Add self improvement pattern with two new prompt nodes

agent/agent.ts

@@ -11,13 +11,18 @@ 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", verificationSetup.name);
+const triggerEventToolConditional = createToolConditional("triggerEventToolNode", selfEvalSetup.name);
+const peToolConditional = createToolConditional("peToolNode", verificationSetup.name);
 
 const roNode = createEnsembleNode("ROBERTA", "roberta");
 const flNode = createEnsembleNode("FLAN", "flan");
@@ -33,6 +38,12 @@ 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)
@@ -49,9 +60,16 @@ const agent = new StateGraph(MessagesState)
   .addEdge(triggerEventSetup.name, "triggerEventModel")
   
   // @ts-expect-error
-  .addConditionalEdges("triggerEventModel", triggerEventToolConditional, ["triggerEventToolNode", verificationSetup.name])
+  .addConditionalEdges("triggerEventModel", triggerEventToolConditional, ["triggerEventToolNode", selfEvalSetup.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/selfEvalSetup.ts

@@ -0,0 +1,21 @@
+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

@@ -13,15 +13,6 @@ export const verificationSetup: 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;
-      parsed[i].context = "NONE"
-    }
-    
     return { proposedTriggerEvent: parsed, proposedTriggerEventIndex: 0 };
   }
   else {

agent/prompts/eval.txt

@@ -0,0 +1,30 @@
+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

@@ -15,6 +15,10 @@ 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);
     }
@@ -33,5 +37,12 @@ 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

@@ -0,0 +1,40 @@
+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

@@ -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.

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()
 })
@@ -20,6 +21,7 @@ export const MessagesState = new StateSchema({
   date: z.string(),
   messages: MessagesValue,
   proposedTriggerEvent: ProposedTriggerEventArray,
+  evalTriggerEvent: ProposedTriggerEventArray,
   proposedTriggerEventIndex: z.int(),
   normalizedClaim: z.string(),
 });

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[]>{
-    const options = new firefox.Options();
-    options.addArguments("--headless");
-
-    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);
-
-        const readableText = await driver.executeScript(
-            "return document.body.innerText;"
-        ) as string;
-
-        const filteredLines = readableText
-            .split(/\r?\n/)
-            .map(line => line.trim())
-            .filter(line => line.split(/\s+/).length > 1); 
-        
-        return filteredLines;
-    } finally {
-        await driver.quit()
-    }
+export async function extractWebpageContent(url: string): Promise<string[]> {
+  try {
+    const response = await backOff(async () => {
+      return await extractWebpageContentWorker(url);
+    }, {
+      numOfAttempts: 10,
+      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"];
+  }
 }
 
-//console.log(await extractWebpageContent("https://www.bbc.co.uk/news/live/c74wd01egvyt"))
+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}": ${err.message}`;
+      logger.error(desc);
+      throw new Error(desc);
+    }
+
+    try {
+      await driver.wait(async () => {
+        return await driver.executeScript(
+          "return document.readyState === 'complete'"
+        );
+      }, 5000);
+    } catch (err: any) {
+      logger.error(`Page load timed out for "${url}", attempting to read partial content: ${err.message}`);
+      // do not throw, attempt to read
+    }
+
+    let readableText: string;
+    try {
+      readableText = await driver.executeScript(
+        "return document.body.innerText;"
+      ) as string;
+    } catch (err: any) {
+      const desc = `Failed to extract page text from "${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 {
+    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
       }
     });