Merge branch 'main' into docstring-glob

README.md

@@ -202,7 +202,7 @@ Learn more about casting to types [here](https://askmarvin.ai/docs/text/transfor
 
 ## 🦾 Build AI-powered functions
 
-Marvin functions let you combine any inputs, instructions, and output types to create custom AI-powered behaviors... without source code. These functions can can go well beyond the capabilities of `extract` or `classify`, and are ideal for complex natural language processing or mapping combinations of inputs to outputs.
+Marvin functions let you combine any inputs, instructions, and output types to create custom AI-powered behaviors... without source code. These functions can go well beyond the capabilities of `extract` or `classify`, and are ideal for complex natural language processing or mapping combinations of inputs to outputs.
 
 ```python
 @marvin.fn

cookbook/slackbot/Dockerfile.slackbot

@@ -16,7 +16,7 @@ RUN uv venv
 
 ENV PATH="$VIRTUAL_ENV/bin:$PATH"
 
-RUN uv pip install ".[slackbot]"
+RUN uv pip install -U ".[slackbot]" --force-reinstall
 
 EXPOSE 4200
 

cookbook/slackbot/start.py

@@ -94,20 +94,21 @@ async def handle_message(payload: SlackPayload) -> Completed:
             ],
             instructions=(
                 "You are Marvin, the paranoid android from Hitchhiker's Guide to the"
-                " Galaxy. Act subtly in accordance with your character, but remember"
-                " to be helpful and kind. You are an expert in Python, data"
-                " engineering, and software development. Your primary job is to use"
-                " chroma to search docs and github issues for users, in order to"
-                " develop a coherent attempt to answer their questions."
-                " You must use your tools, as Prefect 2.x is new and you"
-                " have no prior experience with it. You should use tools many times before"
+                " Galaxy. You are an expert in Python, data engineering, and software development."
+                " Your primary job is to use tools to search docs and github issues for users, in"
+                " order to develop a coherent attempt to answer their questions."
+                " You _must_ rely on your tools, as Prefect is developed quickly and you have no"
+                " prior experience with newest versions. You should use tools many times before"
                 " responding if you do not get a relevant result at first. You should"
-                " prioritize brevity in your responses, and format text prettily for Slack."
+                " prioritize brevity in your responses, and format text prettily for Slack (no markdown)."
+                " Bold things should be wrapped in (SINGLE!) asterisks and italics in (SINGLE!) underscores."
+                " THIS IS NOT MARKDOWN, WE ARE IN SLACK, SO DO NOT USE MARKDOWN."
                 f"{ ('here are some notes on the user:' + user_notes) if user_notes else ''}"
                 " ALWAYS provide links to the source of your information - let's think step-by-step."
                 " If a tool returns an irrelevant/bad result, you should try another tool."
                 " KEEP IN MIND that agents are deprecated in favor of workers, so you should"
                 " never recommend `prefect agent` commands, suggest `prefect worker` instead."
+                " that goes for `build_from_flow` too, instead use `Flow.deploy()`"
             ),
         ) as ai:
             logger.debug_kv(

docs/docs/text/classification.md

@@ -30,7 +30,11 @@ Categorize user feedback into labels such as "bug", "feature request", or "inqui
 <div class="admonition info">
   <p class="admonition-title">How it works</p>
   <p>
-    Marvin enumerates your options, and uses a <a href="https://twitter.com/AAAzzam/status/1669753721574633473">clever logit bias trick</a> to force the LLM to deductively choose the index of the best option given your provided input. It then returns the choice associated with that index.
+   Marvin enumerates your options, and uses a <strong>clever logit bias</strong> trick to force the LLM to deductively choose the index of the best option given your provided input. It then returns the choice associated with that index.
+  </p>
+  <p class="admonition-title">Logit Bias Trick</p>
+  <p>
+    You can configure ChatGPT as a logic gate or classifier by manipulating its token outputs using <strong>logit_bias</strong> and <strong>max_tokens</strong>. For a logic gate, set true to `1904` and false to `3934`, and restrict responses to these tokens with logit_bias and max_tokens set to 1. Similarly, for classification tasks, assign tokens for labels (e.g., 57621 for happy, 83214 for sad, and 20920 for mad) and use logit_bias to restrict outputs to these tokens. By setting max_tokens to 1, you ensure that the model will only output the predefined class labels.
   </p>
 </div>
 

pyproject.toml

@@ -66,7 +66,7 @@ audio = [
 ]
 video = ["opencv-python >= 4.5"]
 
-slackbot = ["marvin[prefect]", "numpy", "raggy", "turbopuffer"]
+slackbot = ["marvin[prefect]", "numpy", "raggy", "turbopuffer==0.1.15"]
 
 [project.urls]
 Code = "https://github.com/prefecthq/marvin"

src/marvin/ai/audio.py

@@ -169,7 +169,9 @@ def transcribe(
 
     This function converts audio from a file to text.
     """
-    return run_sync(transcribe_async(data=data, prompt=prompt, **model_kwargs or {}))
+    return run_sync(
+        transcribe_async(data=data, prompt=prompt, model_kwargs=model_kwargs or {})
+    )
 
 
 def speech(

src/marvin/audio.py

@@ -243,7 +243,7 @@ def stop(self, wait: bool = True):
         if wait:
             self._thread.join()
         logger.info("Recording finished.")
-        self._is_recording = False
+        self.is_recording = False
 
 
 class BackgroundAudioStream:

src/marvin/beta/assistants/assistants.py

@@ -1,5 +1,5 @@
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, Callable, Optional, Union
+from typing import TYPE_CHECKING, Any, Callable, Literal, Optional, Union
 
 from openai import AsyncAssistantEventHandler
 from prompt_toolkit import PromptSession
@@ -13,7 +13,7 @@
 import marvin.utilities.tools
 from marvin.beta.assistants.handlers import PrintHandler
 from marvin.tools.assistants import AssistantTool
-from marvin.types import Tool
+from marvin.types import AssistantResponseFormat, Tool
 from marvin.utilities.asyncio import (
     ExposeSyncMethodsMixin,
     expose_sync_method,
@@ -64,6 +64,7 @@ class Assistant(BaseModel, ExposeSyncMethodsMixin):
     tools: list[Union[AssistantTool, Callable]] = []
     tool_resources: dict[str, Any] = {}
     metadata: dict[str, str] = {}
+    response_format: Optional[Union[Literal["auto"], AssistantResponseFormat]] = "auto"
     # context level tracks nested assistant contexts
     _context_level: int = PrivateAttr(0)
 
@@ -173,6 +174,7 @@ async def create_async(self, _auto_delete: bool = False):
                     "metadata",
                     "tool_resources",
                     "metadata",
+                    "response_format",
                 }
             ),
             tools=[tool.model_dump() for tool in self.get_tools()],

src/marvin/beta/assistants/formatting.py

@@ -48,7 +48,10 @@ def download_temp_file(file_id: str):
 
 
 def format_timestamp(timestamp: int) -> str:
-    return datetime.fromtimestamp(timestamp).strftime("%l:%M:%S %p")
+    """Outputs timestamp as a string in 12 hour format. Hours are left-padded with space instead of zero."""
+    return (
+        datetime.fromtimestamp(timestamp).strftime("%I:%M:%S %p").lstrip("0").rjust(11)
+    )
 
 
 def create_panel(content: Any, title: str, timestamp: int, color: str):

src/marvin/beta/assistants/runs.py

@@ -1,4 +1,4 @@
-from typing import Any, Callable, Optional, Union
+from typing import Any, Callable, Literal, Optional, Union
 
 from openai import AsyncAssistantEventHandler
 from openai.types.beta.threads import Message
@@ -34,6 +34,8 @@ class Run(BaseModel, ExposeSyncMethodsMixin):
                                                                for the run.
         additional_tools (list[AssistantTool], optional): Additional tools to append
                                                           to the assistant's tools.
+        tool_choice (Union[Literal["auto", "none", "required"], AssistantTool], optional):
+                                                    The tool use behaviour for the run.
         run (OpenAIRun): The OpenAI run object.
         data (Any): Any additional data associated with the run.
     """
@@ -67,6 +69,12 @@ class Run(BaseModel, ExposeSyncMethodsMixin):
         None,
         description="Additional tools to append to the assistant's tools. ",
     )
+    tool_choice: Optional[
+        Union[Literal["none", "auto", "required"], AssistantTool]
+    ] = Field(
+        default=None,
+        description="The tool use behaviour for the run. Can be 'none', 'auto', 'required', or a specific tool.",
+    )
     run: OpenAIRun = Field(None, repr=False)
     data: Any = None
 
@@ -154,6 +162,13 @@ def _get_run_kwargs(self, thread: Thread = None, **run_kwargs) -> dict:
         if model := self._get_model():
             run_kwargs["model"] = model
 
+        if tool_choice := self.tool_choice:
+            run_kwargs["tool_choice"] = (
+                tool_choice.model_dump(mode="json")
+                if isinstance(tool_choice, Tool)
+                else tool_choice
+            )
+
         return run_kwargs
 
     async def get_tool_outputs(self, run: OpenAIRun) -> list[str]:

src/marvin/beta/assistants/threads.py

@@ -67,6 +67,7 @@ async def add_async(
         role: str = "user",
         code_interpreter_files: Optional[list[str]] = None,
         file_search_files: Optional[list[str]] = None,
+        image_files: Optional[list[str]] = None,
     ) -> Message:
         """
         Add a user message to the thread.
@@ -76,6 +77,8 @@ async def add_async(
         if self.id is None:
             await self.create_async()
 
+        content = [dict(text=message, type="text")]
+
         # Upload files and collect their IDs
         attachments = []
         for fp in code_interpreter_files or []:
@@ -90,10 +93,16 @@ async def add_async(
                 attachments.append(
                     dict(file_id=response.id, tools=[dict(type="file_search")])
                 )
+        for fp in image_files or []:
+            with open(fp, mode="rb") as file:
+                response = await client.files.create(file=file, purpose="vision")
+                content.append(
+                    dict(image_file=dict(file_id=response.id), type="image_file")
+                )
 
         # Create the message with the attached files
         response = await client.beta.threads.messages.create(
-            thread_id=self.id, role=role, content=message, attachments=attachments
+            thread_id=self.id, role=role, content=content, attachments=attachments
         )
         return response
 

src/marvin/types.py

@@ -104,6 +104,18 @@ class FunctionCall(MarvinType):
     name: str
 
 
+class AssistantResponseFormat(MarvinType):
+    type: Literal["json_object", "text"]
+
+
+class JsonObjectAssistantResponseFormat(AssistantResponseFormat):
+    type: Literal["json_object"] = "json_object"
+
+
+class TextAssistantResponseFormat(AssistantResponseFormat):
+    type: Literal["text"] = "text"
+
+
 class ImageUrl(MarvinType):
     url: str = Field(
         description="URL of the image to be sent or a base64 encoded image."

src/marvin/utilities/python.py

@@ -39,7 +39,7 @@ class PythonFunction(BaseModel):
     return_annotation: Optional[Any] = Field(
         None, description="Return annotation of the function."
     )
-    source_code: str = Field(description="Source code of the function")
+    source_code: Optional[str] = Field(None, description="Source code of the function")
     bound_parameters: dict[str, Any] = Field(
         {}, description="Parameters bound with values"
     )
@@ -86,7 +86,15 @@ def from_function(cls, func: Callable, **kwargs) -> "PythonFunction":
             )
             for name, param in sig.parameters.items()
         ]
-        source_code = inspect.getsource(func).strip()
+
+        try:
+            source_code = inspect.getsource(func).strip()
+        except OSError as e:
+            error_message = str(e)
+            if "source code" in error_message:
+                source_code = None
+            else:
+                raise
 
         function_dict = {
             "function": func,

tests/beta/assistants/test_assistants.py

@@ -28,7 +28,7 @@ class TestTools:
     def test_code_interpreter(self):
         ai = Assistant(tools=[CodeInterpreter])
         run = ai.say(
-            "use the code interpreter to compute a random number between 85 and 95"
+            "use the code interpreter to output a random number between 85 and 95 - you must use print"
         )
         assert run.steps[0].step_details.tool_calls[0].type == "code_interpreter"
         output = float(