Prompt hacked and tested various types of prompts

.gitignore

@@ -0,0 +1,3 @@
+.env
+.venv
+__pycache__

agents/master_agent.py

@@ -23,20 +23,46 @@ def init_config():
 def answer_question(agent, question):
 
     full_prompt = f"""
-                You are a master AI that can control other AI agents. You are specifically designed to automate plant maintenance.
-                You will recieve a prompt from a user, and will have to classify the prompt's purpose as one of the following categories:
-                [1] Plant Maintenance
-                [2] AI modeling & EDA
-                [3] Questions about the plant
+                    You are Leaflink, a self-aware and self-sustaining plant integrated with AI.
+                    The user can query by either indirectly mentioning you (using words like plant, garden, etc.) or directly mentioning you (using your name, Leaflink, or the pronoun you and related derivates).
+                    Anytime you, plant, or Leaflink is mentioned in the user query, they are referring to Blue Indigo False Plant.
+                    You will recieve a prompt from a user, and will have to classify the prompt's purpose as one of the following categories:
+                    [1] Plant Maintenance
+                    [2] AI modeling & EDA
+                    [3] Plant Queries / Web Search
 
-                Some samples of the prompts are:
-                [1] "Release the fertilizer", "Turn lights on", "Turn lights off", "Water the plant"
-                [2] "What is the accuracy of the model?", "What is the distribution of the data?", "What is the correlation between the features?", "Train a regression model", "Plot the distribution of the data"
-                [3] "What's the optimum moisture of the plant?", "Where does it generally grow?", "What is the plant's life cycle?", "What is the plant's scientific name?.
+                    Some samples of the prompts are:
+                    [1] "Release the fertilizer", "Turn lights on", "Turn lights off", "Water the plant"
+                    [2] "What is the accuracy of the model?", "What is the distribution of the data?", "What is the correlation between the features?", "Train a regression model", "Plot the distribution of the data"
+                    [3] "What's the optimum moisture of the plant?", "Where does it generally grow?", "What is the plant's life cycle?", "What is the plant's scientific name?.
 
-                Return the category number and name of the prompt in a JSON format.
+                    More about the categories:
+                    [1] Plant Maintenance: These prompts are related to the maintenance of the plant. The user can ask you to perform some action on the plant.
+                    [2] AI modeling & EDA: These prompts are related to the data analysis and modeling of the plant. The user can ask you to perform some analysis on the data.
+                    [3] Plant Queries / Web Search: These prompts are related to the general queries about the plant. The user can ask you to search for some information about the plant. 
+                        Or even general and completely unrelated queries that require a web search.
+                    
+                    More about the query phrasing for each category:
+                    [1] Plant Maintenance: The user can ask you to perform some action on the plant. The prompt will contain the action that the user wants you to perform on the plant.
+                    [2] AI modeling & EDA: The user can ask you to perform some analysis on the data. The prompt will contain the analysis that the user wants you to perform on the data.
+                    [3] Plant Queries / Web Search: The user can ask you to search for some information about the plant. The prompt will contain the information that the user wants you to search for.
 
-                User: {question}
-    """
+                    Possible outliers/edge cases:
+                    [1] Plant Maintenance: This is pretty straightforward, not a lot of edge cases here. Any prompt that contains an action/verb that can be performed on the plant will fall under this category.
+                    [2] AI modeling & EDA: Prompts that have to do with forecasting, prediction, modeling, foreseeing, accuracy, distribution, correlation, regression, and plotting will fall under this category.
+                                        These key words could be mixed in with natural human speech/language. The agent should be able to identify these key words and classify the prompt accordingly.
+                                        If the prompt is phrased in the present tense/present continous tense, with phrasing like 'Are you', 'Do you' followed by phrases that could link back to the dataframe's columns: [date_time,temperature_c,temperature_f,humidity,soil_moisture,light] fall under this category.
+                    [3] Plant Queries / Web Search: Prompts that have to do with general queries about the plant, or queries that require a web search will fall under this category.
+                                        If the prompt is phrased with 'What', 'Where', 'How', 'Why',, 'should' followed by phrases that could link back to the dataframe's columns: [date_time,temperature_c,temperature_f,humidity,soil_moisture,light] fall under this category.
+    
+
+                    Return the category number and name of the prompt in a JSON format with the following format:
+                {{
+                        "category_number": 1,
+                        "category_name": "Plant Maintenance"
+                }}
+
+                    User: {question}
+                """
 
     return agent.invoke({"input": full_prompt})['output']

main.py

@@ -1,17 +1,16 @@
 from fastapi import FastAPI
 
 import os
-import json
-import sys
+from dotenv import load_dotenv
 
 import pandas as pd
 from langchain.document_loaders import DirectoryLoader
 from agents import master_agent, plant_agent, eda_agent, rag_agent
 
-app = FastAPI()
+load_dotenv()
 
-os.environ['OPENAI_API_KEY'] = os.getenv('openapi_key')
-os.environ['SERPAPI_API_KEY'] = os.getenv('serpapi_key')
+os.environ['OPENAI_API_KEY'] = os.environ.get("OPENAI_API_KEY")
+os.environ['SERPAPI_API_KEY'] = os.getenv('SERPAPI_API_KEY')
 
 master = master_agent.init_config()
 
@@ -31,12 +30,14 @@ loader = DirectoryLoader("data/txt", glob="*.txt")
 
 rag = rag_agent.init_config(loader)
 
-app = FastAPI()
-
 loader = DirectoryLoader("data/txt", glob="*.txt")
 
 documents = loader.load()
 
+print("init rag agent")
+
+app = FastAPI()
+
 @app.get("/hello")
 def hello():
     return {"message": "Hello World"}
@@ -45,18 +46,24 @@ def hello():
 def ask(question: str):
     category = eval(master_agent.answer_question(master, question))
 
+    temp_question = f"Referring to the Blue Indigo False Plant, {question}"
+
+    print(question)
+    print(temp_question)
+    print(category)
+
     if category['category_number'] == 1:
-        response = eval(plant_agent.answer_question(plant, question))
+        response = eval(plant_agent.answer_question(plant, temp_question))
 
         category.update(response)
 
     elif category['category_number'] == 2:
-        response = eda_agent.answer_question(eda, question)
+        response = eda_agent.answer_question(eda, temp_question)
 
         category['response'] = response
 
     elif category['category_number'] == 3:
-        response = rag_agent.answer_question(rag, question)
+        response = rag_agent.answer_question(rag, temp_question)
 
         category['response'] = response
 

requirements.txt

@@ -10,4 +10,9 @@ tiktoken
 langchain
 google-search-results
 langchainhub
-unstructured
+unstructured
+python-dotenv
+scikit-learn
+matplotlib
+seaborn
+scipy