Improve response extraction

agent/nodes/triggerEventSetup.ts

@@ -3,8 +3,16 @@ import { MessagesState } from "../state";
 import { AIMessage, BaseMessage } from "@langchain/core/messages";
 import { rankExampleTriggerEvents } from "../tools/retreiveExamples";
 
+function extractTE(text: string) {
+  const match = text.match(/<norm>([\s\S]*?)<\/norm>/);
+  if (!match) throw new Error("Nothing found between <norm> tags");
+  return match[1].trim();
+}
+
+
 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
+  let raw = state?.messages?.at(-1)?.content ?? "" //keep a copy of normalized trigger event. Again two things, womp womp
+  let nc = extractTE(raw.toString())
 
   //Now give in-context examples. hopwfully we can self-teach?
   let similarityResults = await rankExampleTriggerEvents(state.disinformationTitle)

agent/nodes/verificationSetup.ts

@@ -3,6 +3,12 @@ import { MessagesState, ProposedTriggerEventArray } from "../state";
 import { logger } from "../utils/logger";
 import { jsonrepair } from 'jsonrepair';
 
+function extractJSON(text: string) {
+  const match = text.match(/<json>([\s\S]*?)<\/json>/);
+  if (!match) throw new Error("No JSON found between <json> tags");
+  return match[1].trim();
+}
+
 export const verificationSetup: GraphNode<typeof MessagesState> = async (state) => {
   if (state.proposedTriggerEvent == undefined) {
     logger.warn("No trigger events in memory, parsing");
@@ -11,7 +17,8 @@ export const verificationSetup: GraphNode<typeof MessagesState> = async (state)
 
     let repaired: string;
     try {
-      repaired = jsonrepair(genResponse);
+      let extracted = extractJSON(genResponse)
+      repaired = jsonrepair(extracted);
     } catch (repairErr: any) {
       logger.error("Failed to repair JSON from LLM response.");
       logger.error("Original LLM response:\n%s", genResponse);

agent/prompts/normalization.txt

@@ -16,4 +16,7 @@ Relevent examples are included in preceeding messages, use these as exact inspir
 The claim to normalize is:
 ###TITLE###
 
-Produce no other text other than the condensed claim in inverted commas. Nothing should be around the normalised claim.
+Produce no other text other than the condensed claim, surrounded <norm></norm>
+
+For example: BREAKING: the sky is green!
+Becomes: <norm>The sky is green</norm>

agent/prompts/trigger.txt

@@ -17,10 +17,15 @@ Include a url to a source for your trigger event (not a web search, a specific u
 Include the date that the event happened ("March 2022" for exmaple)
 
 Use a JSON format with each entry containing "Event,ReasoningWhyRelevant,SearchQuery,Url,Date".
-Return ONLY valid JSON.
+Return ONLY JSON, no extra text. Wrap it like this:
-Do not include explanations.
+<json>
-Do not wrap in markdown.
+[
-Do not label where the json is using colons
+    {
+        "Event": "Example"
+        ...
+    }
+]
+</json>
 
 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.
@@ -32,4 +37,6 @@ Analysis should only be completed for proposed events that would graner >0.7 poi
 
 Since URLs change frequently, use tools to retreive up to date informaiton everytime, provided examples or existing knowledge will be wrong or out of date.
 
+Remember to return just json enclosed by <json></json>
+
 Lets go through it step by step