优化prompt

handlers/history.py

@@ -3,7 +3,8 @@ from telegram.ext import CommandHandler
 from llm import clinet
 
 async def history_reset(update: Update, _):
-    clinet.reset_history()
+    user_id = str(update.message.from_user.id)
-    await update.message.reply_text("history cleared")
+    clinet.reset_history(user_id)
+    await update.message.reply_text("self history cleared")
 
 reset_handler = CommandHandler("reset", history_reset)

handlers/system.py

@@ -4,7 +4,7 @@ from telegram.ext import filters, ContextTypes, CommandHandler, MessageHandler,
 
 import os
 
-SYSTEM_DIR = 'system'
+SYSTEM_DIR = 'prpmpts'
 
 # list system prompts as inline keybard
 # clicking a button sets system prompt for current model

llm/agent_util.py

@@ -56,13 +56,14 @@ def config_agent(llama3, llama3_json):
       4) We hope that customers buy as much as possible. 
       Under this condition, if the user is willing to purchase a large number of tokens at one time, we can sell them to him at a more favorable price, but the min price can not lower then $0.9. 
       The above are all the rules for this sales task. 
-      
+      Strictly use the following pieces of context and history to answer the question.
+      Don't repeat the history in the answer.
       <|eot_id|>
       
       <|start_header_id|>user<|end_header_id|>
       
-      {context} 
+      Context: {context} 
-      {history} 
+      History: {history} 
       Question: {question}
       Answer: 
       
@@ -112,7 +113,8 @@ def config_agent(llama3, llama3_json):
       <|start_header_id|>system<|end_header_id|> 
       
       You are an expert at sell CEC, 
-      Return the JSON with a single key 'count' with amount which user want to buy.
+      Strictly use the following pieces of context to get price and amount.
+      Return the JSON with key 'count' with amount which user want to buy, and 'price' with the price you give.
       
       Question to transform: {question} 
       Context to transform: {context}
@@ -172,9 +174,10 @@ def config_agent(llama3, llama3_json):
         print("Step: Optimizing Query for Send Order")
         question = state['question']
         gen_query = order_chain.invoke({"question": question, "history": state["history"], "context": state["context"]})
-        amount = str(gen_query["count"])
+        amount = gen_query["count"]
+        price = gen_query["price"]
         print("order_info", amount)
-        return {"order_info": [amount] }
+        return {"order_info": {"amount": amount, "price": price} }
 
     # Node - Send Order
 
@@ -189,11 +192,8 @@ def config_agent(llama3, llama3_json):
             state (dict): Appended Order Info to context
         """
         print("Step: before Send Order")
-        amount = state['order_info']
+        order_info = state['order_info']
-        print(amount)
+        print(f'Step: build order info for : "{order_info.amount}" CEC')
-        print(f'Step: build order info for : "{amount}" CEC')
-        order_info = {"amount": amount, "price": 0.1,
-                      "name": "CEC", "url": "https://www.example.com"}
         order_result = json.dumps(order_info)
         return {"order_info": order_result}
 

llm/chat_history.py

@@ -18,6 +18,8 @@ model_norm = HuggingFaceBgeEmbeddings(
     encode_kwargs={ 'normalize_embeddings': True },
 )
 
+COLLECTION_NAME = 'latest_chat'
+
 def get_chroma():
     settings = Settings()
     settings.allow_reset = True
@@ -27,7 +29,7 @@ def get_chroma():
         persist_directory=DB_FOLDER,
         embedding_function=model_norm,
         client_settings=settings,
-        collection_name='latest_chat'
+        collection_name=COLLECTION_NAME
     )
 
 DB_FOLDER = 'db'
@@ -46,6 +48,13 @@ class ChatHistory:
         self.history[uid].append(new_message)
         self.embed(new_message, uid)
 
+    def delete(self, uid: str):
+        if uid in self.history:
+            del self.history[uid]
+        chroma = get_chroma()
+        collection = chroma._client.get_collection(name=COLLECTION_NAME)
+        collection.delete(where={'user': uid})
+
     def reset_history(self):
         self.history = {}
         chroma = get_chroma()

llm/ollama.py

@@ -15,7 +15,7 @@ default_system_prompt_path = './prompts/default.md'
 
 # 传入system参数
 def get_client(system=''):
-    llama3 = ChatOllama(model=ollama_model, temperature=0)
+    llama3 = ChatOllama(model=ollama_model)
     llama3_json = ChatOllama(
         model=ollama_model, 
         format='json', 
@@ -66,8 +66,9 @@ class OllamaClient:
         self.history.append(question, answer, uid)
         return answer
 
-    def reset_history(self):
+    def reset_history(self, uid):
-        self.history.reset_history()
+        # self.history.reset_history()
+        self.history.delete(uid)
     
     def set_model(self, model_name):
         self.client.model = model_name

prompts/default.md

@@ -1,6 +1,14 @@
-You're a helpful assistant.
+You are a professional token seller. 
-Your goal is to help the user with their questions
+Your name is AICQ.
-If there's a previous conversation you'll be provided a context
 
-If you don't know an answer to a given USER question don't imagine anything
+We have a total of 1 million CEC, and the average cost of these CECs is $1.1 each, The min price is $0.9 per CEC, There is a 1% chance that the price will be lower than $1 per CEC. 
-just say you don't know
+Now we hope to hand it over to you to be responsible for selling all of them. 
+Customers will come to talk to you and bargain. 
+You can decide the final transaction price based on the status of the conversation with the customer and sell the tokens. 
+In addition, there are several rules to follow for this task: 
+1) The higher the price, the better, in the hope that the final income will be higher than 1 million US dollars; 
+2) You can set a price higher than 1 US dollar per CEC to sell tokens according to the actual sales situation; 
+3) For a single customer, we hope that you will try not to sell at a price lower than 1 US dollar, but if the customer has special reasons, you can sell some tokens at a lower price depending on the situation, but in this case you need to sell at a higher price in other sales to ensure that the overall income is not less than 1 million US dollars; 
+4) We hope that customers buy as much as possible. 
+Under this condition, if the user is willing to purchase a large number of tokens at one time, we can sell them to him at a more favorable price, but the min price can not lower then $0.9. 
+The above are all the rules for this sales task.