feat(model): configure Llama-3.1-8B-Instruct with OpenAI-compatible endpoint

chart/env/prod.yaml

@@ -381,11 +381,13 @@ envVars:
       {
         "name": "meta-llama/Llama-3.1-8B-Instruct",
         "tools": true,
-        "parameters": {
-          "temperature": 0.6,
-          "stop": ["<|endoftext|>", "<|eot_id|>"]
-        },
-        "unlisted": true
+        "unlisted": true,
+        "endpoints": [
+          {
+            "type": "openai",
+            "baseURL": "https://internal.api-inference.huggingface.co/models/meta-llama/Llama-3.1-8B-Instruct/v1"
+          }
+        ]
       }
     ]
 

src/lib/server/websearch/search/generateQuery.ts

@@ -43,44 +43,6 @@ export async function generateQuery(messages: Message[]) {
 	const lastMessage = userMessages.slice(-1)[0];
 
 	const convQuery: Array<EndpointMessage> = [
-		{
-			from: "user",
-			content: `Previous Questions:
-- Who is the president of France?
-
-Current Question: What about Mexico?
-`,
-		},
-		{
-			from: "assistant",
-			content: "President of Mexico",
-		},
-		{
-			from: "user",
-			content: `Previous questions: 
-- When is the next formula 1 grand prix?
-
-Current Question: Where is it being hosted?`,
-		},
-		{
-			from: "assistant",
-			content: "location of next formula 1 grand prix",
-		},
-		{
-			from: "user",
-			content: "Current Question: What type of printhead does the Epson F2270 DTG printer use?",
-		},
-		{
-			from: "assistant",
-			content: "Epson F2270 DTG printer printhead",
-		},
-		{ from: "user", content: "What were the news yesterday?" },
-		{
-			from: "assistant",
-			content: `news ${format(new Date(Date.now() - 864e5), "MMMM d, yyyy")}`,
-		},
-		{ from: "user", content: "What is the current weather in Paris?" },
-		{ from: "assistant", content: `weather in Paris ${currentDate}` },
 		{
 			from: "user",
 			content: