.
-Output requirements:
-- It should feel like a slightly polished or corrected version of the existing response, not a fundamentally
-different answer.
-- Do not mention the critique, scores, or the editing process. The output should be clean json that passes validation checks
-
-Again, use a JSON format with each entry containing "Event,ReasoningWhyRelevant,SearchQuery,Url,Date".
-Use tools available to you if further information is required
-
-Add no new events, only improve the existing items
-
-Disinformation query: 
-###NTITLE###
-Disinformation date:
-###CDATE###
-
-LLM’s response to improve:
-###L2M###
-
-Citique:
-###LM###
-
-This contains specific feedback, justifications, scores from 1 to 10, and potentially a key improvement
-suggestion. Focus on the justifications for low scores and the key suggestion.
-
-Let's think it through step by step

agent/prompts/trigger.txt

@@ -26,4 +26,8 @@ Events will be reordered as part of processing, each statement must stand alone
 The preceeding messages act as examples of previous responses to potentially ficitonal events and scores given.
 Analysis should only be completed for proposed events that would graner >0.7 points
 
+First, consider a range of directions in which the proposed disinformation could have been influenced by.
+Then, research these directions in turn, using the tools at hand.
+Finally, refine your proposed "trigger event" until it is specific, quantifiable and backed up by evidence.
+
 Lets go through it step by step

agent/state.ts

@@ -21,7 +21,6 @@ export const MessagesState = new StateSchema({
   date: z.string(),
   messages: MessagesValue,
   proposedTriggerEvent: ProposedTriggerEventArray,
-  evalTriggerEvent: ProposedTriggerEventArray,
   proposedTriggerEventIndex: z.int(),
   normalizedClaim: z.string(),
 });

agent/tools/webpageFetch.ts

@@ -8,7 +8,7 @@ export async function extractWebpageContent(url: string): Promise<string[]> {
     const response = await backOff(async () => {
       return await extractWebpageContentWorker(url);
     }, {
-      numOfAttempts: 10,
+      numOfAttempts: 5,
       startingDelay: 500,
       timeMultiple: 2,
       jitter: "full",