Merge pull request #25 from testpower4/main

fix token exceeding LLM limitation issue

Author: pinegreene

agents/application/executor.py

@@ -2,6 +2,9 @@
 import json
 import ast
 import re
+from typing import List, Dict, Any
+
+import math
 
 from dotenv import load_dotenv
 from langchain_core.messages import HumanMessage, SystemMessage
@@ -13,14 +16,27 @@
 from agents.application.prompts import Prompter
 from agents.polymarket.polymarket import Polymarket
 
+def retain_keys(data, keys_to_retain):
+    if isinstance(data, dict):
+        return {
+            key: retain_keys(value, keys_to_retain)
+            for key, value in data.items()
+            if key in keys_to_retain
+        }
+    elif isinstance(data, list):
+        return [retain_keys(item, keys_to_retain) for item in data]
+    else:
+        return data
 
 class Executor:
-    def __init__(self) -> None:
+    def __init__(self, default_model='gpt-3.5-turbo-16k') -> None:
         load_dotenv()
+        max_token_model = {'gpt-3.5-turbo-16k':15000, 'gpt-4-1106-preview':95000}
+        self.token_limit = max_token_model.get(default_model)
         self.prompter = Prompter()
         self.openai_api_key = os.getenv("OPENAI_API_KEY")
         self.llm = ChatOpenAI(
-            model="gpt-3.5-turbo",
+            model=default_model, #gpt-3.5-turbo"
             temperature=0,
         )
         self.gamma = Gamma()
@@ -43,9 +59,12 @@ def get_superforecast(
         result = self.llm.invoke(messages)
         return result.content
 
-    def get_polymarket_llm(self, user_input: str) -> str:
-        data1 = self.gamma.get_current_events()
-        data2 = self.gamma.get_current_markets()
+
+    def estimate_tokens(self, text: str) -> int:
+        # This is a rough estimate. For more accurate results, consider using a tokenizer.
+        return len(text) // 4  # Assuming average of 4 characters per token
+
+    def process_data_chunk(self, data1: List[Dict[Any, Any]], data2: List[Dict[Any, Any]], user_input: str) -> str:
         system_message = SystemMessage(
             content=str(self.prompter.prompts_polymarket(data1=data1, data2=data2))
         )
@@ -54,6 +73,55 @@ def get_polymarket_llm(self, user_input: str) -> str:
         result = self.llm.invoke(messages)
         return result.content
 
+
+    def divide_list(self, original_list, i):
+        # Calculate the size of each sublist
+        sublist_size = math.ceil(len(original_list) / i)
+        
+        # Use list comprehension to create sublists
+        return [original_list[j:j+sublist_size] for j in range(0, len(original_list), sublist_size)]
+    
+    def get_polymarket_llm(self, user_input: str) -> str:
+        data1 = self.gamma.get_current_events()
+        data2 = self.gamma.get_current_markets()
+        
+        combined_data = str(self.prompter.prompts_polymarket(data1=data1, data2=data2))
+        
+        # Estimate total tokens
+        total_tokens = self.estimate_tokens(combined_data)
+        
+        # Set a token limit (adjust as needed, leaving room for system and user messages)
+        token_limit = self.token_limit
+        if total_tokens <= token_limit:
+            # If within limit, process normally
+            return self.process_data_chunk(data1, data2, user_input)
+        else:
+            # If exceeding limit, process in chunks
+            chunk_size = len(combined_data) // ((total_tokens // token_limit) + 1)
+            print(f'total tokens {total_tokens} exceeding llm capacity, now will split and answer')
+            group_size = (total_tokens // token_limit) + 1 # 3 is safe factor
+            keys_no_meaning = ['image','pagerDutyNotificationEnabled','resolvedBy','endDate','clobTokenIds','negRiskMarketID','conditionId','updatedAt','startDate']
+            useful_keys = ['id','questionID','description','liquidity','clobTokenIds','outcomes','outcomePrices','volume','startDate','endDate','question','questionID','events']
+            data1 = retain_keys(data1, useful_keys)
+            cut_1 = self.divide_list(data1, group_size)
+            cut_2 = self.divide_list(data2, group_size)
+            cut_data_12 = zip(cut_1, cut_2)
+
+            results = []
+
+            for cut_data in cut_data_12:
+                sub_data1 = cut_data[0]
+                sub_data2 = cut_data[1]
+                sub_tokens = self.estimate_tokens(str(self.prompter.prompts_polymarket(data1=sub_data1, data2=sub_data2)))
+
+                result = self.process_data_chunk(sub_data1, sub_data2, user_input)
+                results.append(result)
+            
+            combined_result = " ".join(results)
+            
+        
+            
+            return combined_result
     def filter_events(self, events: "list[SimpleEvent]") -> str:
         prompt = self.prompter.filter_events(events)
         result = self.llm.invoke(prompt)