Fix codegen

.gitignore

@@ -168,4 +168,5 @@ scripts
 # Polymind learned facts and tools
 knowledge/facts/**
 knowledge/tools/**
-use_cases
+use_cases
+polymind/example_tools/**

polymind/core/task.py

@@ -163,7 +163,7 @@ class AtomTask(BaseTask):
         {input}
         ---
 
-        Please put the result into the ```json blob```.
+        Please put the result into the ```json blob```, and the key should be "output".
     """
 
     def __init__(self, tool_manager: ToolManager, tool_retriever: RetrieveTool, **kwargs):
@@ -186,7 +186,7 @@ async def _use_tool(self, objective: str, tool_description: str) -> Message:
             tool_description (str): The description of the tool.
         """
         # Retrieve the tool using ToolRetriever.
-        tool_retrieve_message = Message(content={self.tool_retrieve_query_key: tool_description, "top_k": 1})
+        tool_retrieve_message = Message(content={self.tool_retrieve_query_key: tool_description, "top_k": 3})
         tool_retrieve_result_message = await self._tool_retriever(tool_retrieve_message)
         self._logger.debug(f"Tool retrieve result: {tool_retrieve_result_message.content}")
         tool_name = tool_retrieve_result_message.content["results"][0]
@@ -263,7 +263,7 @@ async def _execute(self, input: Message) -> Message:
             "input"
         ] = f"""
             Context: {self.task_context}
-            Input from the previous step:
+            Input from the previous steps:
             {input_field}
             Objective: {self.task_name}
         """

polymind/core_tools/llm_tool.py

@@ -6,12 +6,14 @@
 import os
 from typing import List
 
+import anthropic
 from openai import AsyncOpenAI
 from pydantic import Field
 
 from polymind.core.logger import Logger
 from polymind.core.message import Message
-from polymind.core.tool import BaseTool, Embedder, LLMTool, Param
+from polymind.core.tool import (BaseTool, CodeGenerationTool, Embedder,
+                                LLMTool, Param)
 from polymind.core_tools.rest_api_tool import RestAPITool
 
 
@@ -220,3 +222,122 @@ async def _embedding(self, input: List[str]) -> List[List[float]]:
         embedding_list = [entry.get("embedding", []) for entry in response.get("data", [])]
         embeddings: List[List[float]] = [embedding[: self.embed_dim] for embedding in embedding_list]
         return embeddings
+
+
+class OpenAICodeGenerationTool(CodeGenerationTool):
+    """Use OpenAI to generate code snippets based on the input prompt."""
+
+    tool_name: str = "open-ai-code-generation"
+
+    def _set_llm_client(self):
+        model_name = os.environ.get("MODEL_NAME", "gpt-3.5-turbo")
+        self._llm_tool = OpenAIChatTool(model_name=model_name)
+        # self._llm_tool = AnthropicClaudeTool()
+
+
+class AnthropicClaudeTool(LLMTool):
+    """AnthropicClaudeTool is a bridge to Anthropic's Claude API.
+    The tool can be initialized with system_prompt, max_tokens, and temperature.
+    The input message of this tool should contain a "prompt", and optionally a "system_prompt".
+    The "system_prompt" in the input message will override the default system_prompt.
+    The tool will return a message with key "answer" with the response from the Claude API.
+    """
+
+    model_config = {
+        "arbitrary_types_allowed": True,  # Allow arbitrary types
+    }
+    tool_name: str = "anthropic-claude"
+    model_name: str = Field(default="claude-3-opus-20240229", description="The name of the Claude model.")
+    descriptions: List[str] = [
+        "This tool is used to chat with Anthropic's Claude language model.",
+        "This tool can be used as the orchestrator to control the conversation and problem solving.",
+        "This tool can be used to breakdown the problem into smaller parts and solve them.",
+        "This tool can be used to generate the response from the chat.",
+        "This tool can be used to generate the code of new tools.",
+        "This tool can do simple calculation.",
+        "Simple calculator that does basic arithmetic calculation.",
+    ]
+
+    client: anthropic.Client = Field(default=None)
+    system_prompt: str = Field(default="You are a helpful AI assistant.")
+    max_tokens: int = Field(default=4000)
+    temperature: float = Field(default=0.7)
+    stop: str = Field(default=None)
+    response_format: str = Field(default="text", description="The format of the response from the chat.")
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self._logger = Logger(__file__)
+        self._set_client()
+
+    def _set_client(self):
+        """Set the client for the language model."""
+        self.client = anthropic.Anthropic()
+
+    def input_spec(self) -> List[Param]:
+        """Return the input specification of the tool."""
+        return [
+            Param(
+                name="system_prompt",
+                type="str",
+                required=False,
+                example="You are a helpful AI assistant.",
+                description="The system prompt for the chat.",
+            ),
+            Param(
+                name="input",
+                type="str",
+                required=True,
+                example="hello, how are you?",
+                description="The prompt for the chat.",
+            ),
+            Param(
+                name="max_tokens",
+                type="int",
+                required=False,
+                example="1500",
+                description="The maximum number of tokens for the chat.",
+            ),
+            Param(
+                name="temperature",
+                type="float",
+                required=False,
+                example="0.7",
+                description="The temperature for the chat.",
+            ),
+        ]
+
+    def output_spec(self) -> List[Param]:
+        """Return the output specification of the tool."""
+        return [
+            Param(
+                name="output",
+                type="str",
+                required=True,
+                example="I'm good, how are you?",
+                description="The response from the chat.",
+            ),
+        ]
+
+    async def _invoke(self, input: Message) -> Message:
+        """Execute the tool and return the result."""
+        prompt = input.get("input", "")
+        system_prompt = input.get("system_prompt", self.system_prompt)
+        prompt = f"{system_prompt}\n{prompt}"
+        temperature = input.get("temperature", self.temperature)
+        max_tokens = input.get("max_tokens", self.max_tokens)
+        messages = [
+            {"role": "user", "content": prompt},
+        ]
+        response = self.client.messages.create(
+            model=self.model_name,
+            max_tokens=max_tokens,
+            temperature=temperature,
+            messages=messages,
+        )
+        content = response.content[0].text
+        self._logger.tool_log(f"[{self.tool_name}], System Prompt: [{system_prompt}]")
+        self._logger.tool_log(f"[{self.tool_name}], Prompt: [{prompt}]")
+        self._logger.tool_log(f"[{self.tool_name}], Response from Claude: [{content}]")
+        response_message = Message(content={"output": content})
+        return response_message

polymind/core_tools/retrieve_tool.py

@@ -163,6 +163,7 @@ def _set_client(self):
             self._client.create_collection(self.collection_name, dimension=self.embed_dim, auto_id=True)
         self.embedder = OpenAIEmbeddingTool(embed_dim=self.embed_dim)
         self._llm_tool: LLMTool = OpenAIChatTool()
+        self.top_k = self.top_k or 3
         self._logger = Logger(__file__)
 
     def _extra_input_spec(self) -> List[Param]:

polymind/thought_process/chain_of_tasks.py

@@ -31,15 +31,13 @@ class ChainOfTasks(ThoughtProcess):
 
     problem_decomposition_prompt: str = """
         Please read the requirement carefully, and think step-by-step before answering the question.
-        Please decompose the problem into 1-5 steps, depending on the complexity of the problem.
-
-        Each of the following sub-task will use the output of the previous task as input.
-
-        Please write down your decomposition into the json blob.
-        For each step, please give it an "objective", "input" and "output".
-        FOr the objective, please make it less ambiguous and more specific.
-        And make it to explain how to use the input.
-        For input and output, please use declarative name and please describe the type as well.
+        Follow the below rules:
+        1. Please decompose the problem into 1-5 steps, depending on the complexity of the problem.
+            Each of the following sub-task will use the output of the previous task as input.
+        2. For each step, please give it an "objective", "input" and "output".
+            Objectives: Make it less ambiguous and more specific to the requirement, e.g. including date if provided.
+            Input: Make it to explain how to use the input. Use declarative name and please describe the type as well.
+        3. Please write down the decomposition into the json blob.
 
         An example of the decomposition is as follows:
 

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "polymind"
-version = "0.0.42"  # Update this version before publishing to PyPI
+version = "0.0.43"  # Update this version before publishing to PyPI
 description = "PolyMind is a customizable collaborative multi-agent framework for collective intelligence and distributed problem solving."
 authors = ["TechTao"]
 license = "MIT License"
@@ -18,6 +18,7 @@ pymilvus = "2.3.7"
 faiss-cpu = "1.8.0"
 colorama = "^0.4.6"
 tavily-python = "^0.3.3"
+anthropic = "^0.25.6"
 
 [tool.poetry.group.dev.dependencies]
 black = "^24.2.0"

tests/polymind/core/test_tool.py

@@ -481,135 +481,3 @@ async def test_invoke_tool(self, monkeypatch):
         response = await manager.invoke_tool("test_tool", input=params)
         assert response.get("result") == "tset", "The result should be the reverse of the input query"
         assert response.get("result2") == "olleh", "The result should be the reverse of the input query2"
-
-
-class TestCodeGenerationTool:
-    @pytest.fixture
-    def llm_tool_mock(self):
-        return AsyncMock()
-
-    @pytest.fixture
-    def code_gen_tool(self, llm_tool_mock):
-        return CodeGenerationTool(llm_tool=llm_tool_mock)
-
-    # @pytest.mark.asyncio
-    # async def test_execute_successful(self, code_gen_tool, llm_tool_mock):
-    #     # Setup mock response
-    #     llm_tool_mock.return_value = AsyncMock(
-    #         content={
-    #             "output": """
-    #             ```python
-    #             output = {"result": 42}
-    #             ```
-    #             """
-    #         }
-    #     )
-    #     expected_code = 'output = {"result": 42}'
-    #     expected_output = '{"result": 42}'
-    #     input_message = Message(content={"input": "Sum two numbers"})
-
-    #     # Execute the test
-    #     result = await code_gen_tool(input_message)
-    #     actual_output = result.content["output"]
-
-    #     assert result.content["code"] == expected_code
-    # assert json.loads(actual_output) == json.loads(expected_output), actual_output
-
-    @pytest.mark.asyncio
-    async def test_execute_failure_max_attempts(self, code_gen_tool, llm_tool_mock):
-        # Simulate failures
-        llm_tool_mock.side_effect = [AsyncMock(side_effect=Exception("Error")) for _ in range(3)]
-        input_message = Message(content={"input": "Sum two numbers"})
-
-        with pytest.raises(Exception):
-            await code_gen_tool(input_message)
-
-    @pytest.mark.asyncio
-    async def test_code_gen_success(self, code_gen_tool, llm_tool_mock):
-        # Setup mock response
-        llm_tool_mock.return_value = AsyncMock(
-            content={
-                "output": """
-                ```python
-                a = 10
-                b = 32
-                def total(a, b):
-                    return a + b
-                output = {"result": total(a, b)}
-                ```
-                """
-            }
-        )
-        requirement = "Sum two numbers"
-        previous_errors = []
-
-        code = await code_gen_tool._code_gen(requirement=requirement, previous_errors=previous_errors)
-        expected_code = """
-            a = 10
-            b = 32
-            def total(a, b):
-                return a + b
-            output = {"result": total(a, b)}
-        """
-        expected_code = re.sub("\s+", "", expected_code)
-        assert re.sub("\s+", "", code) == expected_code
-
-    @pytest.mark.asyncio
-    async def test_code_gen_failure_no_code(self, code_gen_tool, llm_tool_mock):
-        # Setup mock response
-        llm_tool_mock.return_value = AsyncMock(content={"output": ""})
-        requirement = "Sum two numbers"
-        previous_errors = []
-
-        with pytest.raises(ValueError):
-            await code_gen_tool._code_gen(requirement, previous_errors)
-
-    def test_extract_required_packages(self, code_gen_tool):
-        code = textwrap.dedent(
-            """
-            import numpy
-            import pandas as pd
-            import matplotlib.pyplot as plt
-            import yfinance as yf
-            from sklearn.linear_model import LinearRegression
-            """
-        )
-        expected_packages = set(["numpy", "pandas", "matplotlib", "yfinance", "sklearn"])
-        packages = code_gen_tool._extract_required_packages(code)
-        assert set(packages) == expected_packages
-
-    @pytest.mark.asyncio
-    async def test_code_run_valid_python(self, code_gen_tool):
-        code = 'output = {"result": 100, "price": 200, "name": "test", "ids": [1, 2, 3]}'
-        result = await code_gen_tool._code_run(code)
-        assert result == {"result": 100, "price": 200, "name": "test", "ids": [1, 2, 3]}
-
-    @pytest.mark.asyncio
-    async def test_code_run_invalid_python(self, code_gen_tool):
-        code = "for i in range(10 print(i)"
-        with pytest.raises(Exception):
-            await code_gen_tool._code_run(code)
-
-    @pytest.mark.asyncio
-    async def test_output_parse_success(self, code_gen_tool, llm_tool_mock):
-        # Setup mock response
-        llm_tool_mock.return_value = AsyncMock(
-            content={"output": json.dumps({"status": "success", "output": '{"result": 42}'})}
-        )
-        requirement = "Sum two numbers"
-        output = '{"result": 42}'
-
-        parsed_output = await code_gen_tool._output_parse(requirement=requirement, output=output)
-        assert json.loads(parsed_output) == output, parsed_output
-
-    @pytest.mark.asyncio
-    async def test_output_parse_failure(self, code_gen_tool, llm_tool_mock):
-        # Setup mock response
-        llm_tool_mock.return_value = AsyncMock(
-            content={"output": json.dumps({"status": "failure", "reason": "Error: Invalid input"})}
-        )
-        requirement = "Sum two numbers"
-        output = {"result": 42}
-
-        with pytest.raises(ValueError):
-            await code_gen_tool._output_parse(requirement=requirement, output=output)