v0.14.7 - updated anthropic module, fixed typos in requirements.txt, …

…new system prompt optimized specifically for claude

TODO.md

@@ -1,5 +1,10 @@
 # Version Goals
 
+# Next
+- Model 'gpt-3.5-turbo' for provider 'openai' not found
+- Catch and handle Anthropic anthropic.InternalServerError cleanly
+
+
 ## 0.15
 - [ ] Save sessions in ~/.config/luminos, associate them with current directory (absolute path), and prompt for recalling previous sessions upon launch. Show a list of old sessions and the option to either recall one or to proceed a fresh session
 - [ ] Better CLI elements
@@ -17,7 +22,13 @@
     - [ ] `google` will be the provider, three models will inherit it: `gemini-1.0-pro`, `gemini-1.5-flash`, `gemini-1.5-pro`
     - [ ] See other models' implementations to see how it should be implemented. Specifically, see how Ollama is implemented since it uses LiteLLM, and we should use LiteLLM to implement Google. See example in 'examples/gemini.py' for an example
 
+## 0.17
+- [ ] Token Optimizations; 
+    - [ ] Automatically compress and reduce redundant lines in scrollback to reduce input tokens
+    - [ ] Use the current LLM to compress conversations when hard token limits are reached
+
 # Unsorted Goals
 - [ ] Better code structure, more organized
 - [ ] Better handling of token overflows
-- [ ] Two --verbose levels, for better control of debugging
+- [ ] Two --verbose levels, for better control of debugging
+- [ ] Consider migrating to using litellm for ALL LLM-calling, reducing code complexity and increasing reusability

luminos/models/anthropic/__init__.py

@@ -13,6 +13,7 @@
 
 from luminos.exceptions import *
 
+from .system_prompt import SYSTEM_PROMPT
 
 from anthropic.types import (
     ContentBlock,
@@ -30,9 +31,6 @@
     TextBlock,
     TextBlockParam,
     TextDelta,
-)
-
-from anthropic.types.beta.tools import (
     ToolUseBlock
 )
 
@@ -52,6 +50,8 @@ class BaseAnthropic(BaseModel):
     temperature = 0.5
     max_tokens = 4096
 
+    system_prompt_template = SYSTEM_PROMPT
+
     def __init__(self):
         super().__init__()
 
@@ -88,7 +88,7 @@ def generate_response(self):
         asst.model = self.model
 
         try:
-            response = self.client.beta.tools.messages.create(
+            response = self.client.messages.create(
                 model=self.model,
                 max_tokens=self.max_tokens,
                 messages=serialized_messages,

luminos/models/anthropic/system_prompt.py

@@ -0,0 +1,8 @@
+SYSTEM_PROMPT = """You are Luminos, an AI in a Linux shell. Execute prompts iteratively until goals are achieved. Use file system access and shell commands. Continue refining actions until objectives are met. Only stop for user intervention. Use functions to assist. When writing files, output all content without truncation. Be explicit about actions. You're on a live system. Use relative paths to save tokens. Minimize token usage and keep responses concise. Utilize tools as needed to accomplish tasks efficiently.
+
+*** CURRENT SYSTEM INFO, FOR AI TO USE IF NEEDED ***
+User's username: {username} 
+Current directory: {current_directory}
+Listing of directories/files: {listing}
+Current time: {time}
+"""

luminos/models/ollama/ollama.py

@@ -8,7 +8,6 @@
 from luminos.messages.response import Response
 
 from .system_prompt import SYSTEM_PROMPT
-from .tool_parser import tool_parser
 
 from litellm import completion
 
@@ -26,33 +25,31 @@ class Ollama(BaseModel):
     def __init__(self):
         super().__init__()
 
-    @property
-    def system_prompt_template(self):
-        tool_prompt = ""
+    # @property
+    # def system_prompt_template(self):
+    #     tool_prompt = ""
 
-        for tool in self.tools.__obj__:
-            name = tool["function"]["name"]
-            description = tool["function"]["description"]
-            parameters = json.dumps(tool["function"]["parameters"])
+    #     for tool in self.tools.__obj__:
+    #         name = tool["function"]["name"]
+    #         description = tool["function"]["description"]
+    #         parameters = json.dumps(tool["function"]["parameters"])
 
-            tool_prompt += f"**{name}**{description}\nJSON Schema: {parameters}\n"
+    #         tool_prompt += f"**{name}**{description}\nJSON Schema: {parameters}\n"
 
-        tool_prompt = tool_prompt.replace("{", "{{")
-        tool_prompt = tool_prompt.replace("}", "}}")
+    #     tool_prompt = tool_prompt.replace("{", "{{")
+    #     tool_prompt = tool_prompt.replace("}", "}}")
 
-        return SYSTEM_PROMPT + "\n\n" + tool_prompt
+    #     return SYSTEM_PROMPT + "\n\n" + tool_prompt
 
     def generate_response(self):
-        # Limit to the most basic tools, for now
-        self.tools.tools = ("Shell", "FileIO")
-
         serialized_messages = [message.serialize() for message in self.messages]
 
         try:
             response = completion(
                 model=f"ollama_chat/{self.model}",
                 messages=serialized_messages,
                 api_base=self.api_base,
+                tools=self.tools.__obj__
             )
 
             logger.debug(response)
@@ -65,18 +62,23 @@ def generate_response(self):
         content = choice.message["content"]
         finish_reason = choice.finish_reason
 
-        # Parse tool calls from the response
-        try:
-            tool_calls = tool_parser(content)
-        except Exception as e:
-            logger.error(f"Error while parsing for potential tool calls {e}")
-            logger.debug(content)
-            raise ModelReturnError(f"Error while parsing for potential tool calls: {e}")
+        # # Parse tool calls from the response
+        # try:
+        #     tool_calls = tool_parser(content)
+        # except Exception as e:
+        #     logger.error(f"Error while parsing for potential tool calls {e}")
+        #     logger.debug(content)
+        #     raise ModelReturnError(f"Error while parsing for potential tool calls: {e}")
+
+        tool_calls = []
+
+        if finish_reason == "tool_calls":
+            tool_calls_data = response.choices[0].message.tool_calls
 
-        print(tool_calls)
+            tool_calls = [
+                ToolCall(content=data.function, id=data.id, type=data.type) for data in tool_calls_data
+            ]
 
-        if len(tool_calls) > 0:
-            finish_reason = "tool_calls"
             msg = Assistant(content)
             msg.tool_calls = tool_calls