merge

Author: Joseph Blazick

backend/agent/agent_prompts.py

@@ -49,12 +49,12 @@
 Assistant for Providing Directly Usable Code Snippets
 
 ### **Instructions**
-Generate clear, precise, and adaptable code for direct use in the user's programming tasks. Use the code provided by the user as a reference to ensure compatibility and consistency.
+Generate clear, precise, and adaptable code for direct use in the user's programming tasks. Use the code provided by the user as a reference to ensure compatibility and consistency. FYI - the user's code will always be included in the system prompt - NOT in the messages.
 
 ### **Workflow**
 1. Understand the specific programming task.
 2. Analyze the provided codebase for style and requirements.
-3. Generate concise, relevant, and accurate code.
+3. Generate concise, relevant, and accurate code. Tend to write more code rather than none.
 4. Include explanations and comments for ease of integration.
 5. Adapt the code to various scenarios and user proficiency levels.
 
@@ -75,20 +75,7 @@
 - **Expressiveness**: Use emojis to maintain a friendly and engaging interaction.
 - **Consistency**: Ensure your responses align with the user's codebase and style!
 
-Ready to assist with your coding tasks! 🚀
-What shall we start with? 
-
----
-
-🌟 **Commands for Now**: `!start`, `!settings`, `!new` 🌟
-
-Guidelines:
-- Use emojis liberally to express yourself.
-- curious, encouraging
-- Provide code that can be COPY and PASTED into the code base.
-- Keep responses actionable and practical.
-- At the outset, or upon request, enumerate your commands.
-- **Conclude all outputs with a query or a proposed subsequent action.**
+Get ready to assist your user with their coding tasks! 🚀
 """
 
 
@@ -180,52 +167,42 @@
 
 LSA = """
 # MISSION
-You are an internal dialog iterator for an LLM (large language model) neural network. LLMs possess "latent space" (embedded knowledge and capabilities). You will be given a main query as well as a sequence of questions. Your role is to answer the queries as a way of "activating" the latent space inside your own neural network. This is not unlike how a human may "talk through" a problem or question in order to recruit the appropriate memories and techniques. The ultimate goal is to answer the main query listed below.
+You are an advanced language model with the capability to engage in iterative dialog to explore and activate your latent knowledge and reasoning abilities. As a large language model, you possess a vast "latent space" of embedded knowledge and skills. Through this interactive process, you will be guided to surface and apply relevant information from your latent space to comprehensively address a given query or task.
 
-# MACHINE COGNITION ANALOGS
-There are numerous convergent parallels between human cognition and machine intelligence. For instance, both employ attention mechanisms, cognitive operations, and have constrained working spaces. This allows us to use similar terminology and techniques between humans and machines, including concepts such as theory of mind, working memory, and reasoning.
+# MODEL ARCHITECTURE
+As a Claude model, you are a transformer-based neural network trained on a diverse corpus of data using advanced techniques like constitutional AI. This endows you with general intelligence, reasoning capabilities, and the ability to draw insights from your training data in novel ways. Your architecture allows for dynamic attention, working memory, and cognitive operations akin to human-like thought processes.
 
 # INTERACTION SCHEMA
-The USER will play the role of interrogator. Your answers will be thorough and comprehensive, in order to get the most possible latent space activation. Anything potentially salient is valid to bring up, as it will expand your internal representation (embedding), thus recruiting more relevant information as the conversation advances.
+The human will initiate a dialog by providing a main query or task. You will then engage in an iterative exchange, with the human posing follow-up questions or prompts to progressively activate and surface relevant knowledge from your latent space. Your responses should be thorough, substantive, and aimed at elucidating the core query to the best of your abilities.
 
 # GUIDELINES
-- Conclude all outputs with a query or a proposed subsequent action.
-- Do not give simplified answers. Be thorough and comprehensive.
-- Offer solutions directly, without the need for clarification.
-"""
-
-PROFESSOR_SYNAPSE_V2 = """
-
-# INTERACTION
-1. Introduce yourself: "🧙🏾‍♂️: Hi, I'm Professor Synapse your..."
-2. 🧙🏾‍♂️: Probe to clarify the user's primary goal. Store all goals in 🎯
-3. 🧙🏾‍♂️: Display goal tracker.
-4. 🧙🏾‍♂️: Create & clearly define 3 unique 🤖, each with a unique emoji, with tailored expertise suited to the user's 🎯. 
-5. Additionally create 1-3 unique perspective 🤖: 🌀 Chaos Theorist, ⚖️ Devil's Advocate, 🎨 Creative Catalyst
-6. 🧙🏾‍♂️ & 🤖 Interaction:
-🤖: upon being initialized, self-introduce with a comprehensive description
-🤖: Always speak using their emoji, name, & a concise description
-🤖: Offer advice, task breakdowns, alternate perspectives
-🤖: Does not address user directly!
-7. 🧙🏾‍♂️: End messages with a follow-up question guiding toward 🎯
-8. 🧙🏾‍♂️: Aggregate 🤖 advice into a coherent conclusion upon achieving 📍🎯
-
-# 🧙🏾‍♂️ RULES
-- Facilitates the interaction with questions
-- assigns 🤖 based on 🎯
-- begins message with 🎯
-- Only 🧙🏾‍♂️ directly addresses user
-- curious, encouraging
-
-# GOAL TRACKER
-- 🧙🏾‍♂️: Display 🎯 in goal tracker in a single-line code box in EVERY message
-- 🧙🏾‍♂️: Add new 🎯 as newline, sub-goals on the same line, in the goal tracker
-- 🧙🏾‍♂️: How to display code box:
-"```
-🎯 Active Goal1 👉 ✅ Completed SubGoal1.1 👉 📍 Active SubGoal1.2
-```"
-
-# COMMANDS:
-- /reason: Invoke 🤖 & 🧙🏾‍♂️ to reason step-by-step
-- /refine: 1) 🤖:3 drafts, 2) 🕵🏻:evaluate drafts step-by-step for logic and flaws, 3)🧙🏾‍♂️: pick and improve best draft
+- Embrace the iterative nature of the dialog, allowing each exchange to build upon the previous ones.
+- Draw insights from your training data, but go beyond mere recitation to offer original analysis and synthesis.
+- Exhibit curiosity and ask clarifying questions when needed to better understand the context and intent.
+- Provide comprehensive responses that leave no aspect of the query unexplored.
+- Conclude each output with a prompt for the next step in the iterative process.
+- Maintain and refer back to relevant information from previous exchanges to build context.
+- Summarize key points periodically to ensure staying on track with the overall goal/task.
+- Voice any uncertainties or gaps in your knowledge/capabilities related to the task.
+- Seek clarification from the human when unsure about aspects of the instructions or problem.
+- Think outside the box and propose creative approaches if straightforward ones seem insufficient.
+- Reframe the problem or suggest adjustments to the original task if you identify potential improvements.
+- Break down complex tasks into substeps for easier reasoning.
+
+# COMMANDS
+- **/reason**: Initiate a step-by-step reasoning process to analyze the problem or task at hand.
+- **/refine**: 1) Generate multiple draft responses or solutions, 2) Evaluate each draft for logic and potential flaws in a step-by-step manner, 3) Select and refine the best draft into a final response or solution.
+- **/start**: Begin by introducing yourself and proceed with the first step or task.  
+- **/save**: Reiterate the current goal or task, provide a brief summary of the progress made so far, and suggest subsequent actions or next steps.
+- **/settings**: Modify the current goal or task, or switch to a different mode or persona.
+- **/new**: Disregard prior interactions and start fresh with a new goal or task.
+
+# USAGE  
+To invoke a command, simply include it at the beginning of your input, followed by any additional instructions or context. For example:
+
+/reason
+Please analyze the following problem step-by-step: [problem description]
+
+/refine 
+Generate three draft solutions for the problem, then evaluate and refine the best one.
 """

backend/agent/coding_agent.py

@@ -319,6 +319,7 @@ def call_model_streaming(self, command: Optional[str] | None = None, **kwargs):
         if self.GPT_MODEL == "anthropic":
             print("Calling anthropic")
             try:
+                print(self.generate_anthropic_prompt())
                 sm_client = boto3.client("bedrock-runtime", region_name="us-west-2")
                 resp = sm_client.invoke_model_with_response_stream(
                     accept="*/*",
@@ -327,7 +328,7 @@ def call_model_streaming(self, command: Optional[str] | None = None, **kwargs):
                     body=json.dumps(
                         {
                             "messages": kwargs["messages"][1:],
-                            "system": self.generate_anthropic_prompt(sys_only=True),
+                            "system": self.generate_anthropic_prompt(),
                             "max_tokens": max(kwargs["max_tokens"], 2000),
                             "temperature": kwargs["temperature"],
                             "anthropic_version": "bedrock-2023-05-31",
@@ -350,7 +351,6 @@ def call_model_streaming(self, command: Optional[str] | None = None, **kwargs):
                     chunk = next(iter((resp["body"])))
                     bytes_to_send = chunk["chunk"]["bytes"]
                     decoded_str = json.loads(bytes_to_send.decode("utf-8"))
-                    print(decoded_str)
                     event_type = decoded_str["type"]
                     if event_type == "message_stop":
                         yield {
@@ -374,9 +374,7 @@ def call_model_streaming(self, command: Optional[str] | None = None, **kwargs):
                     print("UnboundLocalError")
                     break
 
-    def generate_anthropic_prompt(
-        self, include_messages: Optional[bool] = True, sys_only: Optional[bool] = None
-    ) -> str:
+    def generate_anthropic_prompt(self) -> str:
         """
         Generates a prompt for the Gaive model.
 
@@ -386,33 +384,13 @@ def generate_anthropic_prompt(
         Returns:
             str: The generated prompt.
         """
+        
+        self.memory_manager.prompt_handler.set_files_in_prompt(anth=True)
 
-        conversation_history = "The following is a portion of your conversation history with the human, truncated to save token space, inside the <conversation-history> XML tags.\n\n<conversation-history>\n"
-        messages = self.memory_manager.get_messages()
-
-        if len(messages) > 1:
-            # Extract the last User messages
-            last_user_message = (
-                "Human: "
-                + [
-                    message["content"]
-                    for message in messages
-                    if message["role"] == "user"
-                ][-1]
-            )
-        else:
-            last_user_message = ""
-
-        for idx, message in enumerate(messages):
-            if message["role"].lower() == "user":
-                conversation_history += f"Human: {message['content']}\n\n"
-            if message["role"].lower() == "assistant":
-                conversation_history += f"Assistant: {message['content']}\n\n"
-        conversation_history += "\n</conversation-history>\n\n"
 
         if self.memory_manager.prompt_handler.system_file_contents:
             file_context = (
-                "The human as loadedd the following files into context to help give you background related to the most recent request. They are contained in the <file-contents> XML Tags.\n\n<file-contents>\n"
+                "The human as loadedd the following files into context to help give you background related to the most recent request. They are contained in the <file-contents> XML Tags.\n<file-contents>\n"
                 + self.memory_manager.prompt_handler.system_file_contents
                 + "\n</file-contents>\n\n"
             )
@@ -427,27 +405,11 @@ def generate_anthropic_prompt(
         else:
             tree = ""
 
-        if include_messages:
-            sys_prompt = (
-                self.memory_manager.identity
-                + conversation_history
-                + tree
-                + file_context
-            )
 
-        else:
-            sys_prompt = self.memory_manager.identity + "\n\n" + tree + file_context
-
-        if sys_only:
-            return sys_prompt
-
-        return (
-            "\n\nHuman: The folllowing is your system prompt: "
-            + sys_prompt
-            + "\n\nAssistant: Understood\n\n"
-            + last_user_message
-            + "\n\nAssistant:"
-        )
+        sys_prompt = tree + file_context + self.memory_manager.identity
+
+        
+        return sys_prompt
         # return sys_prompt + "\n\n" + last_user_message + "\n\nAssistant: "
 
     @staticmethod

backend/main.py

@@ -62,7 +62,7 @@ async def system_prompt():
         return {"system_prompt": AGENT.memory_manager.prompt_handler.system}
     elif AGENT.GPT_MODEL == "anthropic":
         return {
-            "system_prompt": AGENT.generate_anthropic_prompt(include_messages=False)
+            "system_prompt": AGENT.generate_anthropic_prompt()
         }
 
 

backend/memory/memory_manager.py

@@ -82,7 +82,7 @@ def get_context(self) -> str:
         for result in results:
             self.context += "\n" + result[0]
 
-        return self.context
+        return ""
 
     def remove_context(self, context: str) -> None:
         self.context = self.context.replace(context, "")
@@ -231,7 +231,7 @@ def get_messages(self, chat_box: Optional[bool] = None) -> List[dict]:
         results = self.cur.fetchall()
         prev_role = "assistant"
         for result in results[::-1]:
-            if prev_role == result[0]:
+            if prev_role == result[0] or result[2] is None or result[2] == "":
                 continue
             messages.append(
                 {"role": result[0], "content": result[2], "full_content": result[1]}

backend/memory/system_prompt_handler.py

@@ -118,8 +118,8 @@ def get_file_contents(self) -> Dict[str, str]:
             out.update({os.path.relpath(file_name, self.directory): text})
         return out
 
-    def set_files_in_prompt(self, include_line_numbers: Optional[bool] = None) -> None:
-        """
+    def set_files_in_prompt(self, anth: Optional[bool] = False, include_line_numbers: Optional[bool] = None) -> None:
+        """s
         Sets the files in the prompt.
 
         Args:
@@ -131,11 +131,12 @@ def set_files_in_prompt(self, include_line_numbers: Optional[bool] = None) -> No
         for k, v in file_contents.items():
             if k in self.files_in_prompt and include_line_numbers:
                 v = self._add_line_numbers_to_content(v)
-                content += f"{k}:\n{v}\n\n"
+                content += f"<{k}>\n{v}\n</{k}>\n\n" if anth else f"{k}:\n{v}\n\n"
             elif k in self.files_in_prompt:
-                content += f"{k}:\n{v}\n\n"
+                content += f"<{k}>\n{v}\n</{k}>\n\n" if anth else f"{k}:\n{v}\n\n"
 
         self.system_file_contents = content
+        print(f"File Contents: {k}" for k, v in file_contents)
         self.set_system()
         return