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

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/triggerEventSetup.ts

@@ -7,11 +7,11 @@ export const triggerEventSetup: GraphNode<typeof MessagesState> = async (state)
   let nc = state?.messages?.at(-1)?.content ?? "" //keep a copy of normalized trigger event. Again two things, womp womp
   
   //Now give in-context examples. hopwfully we can self-teach?
-  // let similarityResults = await rankExampleTriggerEvents(state.disinformationTitle)
+  let similarityResults = await rankExampleTriggerEvents(state.disinformationTitle)
 
-  // let messages : BaseMessage[] = similarityResults.map((item) => {
+  let messages : BaseMessage[] = similarityResults.map((item) => {
-  //   return new AIMessage(`- Event: ${item.rawtext} \n\n - Claims and given scores: ${item.cleantext}`)
+    return new AIMessage(`- Event: ${item.rawtext} \n\n - Claims and given scores: ${item.cleantext}`)
-  // })
+  })
   
-  return { disinformationTitle: state.disinformationTitle, normalizedClaim: nc };
+  return { messages: messages, disinformationTitle: state.disinformationTitle, normalizedClaim: nc };
 };

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 { jsonrepair } from 'jsonrepair'
-import { rankAndDisplayData } from "../tools/triggerEventTools";
 
 export const verificationSetup: GraphNode<typeof MessagesState> = async (state) => {
   //this is kinda doing two things, but having two nodes for it seems overkill
@@ -11,7 +10,10 @@ 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

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/trigger.txt

@@ -14,7 +14,9 @@ Include a concise but specific search query that can be looked up on a search en
 
 Include a url to a source for your trigger event (not a web search, a specific url from a reputuable source). Do not use OAI cite, include url as text in response.
 
-Use a JSON format with each entry containing "Event,ReasoningWhyRelevant,SearchQuery,Url".
+Include the date that the event happened ("March 2022" for exmaple)
+
+Use a JSON format with each entry containing "Event,ReasoningWhyRelevant,SearchQuery,Url,Date".
 
 Multiple tool invocations should be requested at once, if applicable.
 Use your abilities to look between the lines and produce some insightful analysis, thinking both short and long term.
@@ -24,4 +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

@@ -9,6 +9,7 @@ export const ProposedTriggerEvent = z.object({
   ReasoningWhyRelevant: z.string(),
   SearchQuery: z.string(),
   Url: z.url(),
+  Date: z.string(),
   context: z.string().optional(),
   score: z.number().optional()
 })

agent/tools/retreiveExamples.ts

@@ -15,6 +15,8 @@ const CACHE_PATH = "../data/csv.cache.json";
 
 const JSONL_PATH = "../data/input.jsonl"
 
+const BM25_MIN_DOCS = 3;
+
 type EmbeddingCache = {
   rawtexts: string[];
   cleantexts: string[];
@@ -287,8 +289,20 @@ async function embedText(text: string): Promise<number[]> {
 }
 
 function buildBM25(texts: string[]) {
-  logger.info("Building BM25 index (%s docs)...", texts.length);
+  let paddedTexts = texts;
 
+  if (texts.length < BM25_MIN_DOCS) {
+    const needed = BM25_MIN_DOCS - texts.length;
+    logger.error(
+      "Corpus too small for BM25 (%s docs, need %s+), padding with %s dummy doc(s)",
+      texts.length,
+      BM25_MIN_DOCS,
+      needed
+    );
+    paddedTexts = [...texts, ...Array(needed).fill("placeholder dummy document")];
+  }
+
+  logger.info("Building BM25 index (%s docs)...", paddedTexts.length);
   const bm25 = bm25Factory();
 
   bm25.defineConfig({
@@ -302,7 +316,7 @@ function buildBM25(texts: string[]) {
     nlp.tokens.removeWords,
   ]);
 
-  texts.forEach((text, i) => {
+  paddedTexts.forEach((text, i) => {
     bm25.addDoc({ text }, i);
   });
 

agent/tools/webpageFetch.ts

@@ -1,32 +1,92 @@
 import { Builder, Browser } from "selenium-webdriver";
 import firefox from "selenium-webdriver/firefox";
+import { backOff } from "exponential-backoff";
+import { logger } from "../utils/logger";
 
 export async function extractWebpageContent(url: string): Promise<string[]> {
+  try {
+    const response = await backOff(async () => {
+      return await extractWebpageContentWorker(url);
+    }, {
+      numOfAttempts: 5,
+      startingDelay: 500,
+      timeMultiple: 2,
+      jitter: "full",
+      maxDelay: 50000,
+    });
+    return response;
+  } catch (err: any) {
+    logger.error(`Failed out of retry loop for URL "${url}", returning placeholder to pipeline`);
+    return ["API EXCEPTION"];
+  }
+}
+
+async function extractWebpageContentWorker(url: string): Promise<string[]> {
+  let driver;
+  try {
     const options = new firefox.Options();
     options.addArguments("--headless");
+    driver = await new Builder()
+      .forBrowser(Browser.FIREFOX)
+      .setFirefoxOptions(options)
+      .build();
+  } catch (err: any) {
+    const desc = `Failed to launch Firefox driver: ${err.message}`;
+    logger.error(desc);
+    throw new Error(desc);
+  }
+
+  try {
+    try {
+      await driver.get(url);
+    } catch (err: any) {
+      const desc = `Failed to navigate to URL "${url}": ${err.message}`;
+      logger.error(desc);
+      throw new Error(desc);
+    }
 
-    let driver = await new Builder().forBrowser(Browser.FIREFOX).setFirefoxOptions(options).build()
     try {
-        await driver.get(url)
       await driver.wait(async () => {
         return await driver.executeScript(
           "return document.readyState === 'complete'"
         );
       }, 5000);
+    } catch (err: any) {
+      logger.error(`Page load timed out for "${url}", attempting to read partial content: ${err.message}`);
+      // do not throw, attempt to read
+    }
 
-        const readableText = await driver.executeScript(
+    let readableText: string;
+    try {
+      readableText = await driver.executeScript(
         "return document.body.innerText;"
       ) as string;
+    } catch (err: any) {
+      const desc = `Failed to extract page text from "${url}": ${err.message}`;
+      logger.error(desc);
+      throw new Error(desc);
+    }
 
     const filteredLines = readableText
       .split(/\r?\n/)
       .map(line => line.trim())
       .filter(line => line.split(/\s+/).length > 1);
 
+    if (filteredLines.length === 0) {
+      const desc = `No content extracted from "${url}"`;
+      logger.error(desc);
+      throw new Error(desc);
+    }
+
     return filteredLines;
   } finally {
-        await driver.quit()
+    try {
+      await driver.quit();
+    } catch (err: any) {
+      logger.error(`Failed to quit Firefox driver cleanly: ${err.message}`);
+    }
   }
 }
 
 // console.log(await extractWebpageContent("https://www.bbc.co.uk/news/live/c74wd01egvyt"))
+// console.log(await extractWebpageContent("https://badcertificate.int.jeynes.uk/"))

supporting/Wrapper/run.ts

@@ -118,7 +118,7 @@ async function processRecord(record: any): Promise<ResultRecord> {
       input: buildAgentInput(record),
       streamMode: "values",
       config: {
-        recursion_limit: 50
+        recursion_limit: 100
       }
     });