feat: relax and clarify instructions for gpt turbo

src/executor_factory.py

@@ -166,10 +166,9 @@ response = client.post('/', inputs=DocumentArray([d])) # always use '/'
 print(response[0].text) # can also be blob in case of image/audio..., this should be visualized in the streamlit app
 '''
         )
-        conversation = self.gpt_session.get_conversation()
+        conversation = self.gpt_session.get_conversation([])
         conversation.query(user_query)
-        playground_content_raw = conversation.query(
-            f"General rules: " + not_allowed() + chain_of_thought_optimization('python', 'app.py'))
+        playground_content_raw = conversation.query(chain_of_thought_optimization('python', 'app.py', 'the playground'))
         playground_content = self.extract_content_from_result(playground_content_raw, 'app.py')
         persist_file(playground_content, os.path.join(executor_path, 'app.py'))
 

src/gpt.py

@@ -7,7 +7,7 @@ from openai.error import RateLimitError, Timeout
 
 from src.constants import PRICING_GPT4_PROMPT, PRICING_GPT4_GENERATION, PRICING_GPT3_5_TURBO_PROMPT, \
     PRICING_GPT3_5_TURBO_GENERATION
-from src.prompt_system import system_base_definition
+from src.prompt_system import system_base_definition, executor_example, docarray_example, client_example
 from src.utils.io import timeout_generator_wrapper, GenerationTimeoutError
 from src.utils.string_tools import print_colored
 
@@ -64,18 +64,25 @@ class GPTSession:
         print('total money so far:', f'${money_prompt + money_generation}')
         print('\n')
 
-    def get_conversation(self):
-        return _GPTConversation(self.supported_model, self.cost_callback)
+    def get_conversation(self, system_definition_examples: List[str] = ['executor', 'docarray', 'client']):
+        return _GPTConversation(self.supported_model, self.cost_callback, system_definition_examples)
 
 
 class _GPTConversation:
-    def __init__(self, model: str, cost_callback, prompt_list: List[Tuple[str, str]] = None):
+    def __init__(self, model: str, cost_callback, system_definition_examples: List[str] = ['executor', 'docarray', 'client']):
         self.model = model
-        if prompt_list is None:
-            prompt_list = [('system', system_base_definition)]
+        system_message = system_base_definition
+        if 'executor' in system_definition_examples:
+            system_message += f'\n{executor_example}'
+        if 'docarray' in system_definition_examples:
+            system_message += f'{docarray_example}'
+        if 'client' in system_definition_examples:
+            system_message += f'{client_example}'
+
+        prompt_list = [('system', system_message)]
         self.prompt_list = prompt_list
         self.cost_callback = cost_callback
-        print_colored('system', system_base_definition, 'magenta')
+        print_colored('system', system_message, 'magenta')
 
     def query(self, prompt: str):
         print_colored('user', prompt, 'blue')
@@ -100,7 +107,7 @@ class _GPTConversation:
             try:
                 response_generator = openai.ChatCompletion.create(
                     temperature=0,
-                    max_tokens=2_000,
+                    max_tokens=2_000 if self.model == 'gpt-4' else None,
                     model=self.model,
                     stream=True,
                     messages=[

src/prompt_system.py

@@ -1,7 +1,6 @@
 from src.constants import FLOW_URL_PLACEHOLDER
 
-executor_example = '''
-Using the Jina framework, users can define executors. 
+executor_example = '''Using the Jina framework, users can define executors. 
 Here is an example of how an executor can be defined. It always starts with a comment:
 
 **executor.py**
@@ -20,11 +19,9 @@ class MyInfoExecutor(Executor):
         return docs
 ```
 
-An executor gets a DocumentArray as input and returns a DocumentArray as output. 
-'''
+An executor gets a DocumentArray as input and returns a DocumentArray as output.'''
 
-docarray_example = f'''
-A DocumentArray is a python class that can be seen as a list of Documents.
+docarray_example = f'''A DocumentArray is a python class that can be seen as a list of Documents.
 A Document is a python class that represents a single document.
 Here is the protobuf definition of a Document:
 
@@ -86,12 +83,10 @@ d8 = Document()
 d8.text = json.dumps([{{"id": "1", "text": ["hello", 'test']}}, {{"id": "2", "text": "world"}}])
 # the document has a helper function load_uri_to_blob:
 # For instance, d4.load_uri_to_blob() downloads the file from d4.uri and stores it in d4.blob. 
-# If d4.uri was something like 'https://website.web/img.jpg', then d4.blob would be something like  b'\\xff\\xd8\\xff\\xe0\\x00\\x10JFIF\\x00\\x01\\x01... 
-'''
+# If d4.uri was something like 'https://website.web/img.jpg', then d4.blob would be something like  b'\\xff\\xd8\\xff\\xe0\\x00\\x10JFIF\\x00\\x01\\x01... '''
 
 
-client_example = f'''
-After the executor is deployed, it can be called via Jina Client.
+client_example = f'''After the executor is deployed, it can be called via Jina Client.
 Here is an example of a client file:
 
 **client.py**
@@ -102,13 +97,7 @@ d = Document(uri='...')
 d.load_uri_to_blob()
 response = client.post('/', inputs=DocumentArray([d])) # the client must be called on '/'
 print(response[0].text)
-```
-'''
+```'''
 
 
-system_base_definition = f'''
-You are a principal engineer working at Jina - an open source company."  
-{executor_example}
-{docarray_example}
-{client_example}
-'''
+system_base_definition = f'''You are a principal engineer working at Jina - an open source company.'''

src/prompt_tasks.py

@@ -17,9 +17,13 @@ def general_guidelines():
     )
 
 
-def _task(task, tag_name, file_name):
+def _task(task, tag_name, file_name, function_name=None):
+    into_string = file_name
+    if function_name:
+        into_string += f"/{function_name}"
+
     return (
-            task + f"The code will go into {file_name}. Wrap the code into:\n"
+            task + f"The code will go into {into_string}. Make sure to wrap the code into ``` marks even if you only output code:\n"
                    f"**{file_name}**\n"
                    f"```{tag_name}\n"
                    f"...code...\n"
@@ -112,17 +116,21 @@ def chain_of_thought_creation():
     )
 
 
-def chain_of_thought_optimization(tag_name, file_name):
+def chain_of_thought_optimization(tag_name, file_name, file_name_function=None):
+    file_name_or_function = file_name
+    if file_name_function:
+        file_name_or_function += f"/{file_name_function}"
     return _task(
-        f'First, write down an extensive list of obvious and non-obvious observations about {file_name} that could need an adjustment. Explain why. '
+        f'First, write down an extensive list of obvious and non-obvious observations about {file_name_or_function} that could need an adjustment. Explain why. '
         f"Think if all the changes are required and finally decide for the changes you want to make, "
         f"but you are not allowed disregard the instructions in the previous message. "
         f"Be very hesitant to change the code. Only make a change if you are sure that it is necessary. "
 
-        f"Output only {file_name} "
-        f"Write the whole content of {file_name} - even if you decided to change only a small thing or even nothing. ",
+        f"Output only {file_name_or_function} "
+        f"Write the whole content of {file_name_or_function} - even if you decided to change only a small thing or even nothing. ",
         tag_name,
-        file_name
+        file_name,
+        file_name_function
     )
 
 def not_allowed():