Baseten integration

.env.example

@@ -25,3 +25,7 @@ CARTESIA_API_KEY=your_cartesia_api_key_here
 
 # Anthropic API credentials
 ANTHROPIC_API_KEY=your_anthropic_api_key_here
+
+# Baseten API credentials
+BASETEN_API_KEY=your_baseten_api_key_here
+BASETEN_BASE_URL=your_baseten_base_url_here

agents-core/vision_agents/core/agents/agents.py

@@ -551,6 +551,8 @@ async def join(self, call: Call) -> "AgentSessionContextManager":
 
         # wait for conversation creation coro at the very end of the join flow
         self.conversation = await create_conversation_coro
+        # Provide conversation to the LLM so it can access the chat history.
+        self.llm.set_conversation(self.conversation)
         return AgentSessionContextManager(self, self._connection)
 
     async def finish(self):

agents-core/vision_agents/core/llm/llm.py

@@ -26,7 +26,7 @@
 from getstream.video.rtc.pb.stream.video.sfu.models.models_pb2 import Participant
 from getstream.video.rtc import AudioStreamTrack, PcmData
 from vision_agents.core.processors import Processor
-from vision_agents.core.utils.utils import parse_instructions
+from vision_agents.core.utils.utils import Instructions, parse_instructions
 from vision_agents.core.events.manager import EventManager
 from .function_registry import FunctionRegistry
 from .llm_types import ToolSchema, NormalizedToolCallItem
@@ -50,7 +50,6 @@ class LLM(abc.ABC):
     before_response_listener: BeforeCb
     after_response_listener: AfterCb
     agent: Optional["Agent"]
-    _conversation: Optional["Conversation"]
     function_registry: FunctionRegistry
 
     def __init__(self):
@@ -59,6 +58,9 @@ def __init__(self):
         self.events = EventManager()
         self.events.register_events_from_module(events)
         self.function_registry = FunctionRegistry()
+        self.instructions: Optional[str] = None
+        self.parsed_instructions: Optional[Instructions] = None
+        self._conversation: Optional[Conversation] = None
 
     async def warmup(self) -> None:
         """
@@ -187,9 +189,20 @@ def _attach_agent(self, agent: Agent):
         Attach agent to the llm
         """
         self.agent = agent
-        self._conversation = agent.conversation
         self._set_instructions(agent.instructions)
 
+    def set_conversation(self, conversation: Conversation):
+        """
+        Provide the Conversation object to the LLM to access the chat history.
+        To be called by the Agent after it joins the call.
+
+        Args:
+            conversation: a Conversation object
+
+        Returns:
+        """
+        self._conversation = conversation
+
     def _set_instructions(self, instructions: str):
         self.instructions = instructions
 

plugins/anthropic/tests/test_anthropic_llm.py

@@ -18,7 +18,7 @@ class TestClaudeLLM:
     async def llm(self) -> ClaudeLLM:
         """Test ClaudeLLM initialization with a provided client."""
         llm = ClaudeLLM(model="claude-sonnet-4-20250514")
-        llm._conversation = InMemoryConversation("be friendly", [])
+        llm.set_conversation(InMemoryConversation("be friendly", []))
         return llm
 
     @pytest.mark.asyncio
@@ -58,7 +58,7 @@ async def test_native_api(self, llm: ClaudeLLM):
     @pytest.mark.integration
     async def test_stream(self, llm: ClaudeLLM):
         streamingWorks = False
-        
+
         @llm.events.subscribe
         async def passed(event: LLMResponseChunkEvent):
             nonlocal streamingWorks
@@ -70,7 +70,6 @@ async def passed(event: LLMResponseChunkEvent):
 
         assert streamingWorks
 
-
     @pytest.mark.integration
     async def test_memory(self, llm: ClaudeLLM):
         await llm.simple_response(

plugins/aws/tests/test_aws.py

@@ -35,7 +35,7 @@ def assert_response_successful(self, response):
     async def llm(self) -> BedrockLLM:
         """Test BedrockLLM initialization with a provided client."""
         llm = BedrockLLM(model="qwen.qwen3-32b-v1:0", region_name="us-east-1")
-        llm._conversation = InMemoryConversation("be friendly", [])
+        llm.set_conversation(InMemoryConversation("be friendly", []))
         return llm
 
     @pytest.mark.asyncio

plugins/baseten/README.md

@@ -0,0 +1,117 @@
+# Qwen3-VL hosted on Baseten 
+Qwen3-VL is the latest open-source Video Language Model (VLM) from Alibaba. This plugin allows developers to easily run the model hosted on [Baseten](https://www.baseten.co/) with Vision Agents. The model accepts text and video and responds with text vocalised with the TTS service of your choice. 
+
+## Features
+
+- **Video understanding**: Automatically buffers and forwards video frames to Baseten-hosted VLM models
+- **Streaming responses**: Supports streaming text responses with real-time chunk events
+- **Frame buffering**: Configurable frame rate and buffer duration for optimal performance
+- **Event-driven**: Emits LLM events (chunks, completion, errors) for integration with other components
+
+## Installation
+
+```bash
+uv add vision-agents[baseten]
+```
+
+## Quick Start
+
+```python
+from vision_agents.core import Agent, User
+from vision_agents.plugins import baseten, getstream, deepgram, elevenlabs, vogent
+
+async def create_agent(**kwargs) -> Agent:
+    # Initialize the Baseten VLM
+    llm = baseten.VLM(model="qwen3vl")
+
+    # Create an agent with video understanding capabilities
+    agent = Agent(
+        edge=getstream.Edge(),
+        agent_user=User(name="Video Assistant", id="agent"),
+        instructions="You're a helpful video AI assistant. Analyze the video frames and respond to user questions about what you see.",
+        llm=llm,
+        stt=deepgram.STT(),
+        tts=elevenlabs.TTS(),
+        turn_detection=vogent.TurnDetection(),
+        processors=[],
+    )
+    return agent
+
+async def join_call(agent: Agent, call_type: str, call_id: str, **kwargs) -> None:
+    await agent.create_user()
+    call = await agent.create_call(call_type, call_id)
+
+    with await agent.join(call):
+        # The agent will automatically process video frames and respond to user input
+        await agent.finish()
+```
+
+## Configuration
+
+### Environment Variables
+
+- **`BASETEN_API_KEY`**: Your Baseten API key (required)
+- **`BASETEN_BASE_URL`**: The base URL for your Baseten API endpoint (required)
+
+### Initialization Parameters
+
+```python
+baseten.VLM(
+    model: str,                    # Baseten model name (e.g., "qwen3vl")
+    api_key: Optional[str] = None,  # API key (defaults to BASETEN_API_KEY env var)
+    base_url: Optional[str] = None, # Base URL (defaults to BASETEN_BASE_URL env var)
+    fps: int = 1,                   # Frames per second to process (default: 1)
+    frame_buffer_seconds: int = 10, # Seconds of video to buffer (default: 10)
+    client: Optional[AsyncOpenAI] = None,  # Custom OpenAI client (optional)
+)
+```
+
+### Parameters
+
+- **`model`**: The name of the Baseten-hosted model to use. Must be a vision-capable model.
+- **`api_key`**: Your Baseten API key. If not provided, reads from `BASETEN_API_KEY` environment variable.
+- **`base_url`**: The base URL for Baseten API. If not provided, reads from `BASETEN_BASE_URL` environment variable.
+- **`fps`**: Number of video frames per second to capture and send to the model. Lower values reduce API costs but may miss fast-moving content. Default is 1 fps.
+- **`frame_buffer_seconds`**: How many seconds of video to buffer. Total buffer size = `fps * frame_buffer_seconds`. Default is 10 seconds.
+- **`client`**: Optional pre-configured `AsyncOpenAI` client. If provided, `api_key` and `base_url` are ignored.
+
+## How It Works
+
+1. **Video Frame Buffering**: The plugin automatically subscribes to video tracks when the agent joins a call. It buffers frames at the specified FPS for the configured duration.
+
+2. **Frame Processing**: When responding to user input, the plugin:
+   - Converts buffered video frames to JPEG format
+   - Resizes frames to 800x600 (maintaining aspect ratio)
+   - Encodes frames as base64 data URLs
+
+3. **API Request**: Sends the conversation history (including system instructions) along with all buffered frames to the Baseten model.
+
+4. **Streaming Response**: Processes the streaming response and emits events for each chunk and completion.
+
+## Events
+
+The plugin emits the following events:
+
+- **`LLMResponseChunkEvent`**: Emitted for each text chunk in the streaming response
+- **`LLMResponseCompletedEvent`**: Emitted when the response stream completes
+- **`LLMErrorEvent`**: Emitted if an API request fails
+
+## Requirements
+
+- Python 3.10+
+- `openai>=2.5.0`
+- `vision-agents` (core framework)
+- Baseten API key and base URL
+
+## Notes
+
+- **Frame Rate**: The default FPS of 1 is optimized for VLM use cases. Higher FPS values will increase API costs and latency.
+- **Frame Size**: Frames are automatically resized to 800x600 pixels while maintaining aspect ratio to optimize API payload size.
+- **Buffer Duration**: The 10-second default buffer provides context for the model while keeping memory usage reasonable.
+- **Tool Calling**: Tool/function calling support is not yet implemented (see TODOs in code).
+
+## Troubleshooting
+
+- **No video processing**: Ensure the agent has joined a call with video tracks available. The plugin automatically subscribes to video when tracks are added.
+- **API errors**: Verify your `BASETEN_API_KEY` and `BASETEN_BASE_URL` are set correctly and the model name is valid.
+- **High latency**: Consider reducing `fps` or `frame_buffer_seconds` to decrease the number of frames sent per request.

plugins/baseten/example/README.md

@@ -0,0 +1 @@
+Please see root plugin readme. 

plugins/baseten/example/pyproject.toml

@@ -0,0 +1,21 @@
+[project]
+name = "qwen3-vl-example"
+version = "0.1.0"
+description = "Example using Qwen3 VL hosted on Baseten with Vision Agents"
+requires-python = ">=3.10"
+dependencies = [
+    "vision-agents",
+    "vision-agents-plugins-baseten",
+    "vision-agents-plugins-getstream",
+    "vision-agents-plugins-deepgram",
+    "vision-agents-plugins-elevenlabs",
+    "python-dotenv",
+]
+
+[tool.uv.sources]
+vision-agents = { workspace = true }
+vision-agents-plugins-baseten = { workspace = true }
+vision-agents-plugins-elevenlabs = { workspace = true }
+vision-agents-plugins-getstream = { workspace = true }
+vision-agents-plugins-deepgram = { workspace = true }
+

plugins/baseten/example/qwen_vl_example.py

@@ -0,0 +1,46 @@
+import asyncio
+
+from dotenv import load_dotenv
+
+from vision_agents.core import Agent, User, cli
+from vision_agents.core.agents import AgentLauncher
+from vision_agents.plugins import baseten, getstream, deepgram, elevenlabs
+from vision_agents.core.events import CallSessionParticipantJoinedEvent
+
+
+load_dotenv()
+
+
+async def create_agent(**kwargs) -> Agent:
+    # Initialize the Baseten VLM
+    llm = baseten.VLM(model="qwen3vl")
+
+    # Create an agent with video understanding capabilities
+    agent = Agent(
+        edge=getstream.Edge(),
+        agent_user=User(name="Video Assistant", id="agent"),
+        instructions="You're a helpful video AI assistant. Analyze the video frames and respond to user questions about what you see.",
+        llm=llm,
+        stt=deepgram.STT(),
+        tts=elevenlabs.TTS(),
+        processors=[],
+    )
+    return agent
+
+async def join_call(agent: Agent, call_type: str, call_id: str, **kwargs) -> None:
+    await agent.create_user()
+    call = await agent.create_call(call_type, call_id)
+
+    @agent.events.subscribe
+    async def on_participant_joined(event: CallSessionParticipantJoinedEvent):
+        if event.participant.user.id != "agent":
+            await asyncio.sleep(2)
+            await agent.simple_response("Describe what you currently see")
+
+    with await agent.join(call):
+        await agent.edge.open_demo(call)
+        # The agent will automatically process video frames and respond to user input
+        await agent.finish()
+
+if __name__ == "__main__":
+    cli(AgentLauncher(create_agent=create_agent, join_call=join_call))

plugins/baseten/pyproject.toml

@@ -0,0 +1,36 @@
+[build-system]
+requires = ["hatchling", "hatch-vcs"]
+build-backend = "hatchling.build"
+
+[project]
+name = "vision-agents-plugins-baseten"
+dynamic = ["version"]
+description = "Baseten plugin for vision agents"
+readme = "README.md"
+requires-python = ">=3.10"
+license = "MIT"
+dependencies = [
+    "vision-agents",
+    "openai>=2.5.0",
+]
+
+[project.urls]
+Documentation = "https://visionagents.ai/"
+Website = "https://visionagents.ai/"
+Source = "https://github.com/GetStream/Vision-Agents"
+
+[tool.hatch.version]
+source = "vcs"
+raw-options = { root = "..", search_parent_directories = true, fallback_version = "0.0.0" }
+
+[tool.hatch.build.targets.wheel]
+packages = ["."]
+
+[tool.uv.sources]
+vision-agents = { workspace = true }
+
+[dependency-groups]
+dev = [
+    "pytest>=8.4.1",
+    "pytest-asyncio>=1.0.0",
+]

plugins/baseten/vision_agents/plugins/baseten/__init__.py

@@ -0,0 +1,4 @@
+from .baseten_vlm import BasetenVLM as VLM
+
+
+__all__ = ["VLM"]