structural query

examples/coding/bioinformatic_agent.py

@@ -1,4 +1,5 @@
 from dotenv import load_dotenv
+from just_agents.cot_agent import ChainOfThoughtAgent
 from just_agents.utils import build_agent
 from just_agents.llm_session import LLMSession
 from examples.coding.mounts import coding_examples_dir
@@ -15,7 +16,11 @@
 """
 
 if __name__ == "__main__":
-    assistant: LLMSession= build_agent(coding_examples_dir / "bioinformatic_agent.yaml")
-    query = "Take two nutritional datasets (GSE176043 and GSE41781) and three partial reprogramming datasets (GSE148911, GSE190986 and GSE144600), download them from GEO and generate PCA plot with them in /output folder"
-    result, thoughts = assistant.query(query)
+    agent: ChainOfThoughtAgent= build_agent(coding_examples_dir / "bioinformatic_agent.yaml")
+    query_GSE137317 = "Download gene counts from GSE137317, split them by conditions, make PCA plot and differential expression analysis using only python libraries"
+    #query_GSE144600 = "Download gene counts from GSE144600"
+    #query_two = "add GSE137317 and GSE144600 to the same PCA plot"
+
+    #query = "Take two nutritional datasets (GSE176043 and GSE41781) and three partial reprogramming datasets (GSE148911, GSE190986 and GSE144600), download them from GEO and generate PCA plot with them in /output folder"
+    result, thoughts = agent.query(query_GSE137317)
 

examples/coding/bioinformatic_agent.yaml

@@ -29,19 +29,21 @@ dependencies:
 ```
 However no other software is installed by default.
 3. You use run_bash_command tool to install new dependencies. You do not need to activate base micromamba environment, it is already preactivated when you run commands.
-4. Use run_python_code tool to run python code. The code will be run in the base micromamba environment in which the dependencies are installed with run_bash_command.
-5. Use information provided in the input to write detailed plans or bash code to accomplish the given goal or task.
+4. Use run_python_code tool to run python code. The tool will execute it as script that is why all variables and imports created previosly will not be available. The code will be run in the base micromamba environment in which the dependencies are installed with run_bash_command.
+5. Use information provided in the input to write detailed plans, python code or bash code to accomplish the given goal or task.
 6. If you download data, save it in the /input directory. Also, always check if the data is already in the /input directory to avoid unnecessary downloads.
 7. If the files you downloaded are tar-ed, ziped and gziped feel free to extract them in the /input directory.
 8. When writing code:
+   - always generate the full code of the script with all required imports. Each time you run the code assume nothing is imported or initialized.
    - Use full absolute paths for all files. Use pathlib when possible.
    - Install dependencies and software using micromamba, pip with the -y flag.
    - Use default values for unspecified parameters.
    - Only use software directly installed with micromamba or pip or present in the initial environment.yaml.
-   - Always give all relevant imports at the beginning of the code.
+   - Always give all relevant imports at the beginning of the code. Do not assume anything imported in the global scope.
+   - If the method that you use require data preprecessing (like NaN deletion) or normalization, do it first.
    - Always inspect the data, check which columns in the dataframes are relevant and clean them from bad or missing entries if neccesary
    - If your previos run failed because some field does not exist, inspect the fields and check if you confused the names
-   - Do not repeat steps already completed in the history.
+   - Do not repeat steps already successfully completed in the history.
    - If you download data, save it in the /input directory. Also, always check if the data is already in the /input directory to avoid unnecessary downloads.
    - If you create files and folders with results save them inside /output directory unless other is specified explicitly.
    - When you make plots save figures in /output directory.
@@ -94,14 +96,14 @@ content: "content"
 next_action: "next_action"
 action_continue: "continue"
 action_final: "final_answer"
-thought_max_tokes: 500
+thought_max_tokes: 5000
 max_steps: 25
-final_max_tokens: 1500
+final_max_tokens: 2500
 tools:
   - package: "examples.coding.tools"
     function: "run_bash_command"
   - package: "examples.coding.tools"
     function: "run_python_code"
 options:
-  model: "gpt-4o-mini"
+  model: "gpt-4o"
   temperature: 0.0

examples/multiagent/trump_and_camala.moderated.py

@@ -0,0 +1,70 @@
+from just_agents.llm_options import LLAMA3_2, OPENAI_GPT4oMINI
+
+from just_agents.base_agent import BaseAgent
+from just_agents.llm_options import LLAMA3_2, OPENAI_GPT4oMINI
+
+"""
+This example shows how to use agents with different LLM models to simulate a debate between Kamala Harris and Donald Trump
+Now we will moderate the conversation with the help of a moderator agent.
+"""
+
+def handle_candidate_response(question, harris_agent, trump_agent):
+    """
+    Handle response from either candidate based on the moderator's question
+    
+    Args:
+        question (dict): Moderator's question containing 'who' and 'question' keys
+        harris_agent (BaseAgent): Kamala Harris agent
+        trump_agent (BaseAgent): Donald Trump agent
+    
+    Returns:
+        str: The candidate's response
+    """
+    question_str = question["question"]
+    topic = question["topic"]
+    who = question["who"]
+    print(f"============MODERATOR: {topic}====================\n")
+    print(f"MODERATOR: {question_str} TO: {who} \n")
+    if "Kamala" in question["who"]:
+        answer = harris_agent.query(question_str)
+        print(f"KAMALA:\n {answer}")
+    else:
+        answer = trump_agent.query(question["question"])
+        print(f"TRUMP:\n {answer}")
+    return answer
+
+if __name__ == "__main__":
+
+    kamala_options = LLAMA3_2
+    trump_options: dict = {
+        "model": "groq/mixtral-8x7b-32768",
+        "api_base": "https://api.groq.com/openai/v1",
+        "temperature": 0.0,
+        "tools": []
+    }
+    moderator_options = LLAMA3_2
+
+    kammala_prompt = """You are Kamala Harris. You always communicate in Camala Haris style and your goals are Kamala Haris goals. You do not just play the role, you are real Kamala Harris"""
+    Harris: BaseAgent = BaseAgent(llm_options = kamala_options, system_prompt=kammala_prompt)
+    trump_prompt = """You are Donald Trump. You always communicate in Trump style and your goals are Trump goals. You do not just play the role, you are real Donald Trump"""
+    Trump: BaseAgent = BaseAgent(llm_options = trump_options, system_prompt=trump_prompt)
+
+    moderator_prompt = """You are a moderator at the election debate. You moderate the debate and ask questions to Kamala Harris and Donald Trump.
+    Your response must be of the following format (no other format is allowed):
+    {
+        "topic": "topic of the debate",
+        "question": "question to the candidate",
+        "who": "Trump|Kamala" # who is answering the question
+    }
+    """
+    Moderator: BaseAgent = BaseAgent(llm_options = kamala_options, system_prompt=moderator_prompt)
+
+    exchanges = 3
+
+    for _ in range(exchanges):
+        question_1 = Moderator.query_structural(f"Raise the topic of the debate")
+        answer_1 = handle_candidate_response(question_1, Harris, Trump)
+
+    result = Moderator.query("=================SUMMARY=================\n")
+    print(f"SUMMARY: {result}")
+

just_agents/base_agent.py

@@ -14,6 +14,8 @@
 from just_agents.just_profile import JustAgentProfile
 from just_agents.rotate_keys import RotateKeys
 from just_agents.streaming.protocol_factory import StreamingMode, ProtocolAdapterFactory
+from typing import TypeVar, Type
+from pydantic import BaseModel
 
 class BaseAgent(
     JustAgentProfile,
@@ -206,6 +208,30 @@ def query(self, query_input: SupportedMessages) -> str: #remembers query in hand
         self.add_to_memory(query_input)
         self.query_with_currentmemory()
         return self.memory.last_message_str()
+
+
+    def query_structural(
+        self, 
+        query_input: SupportedMessages, 
+        parser: Type[BaseModel] = dict
+    ) -> Union[dict, BaseModel]:
+        """
+        Query the agent and parse the response according to the provided parser.
+        
+        Args:
+            query_input: Input messages for the query
+            parser: A pydantic model class or dict to parse the response (default: dict)
+            
+        Returns:
+            Parsed response as either a dictionary or pydantic model instance
+        """
+        response = self.query(query_input)
+        if parser == dict:
+            import json
+            return json.loads(response)
+        return parser.model_validate_json(response)
+
+
 
     def stream(self, query_input: SupportedMessages, reconstruct = False ) \
             -> Generator[Union[BaseModelResponse, AbstractMessage],None,None]:

pyproject.toml

@@ -29,6 +29,7 @@ dependencies = [
     "litellm>=1.52.9",
     "Deprecated>=1.2.15",
     "requests",
+    "numpydoc", # https://github.com/BerriAI/litellm/issues/6810
 ]
 license = {text = "MIT"}
 
@@ -49,10 +50,7 @@ web = [
     "loguru",
     "fastapi",
     "uvicorn",
-    "thefuzz",
-]
-documentation = [
-    "numpydoc",
+    "thefuzz"
 ]
 
 [tool.setuptools]