Use a slightly smaller model. Reduce concurreny. Be more clear in the prompts

agent/nodes/model.ts

@@ -10,8 +10,8 @@ export function createModelNode(tools: any, promptPath: string): GraphNode<typeo
         const sysPrompt = await hydratePrompt(promptPath, state);
 
         const model = new ChatOllama({
-            model: "llama3.1:8b-instruct-fp16",
-            temperature: 0.7,
+            model: "llama3.1:8b-instruct-q4_K_M",
+            temperature: 0.3
         });
 
         const modelWithTools = model.bindTools(Object.values(tools));

agent/prompts/normalization.txt

@@ -16,4 +16,4 @@ 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.
+Produce no other text other than the condensed claim in inverted commas. Nothing should be around the normalised claim.

agent/prompts/relation.txt

@@ -1,9 +0,0 @@
-Could the following real-world event:
-###TECLAIM###
-
-Be a trigger for the following disinformation:
-###TITLE###
-
-Respond with "RELATION", followed by : followed by a confidence score (VERYHIGH, HIGH, MEDIUM, LOW, VERYLOW) followed by : followed by the reason. Use no other words, just return the score and reason in format.
-
-Ignore wether the event happened or not, purely consider the likiness of causation

agent/prompts/trigger.txt

@@ -17,6 +17,10 @@ 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.
+Do not include explanations.
+Do not wrap in markdown.
+Do not label where the json is using colons
 
 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.
@@ -26,4 +30,6 @@ 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
 
+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use tools to retreive up to date informaiton everytime, provided examples or existing knowledge will be wrong or out of date.
+
 Lets go through it step by step

agent/prompts/verify.txt

@@ -1,8 +0,0 @@
-Do the search results cited below
-###TESEARCH###
-Support the idea that the following happened:
-###TECLAIM###
-
-Respond with "CONFIDENCE", followed by : followed by a confidence score (VERYHIGH, HIGH, MEDIUM, LOW, VERYLOW) followed by : followed by the reason. Use no other words, just return the score and reason in format.
-
-Dates can be off by a few days, that would still be valid