Skip to content

Llm

Integration adapters for optional third-party LLM SDKs.

This module exposes universal converter APIs to normalize responses and streaming outputs from popular LLM SDKs (e.g., LiteLLM, OpenAI) for use in PyAgenity agent graphs. Adapters provide unified conversion, streaming, and message normalization for agent workflows.

Exports

BaseConverter: Abstract base class for LLM response converters. ConverterType: Enum of supported converter types. LiteLLMConverter: Converter for LiteLLM responses and streams.

OpenAIConverter: (commented, available if implemented)

Modules:

Name Description
base_converter
litellm_converter
model_response_converter

Classes:

Name Description
BaseConverter

Abstract base class for all LLM response converters.
ConverterType

Enumeration of supported converter types for LLM responses.
LiteLLMConverter

Converter for LiteLLM responses to PyAgenity Message format.

Attributes

__all__ module-attribute 

__all__ = ['BaseConverter', 'ConverterType', 'LiteLLMConverter']

Classes

BaseConverter

Bases: ABC

Abstract base class for all LLM response converters.

Subclasses should implement methods to convert standard and streaming LLM responses into PyAgenity's internal message/event formats.

Attributes:

Name Type Description
state AgentState | None

Optional agent state for context during conversion.

Methods:

Name Description
__init__

Initialize the converter.
convert_response

Convert a standard agent response to a Message.
convert_streaming_response

Convert a streaming agent response to an async generator of EventModel or Message.
Source code in pyagenity/adapters/llm/base_converter.py 
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
class BaseConverter(ABC):
    """
    Abstract base class for all LLM response converters.

    Subclasses should implement methods to convert standard and streaming
    LLM responses into PyAgenity's internal message/event formats.

    Attributes:
        state (AgentState | None): Optional agent state for context during conversion.
    """

    def __init__(self, state: AgentState | None = None) -> None:
        """
        Initialize the converter.

        Args:
            state (AgentState | None): Optional agent state for context during conversion.
        """
        self.state = state

    @abstractmethod
    async def convert_response(self, response: Any) -> Message:
        """
        Convert a standard agent response to a Message.

        Args:
            response (Any): The raw response from the LLM or agent.

        Returns:
            Message: The converted message object.

        Raises:
            NotImplementedError: If not implemented in subclass.
        """
        raise NotImplementedError("Conversion not implemented for this converter")

    @abstractmethod
    async def convert_streaming_response(
        self,
        config: dict,
        node_name: str,
        response: Any,
        meta: dict | None = None,
    ) -> AsyncGenerator[EventModel | Message, None]:
        """
        Convert a streaming agent response to an async generator of EventModel or Message.

        Args:
            config (dict): Node configuration parameters.
            node_name (str): Name of the node processing the response.
            response (Any): The raw streaming response from the LLM or agent.
            meta (dict | None): Optional metadata for conversion.

        Yields:
            EventModel | Message: Chunks of the converted streaming response.

        Raises:
            NotImplementedError: If not implemented in subclass.
        """
        raise NotImplementedError("Streaming not implemented for this converter")

Attributes

state instance-attribute 
state = state

Functions

__init__ 
__init__(state=None)

Initialize the converter.

Parameters:

Name Type Description Default
state
AgentState | None

Optional agent state for context during conversion.

 None
Source code in pyagenity/adapters/llm/base_converter.py 
32
33
34
35
36
37
38
39
def __init__(self, state: AgentState | None = None) -> None:
    """
    Initialize the converter.

    Args:
        state (AgentState | None): Optional agent state for context during conversion.
    """
    self.state = state
convert_response abstractmethod async 
convert_response(response)

Convert a standard agent response to a Message.

Parameters:

Name Type Description Default
response
Any

The raw response from the LLM or agent.

 required

Returns:

Name Type Description
Message Message

The converted message object.

Raises:

Type Description
NotImplementedError

If not implemented in subclass.
Source code in pyagenity/adapters/llm/base_converter.py 
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
@abstractmethod
async def convert_response(self, response: Any) -> Message:
    """
    Convert a standard agent response to a Message.

    Args:
        response (Any): The raw response from the LLM or agent.

    Returns:
        Message: The converted message object.

    Raises:
        NotImplementedError: If not implemented in subclass.
    """
    raise NotImplementedError("Conversion not implemented for this converter")
convert_streaming_response abstractmethod async 
convert_streaming_response(config, node_name, response, meta=None)

Convert a streaming agent response to an async generator of EventModel or Message.

Parameters:

Name Type Description Default
config
dict

Node configuration parameters.

 required
node_name
str

Name of the node processing the response.

 required
response
Any

The raw streaming response from the LLM or agent.

 required
meta
dict | None

Optional metadata for conversion.

 None

Yields:

Type Description
AsyncGenerator[EventModel | Message, None]

EventModel | Message: Chunks of the converted streaming response.

Raises:

Type Description
NotImplementedError

If not implemented in subclass.
Source code in pyagenity/adapters/llm/base_converter.py 
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
@abstractmethod
async def convert_streaming_response(
    self,
    config: dict,
    node_name: str,
    response: Any,
    meta: dict | None = None,
) -> AsyncGenerator[EventModel | Message, None]:
    """
    Convert a streaming agent response to an async generator of EventModel or Message.

    Args:
        config (dict): Node configuration parameters.
        node_name (str): Name of the node processing the response.
        response (Any): The raw streaming response from the LLM or agent.
        meta (dict | None): Optional metadata for conversion.

    Yields:
        EventModel | Message: Chunks of the converted streaming response.

    Raises:
        NotImplementedError: If not implemented in subclass.
    """
    raise NotImplementedError("Streaming not implemented for this converter")

ConverterType

Bases: Enum

Enumeration of supported converter types for LLM responses.

Attributes:

Name Type Description
ANTHROPIC
CUSTOM
GOOGLE
LITELLM
OPENAI
Source code in pyagenity/adapters/llm/base_converter.py 
11
12
13
14
15
16
17
18
class ConverterType(Enum):
    """Enumeration of supported converter types for LLM responses."""

    OPENAI = "openai"
    LITELLM = "litellm"
    ANTHROPIC = "anthropic"
    GOOGLE = "google"
    CUSTOM = "custom"

Attributes

ANTHROPIC class-attribute instance-attribute 
ANTHROPIC = 'anthropic'
CUSTOM class-attribute instance-attribute 
CUSTOM = 'custom'
GOOGLE class-attribute instance-attribute 
GOOGLE = 'google'
LITELLM class-attribute instance-attribute 
LITELLM = 'litellm'
OPENAI class-attribute instance-attribute 
OPENAI = 'openai'

LiteLLMConverter

Bases: BaseConverter

Converter for LiteLLM responses to PyAgenity Message format.

Handles both standard and streaming responses, extracting content, reasoning, tool calls, and token usage details.

Methods:

Name Description
__init__

Initialize the converter.
convert_response

Convert a LiteLLM ModelResponse to a Message.
convert_streaming_response

Convert a LiteLLM streaming or standard response to Message(s).

Attributes:

Name Type Description
state
Source code in pyagenity/adapters/llm/litellm_converter.py 
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
class LiteLLMConverter(BaseConverter):
    """
    Converter for LiteLLM responses to PyAgenity Message format.

    Handles both standard and streaming responses, extracting content, reasoning,
    tool calls, and token usage details.
    """

    async def convert_response(self, response: ModelResponse) -> Message:
        """
        Convert a LiteLLM ModelResponse to a Message.

        Args:
            response (ModelResponse): The LiteLLM model response object.

        Returns:
            Message: The converted message object.

        Raises:
            ImportError: If LiteLLM is not installed.
        """
        if not HAS_LITELLM:
            raise ImportError("litellm is not installed. Please install it to use this converter.")

        data = response.model_dump()

        usages_data = data.get("usage", {})

        usages = TokenUsages(
            completion_tokens=usages_data.get("completion_tokens", 0),
            prompt_tokens=usages_data.get("prompt_tokens", 0),
            total_tokens=usages_data.get("total_tokens", 0),
            cache_creation_input_tokens=usages_data.get("cache_creation_input_tokens", 0),
            cache_read_input_tokens=usages_data.get("cache_read_input_tokens", 0),
            reasoning_tokens=usages_data.get("prompt_tokens_details", {}).get(
                "reasoning_tokens", 0
            ),
        )

        created_date = data.get("created", datetime.now())

        # Extract tool calls from response
        tools_calls = data.get("choices", [{}])[0].get("message", {}).get("tool_calls", []) or []

        logger.debug("Creating message from model response with id: %s", response.id)
        content = data.get("choices", [{}])[0].get("message", {}).get("content", "") or ""
        reasoning_content = (
            data.get("choices", [{}])[0].get("message", {}).get("reasoning_content", "") or ""
        )

        blocks = []
        if content:
            blocks.append(TextBlock(text=content))
        if reasoning_content:
            blocks.append(ReasoningBlock(summary=reasoning_content))
        final_tool_calls = []
        for tool_call in tools_calls:
            tool_id = tool_call.get("id", None)
            args = tool_call.get("function", {}).get("arguments", None)
            name = tool_call.get("function", {}).get("name", None)

            if not tool_id or not args or not name:
                continue

            blocks.append(
                ToolCallBlock(
                    name=name,
                    args=json.loads(args),
                    id=tool_id,
                )
            )

            if hasattr(tool_call, "model_dump"):
                final_tool_calls.append(tool_call.model_dump())
            else:
                final_tool_calls.append(tool_call)

        return Message(
            message_id=generate_id(response.id),
            role="assistant",
            content=blocks,
            reasoning=reasoning_content,
            timestamp=created_date,
            metadata={
                "provider": "litellm",
                "model": data.get("model", ""),
                "finish_reason": data.get("choices", [{}])[0].get("finish_reason", "UNKNOWN"),
                "object": data.get("object", ""),
                "prompt_tokens_details": usages_data.get("prompt_tokens_details", {}),
                "completion_tokens_details": usages_data.get("completion_tokens_details", {}),
            },
            usages=usages,
            raw=data,
            tools_calls=final_tool_calls if final_tool_calls else None,
        )

    def _process_chunk(
        self,
        chunk: ModelResponseStream | None,
        seq: int,
        accumulated_content: str,
        accumulated_reasoning_content: str,
        tool_calls: list,
        tool_ids: set,
    ) -> tuple[str, str, list, int, Message | None]:
        """
        Process a single chunk from a LiteLLM streaming response.

        Args:
            chunk (ModelResponseStream | None): The current chunk from the stream.
            seq (int): Sequence number of the chunk.
            accumulated_content (str): Accumulated text content so far.
            accumulated_reasoning_content (str): Accumulated reasoning content so far.
            tool_calls (list): List of tool calls detected so far.
            tool_ids (set): Set of tool call IDs to avoid duplicates.

        Returns:
            tuple: Updated accumulated content, reasoning, tool calls, sequence,
                and Message (if any).
        """
        if not chunk:
            return accumulated_content, accumulated_reasoning_content, tool_calls, seq, None

        msg: ModelResponseStream = chunk  # type: ignore
        if msg is None:
            return accumulated_content, accumulated_reasoning_content, tool_calls, seq, None
        if msg.choices is None or len(msg.choices) == 0:
            return accumulated_content, accumulated_reasoning_content, tool_calls, seq, None
        delta = msg.choices[0].delta
        if delta is None:
            return accumulated_content, accumulated_reasoning_content, tool_calls, seq, None

        # update text delta
        text_part = delta.content or ""
        content_blocks = []
        if text_part:
            content_blocks.append(TextBlock(text=text_part))
        reasoning_part = getattr(delta, "reasoning_content", "") or ""
        if reasoning_part:
            content_blocks.append(ReasoningBlock(summary=reasoning_part))
        accumulated_content += text_part
        accumulated_reasoning_content += reasoning_part
        # handle tool calls if present
        if getattr(delta, "tool_calls", None):
            for tc in delta.tool_calls:  # type: ignore[attr-defined]
                if not tc:
                    continue
                if tc.id in tool_ids:
                    continue
                tool_ids.add(tc.id)
                tool_calls.append(tc.model_dump())
                content_blocks.append(
                    ToolCallBlock(
                        name=tc.function.name,  # type: ignore
                        args=json.loads(tc.function.arguments),  # type: ignore
                        id=tc.id,  # type: ignore
                    )
                )

        output_message = Message(
            message_id=generate_id(msg.id),
            role="assistant",
            content=content_blocks,
            reasoning=accumulated_reasoning_content,
            tools_calls=tool_calls,
            delta=True,
        )

        return accumulated_content, accumulated_reasoning_content, tool_calls, seq, output_message

    async def _handle_stream(
        self,
        config: dict,
        node_name: str,
        stream: CustomStreamWrapper,
        meta: dict | None = None,
    ) -> AsyncGenerator[Message]:
        """
        Handle a LiteLLM streaming response and yield Message objects for each chunk.

        Args:
            config (dict): Node configuration parameters.
            node_name (str): Name of the node processing the response.
            stream (CustomStreamWrapper): The LiteLLM streaming response object.
            meta (dict | None): Optional metadata for conversion.

        Yields:
            Message: Converted message chunk from the stream.
        """
        accumulated_content = ""
        tool_calls = []
        tool_ids = set()
        accumulated_reasoning_content = ""
        seq = 0

        is_awaitable = inspect.isawaitable(stream)

        # Await stream if necessary
        if is_awaitable:
            stream = await stream

        # Try async iteration (acompletion)
        try:
            async for chunk in stream:
                accumulated_content, accumulated_reasoning_content, tool_calls, seq, message = (
                    self._process_chunk(
                        chunk,
                        seq,
                        accumulated_content,
                        accumulated_reasoning_content,
                        tool_calls,
                        tool_ids,
                    )
                )

                if message:
                    yield message
        except Exception:  # noqa: S110 # nosec B110
            pass

        # Try sync iteration (completion)
        try:
            for chunk in stream:
                accumulated_content, accumulated_reasoning_content, tool_calls, seq, message = (
                    self._process_chunk(
                        chunk,
                        seq,
                        accumulated_content,
                        accumulated_reasoning_content,
                        tool_calls,
                        tool_ids,
                    )
                )

                if message:
                    yield message
        except Exception:  # noqa: S110 # nosec B110
            pass

        # After streaming, yield final message
        metadata = meta or {}
        metadata["provider"] = "litellm"
        metadata["node_name"] = node_name
        metadata["thread_id"] = config.get("thread_id")

        blocks = []
        if accumulated_content:
            blocks.append(TextBlock(text=accumulated_content))
        if accumulated_reasoning_content:
            blocks.append(ReasoningBlock(summary=accumulated_reasoning_content))
        if tool_calls:
            for tc in tool_calls:
                blocks.append(
                    ToolCallBlock(
                        name=tc.get("function", {}).get("name", ""),
                        args=json.loads(tc.get("function", {}).get("arguments", "{}")),
                        id=tc.get("id", ""),
                    )
                )

        logger.debug(
            "Loop done Content: %s  Reasoning: %s Tool Calls: %s",
            accumulated_content,
            accumulated_reasoning_content,
            len(tool_calls),
        )
        message = Message(
            role="assistant",
            message_id=generate_id(None),
            content=blocks,
            delta=False,
            reasoning=accumulated_reasoning_content,
            tools_calls=tool_calls,
            metadata=metadata,
        )
        yield message

    async def convert_streaming_response(  # type: ignore
        self,
        config: dict,
        node_name: str,
        response: Any,
        meta: dict | None = None,
    ) -> AsyncGenerator[Message]:
        """
        Convert a LiteLLM streaming or standard response to Message(s).

        Args:
            config (dict): Node configuration parameters.
            node_name (str): Name of the node processing the response.
            response (Any): The LiteLLM response object (stream or standard).
            meta (dict | None): Optional metadata for conversion.

        Yields:
            Message: Converted message(s) from the response.

        Raises:
            ImportError: If LiteLLM is not installed.
            Exception: If response type is unsupported.
        """
        if not HAS_LITELLM:
            raise ImportError("litellm is not installed. Please install it to use this converter.")

        if isinstance(response, CustomStreamWrapper):  # type: ignore[possibly-unbound]
            stream = cast(CustomStreamWrapper, response)
            async for event in self._handle_stream(
                config or {},
                node_name or "",
                stream,
                meta,
            ):
                yield event
        elif isinstance(response, ModelResponse):  # type: ignore[possibly-unbound]
            message = await self.convert_response(cast(ModelResponse, response))
            yield message
        else:
            raise Exception("Unsupported response type for LiteLLMConverter")

Attributes

state instance-attribute 
state = state

Functions

__init__ 
__init__(state=None)

Initialize the converter.

Parameters:

Name Type Description Default
state
AgentState | None

Optional agent state for context during conversion.

 None
Source code in pyagenity/adapters/llm/base_converter.py 
32
33
34
35
36
37
38
39
def __init__(self, state: AgentState | None = None) -> None:
    """
    Initialize the converter.

    Args:
        state (AgentState | None): Optional agent state for context during conversion.
    """
    self.state = state
convert_response async 
convert_response(response)

Convert a LiteLLM ModelResponse to a Message.

Parameters:

Name Type Description Default
response
ModelResponse

The LiteLLM model response object.

 required

Returns:

Name Type Description
Message Message

The converted message object.

Raises:

Type Description
ImportError

If LiteLLM is not installed.
Source code in pyagenity/adapters/llm/litellm_converter.py 
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
async def convert_response(self, response: ModelResponse) -> Message:
    """
    Convert a LiteLLM ModelResponse to a Message.

    Args:
        response (ModelResponse): The LiteLLM model response object.

    Returns:
        Message: The converted message object.

    Raises:
        ImportError: If LiteLLM is not installed.
    """
    if not HAS_LITELLM:
        raise ImportError("litellm is not installed. Please install it to use this converter.")

    data = response.model_dump()

    usages_data = data.get("usage", {})

    usages = TokenUsages(
        completion_tokens=usages_data.get("completion_tokens", 0),
        prompt_tokens=usages_data.get("prompt_tokens", 0),
        total_tokens=usages_data.get("total_tokens", 0),
        cache_creation_input_tokens=usages_data.get("cache_creation_input_tokens", 0),
        cache_read_input_tokens=usages_data.get("cache_read_input_tokens", 0),
        reasoning_tokens=usages_data.get("prompt_tokens_details", {}).get(
            "reasoning_tokens", 0
        ),
    )

    created_date = data.get("created", datetime.now())

    # Extract tool calls from response
    tools_calls = data.get("choices", [{}])[0].get("message", {}).get("tool_calls", []) or []

    logger.debug("Creating message from model response with id: %s", response.id)
    content = data.get("choices", [{}])[0].get("message", {}).get("content", "") or ""
    reasoning_content = (
        data.get("choices", [{}])[0].get("message", {}).get("reasoning_content", "") or ""
    )

    blocks = []
    if content:
        blocks.append(TextBlock(text=content))
    if reasoning_content:
        blocks.append(ReasoningBlock(summary=reasoning_content))
    final_tool_calls = []
    for tool_call in tools_calls:
        tool_id = tool_call.get("id", None)
        args = tool_call.get("function", {}).get("arguments", None)
        name = tool_call.get("function", {}).get("name", None)

        if not tool_id or not args or not name:
            continue

        blocks.append(
            ToolCallBlock(
                name=name,
                args=json.loads(args),
                id=tool_id,
            )
        )

        if hasattr(tool_call, "model_dump"):
            final_tool_calls.append(tool_call.model_dump())
        else:
            final_tool_calls.append(tool_call)

    return Message(
        message_id=generate_id(response.id),
        role="assistant",
        content=blocks,
        reasoning=reasoning_content,
        timestamp=created_date,
        metadata={
            "provider": "litellm",
            "model": data.get("model", ""),
            "finish_reason": data.get("choices", [{}])[0].get("finish_reason", "UNKNOWN"),
            "object": data.get("object", ""),
            "prompt_tokens_details": usages_data.get("prompt_tokens_details", {}),
            "completion_tokens_details": usages_data.get("completion_tokens_details", {}),
        },
        usages=usages,
        raw=data,
        tools_calls=final_tool_calls if final_tool_calls else None,
    )
convert_streaming_response async 
convert_streaming_response(config, node_name, response, meta=None)

Convert a LiteLLM streaming or standard response to Message(s).

Parameters:

Name Type Description Default
config
dict

Node configuration parameters.

 required
node_name
str

Name of the node processing the response.

 required
response
Any

The LiteLLM response object (stream or standard).

 required
meta
dict | None

Optional metadata for conversion.

 None

Yields:

Name Type Description
Message AsyncGenerator[Message]

Converted message(s) from the response.

Raises:

Type Description
ImportError

If LiteLLM is not installed.
Exception

If response type is unsupported.
Source code in pyagenity/adapters/llm/litellm_converter.py 
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
async def convert_streaming_response(  # type: ignore
    self,
    config: dict,
    node_name: str,
    response: Any,
    meta: dict | None = None,
) -> AsyncGenerator[Message]:
    """
    Convert a LiteLLM streaming or standard response to Message(s).

    Args:
        config (dict): Node configuration parameters.
        node_name (str): Name of the node processing the response.
        response (Any): The LiteLLM response object (stream or standard).
        meta (dict | None): Optional metadata for conversion.

    Yields:
        Message: Converted message(s) from the response.

    Raises:
        ImportError: If LiteLLM is not installed.
        Exception: If response type is unsupported.
    """
    if not HAS_LITELLM:
        raise ImportError("litellm is not installed. Please install it to use this converter.")

    if isinstance(response, CustomStreamWrapper):  # type: ignore[possibly-unbound]
        stream = cast(CustomStreamWrapper, response)
        async for event in self._handle_stream(
            config or {},
            node_name or "",
            stream,
            meta,
        ):
            yield event
    elif isinstance(response, ModelResponse):  # type: ignore[possibly-unbound]
        message = await self.convert_response(cast(ModelResponse, response))
        yield message
    else:
        raise Exception("Unsupported response type for LiteLLMConverter")

Modules

base_converter

Classes:

Name Description
BaseConverter

Abstract base class for all LLM response converters.
ConverterType

Enumeration of supported converter types for LLM responses.

Classes

BaseConverter

Bases: ABC

Abstract base class for all LLM response converters.

Subclasses should implement methods to convert standard and streaming LLM responses into PyAgenity's internal message/event formats.

Attributes:

Name Type Description
state AgentState | None

Optional agent state for context during conversion.

Methods:

Name Description
__init__

Initialize the converter.
convert_response

Convert a standard agent response to a Message.
convert_streaming_response

Convert a streaming agent response to an async generator of EventModel or Message.
Source code in pyagenity/adapters/llm/base_converter.py 
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
class BaseConverter(ABC):
    """
    Abstract base class for all LLM response converters.

    Subclasses should implement methods to convert standard and streaming
    LLM responses into PyAgenity's internal message/event formats.

    Attributes:
        state (AgentState | None): Optional agent state for context during conversion.
    """

    def __init__(self, state: AgentState | None = None) -> None:
        """
        Initialize the converter.

        Args:
            state (AgentState | None): Optional agent state for context during conversion.
        """
        self.state = state

    @abstractmethod
    async def convert_response(self, response: Any) -> Message:
        """
        Convert a standard agent response to a Message.

        Args:
            response (Any): The raw response from the LLM or agent.

        Returns:
            Message: The converted message object.

        Raises:
            NotImplementedError: If not implemented in subclass.
        """
        raise NotImplementedError("Conversion not implemented for this converter")

    @abstractmethod
    async def convert_streaming_response(
        self,
        config: dict,
        node_name: str,
        response: Any,
        meta: dict | None = None,
    ) -> AsyncGenerator[EventModel | Message, None]:
        """
        Convert a streaming agent response to an async generator of EventModel or Message.

        Args:
            config (dict): Node configuration parameters.
            node_name (str): Name of the node processing the response.
            response (Any): The raw streaming response from the LLM or agent.
            meta (dict | None): Optional metadata for conversion.

        Yields:
            EventModel | Message: Chunks of the converted streaming response.

        Raises:
            NotImplementedError: If not implemented in subclass.
        """
        raise NotImplementedError("Streaming not implemented for this converter")
Attributes
state instance-attribute 
state = state
Functions
__init__ 
__init__(state=None)

Initialize the converter.

Parameters:

Name Type Description Default
state AgentState | None

Optional agent state for context during conversion.

 None
Source code in pyagenity/adapters/llm/base_converter.py 
32
33
34
35
36
37
38
39
def __init__(self, state: AgentState | None = None) -> None:
    """
    Initialize the converter.

    Args:
        state (AgentState | None): Optional agent state for context during conversion.
    """
    self.state = state
convert_response abstractmethod async 
convert_response(response)

Convert a standard agent response to a Message.

Parameters:

Name Type Description Default
response Any

The raw response from the LLM or agent.

 required

Returns:

Name Type Description
Message Message

The converted message object.

Raises:

Type Description
NotImplementedError

If not implemented in subclass.
Source code in pyagenity/adapters/llm/base_converter.py 
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
@abstractmethod
async def convert_response(self, response: Any) -> Message:
    """
    Convert a standard agent response to a Message.

    Args:
        response (Any): The raw response from the LLM or agent.

    Returns:
        Message: The converted message object.

    Raises:
        NotImplementedError: If not implemented in subclass.
    """
    raise NotImplementedError("Conversion not implemented for this converter")
convert_streaming_response abstractmethod async 
convert_streaming_response(config, node_name, response, meta=None)

Convert a streaming agent response to an async generator of EventModel or Message.

Parameters:

Name Type Description Default
config dict

Node configuration parameters.

 required
node_name str

Name of the node processing the response.

 required
response Any

The raw streaming response from the LLM or agent.

 required
meta dict | None

Optional metadata for conversion.

 None

Yields:

Type Description
AsyncGenerator[EventModel | Message, None]

EventModel | Message: Chunks of the converted streaming response.

Raises:

Type Description
NotImplementedError

If not implemented in subclass.
Source code in pyagenity/adapters/llm/base_converter.py 
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
@abstractmethod
async def convert_streaming_response(
    self,
    config: dict,
    node_name: str,
    response: Any,
    meta: dict | None = None,
) -> AsyncGenerator[EventModel | Message, None]:
    """
    Convert a streaming agent response to an async generator of EventModel or Message.

    Args:
        config (dict): Node configuration parameters.
        node_name (str): Name of the node processing the response.
        response (Any): The raw streaming response from the LLM or agent.
        meta (dict | None): Optional metadata for conversion.

    Yields:
        EventModel | Message: Chunks of the converted streaming response.

    Raises:
        NotImplementedError: If not implemented in subclass.
    """
    raise NotImplementedError("Streaming not implemented for this converter")
ConverterType

Bases: Enum

Enumeration of supported converter types for LLM responses.

Attributes:

Name Type Description
ANTHROPIC
CUSTOM
GOOGLE
LITELLM
OPENAI
Source code in pyagenity/adapters/llm/base_converter.py 
11
12
13
14
15
16
17
18
class ConverterType(Enum):
    """Enumeration of supported converter types for LLM responses."""

    OPENAI = "openai"
    LITELLM = "litellm"
    ANTHROPIC = "anthropic"
    GOOGLE = "google"
    CUSTOM = "custom"
Attributes
ANTHROPIC class-attribute instance-attribute 
ANTHROPIC = 'anthropic'
CUSTOM class-attribute instance-attribute 
CUSTOM = 'custom'
GOOGLE class-attribute instance-attribute 
GOOGLE = 'google'
LITELLM class-attribute instance-attribute 
LITELLM = 'litellm'
OPENAI class-attribute instance-attribute 
OPENAI = 'openai'

litellm_converter

Classes:

Name Description
LiteLLMConverter

Converter for LiteLLM responses to PyAgenity Message format.

Attributes:

Name Type Description
HAS_LITELLM
logger

Attributes

HAS_LITELLM module-attribute 
HAS_LITELLM = True
logger module-attribute 
logger = getLogger(__name__)

Classes

LiteLLMConverter

Bases: BaseConverter

Converter for LiteLLM responses to PyAgenity Message format.

Handles both standard and streaming responses, extracting content, reasoning, tool calls, and token usage details.

Methods:

Name Description
__init__

Initialize the converter.
convert_response

Convert a LiteLLM ModelResponse to a Message.
convert_streaming_response

Convert a LiteLLM streaming or standard response to Message(s).

Attributes:

Name Type Description
state
Source code in pyagenity/adapters/llm/litellm_converter.py 
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
class LiteLLMConverter(BaseConverter):
    """
    Converter for LiteLLM responses to PyAgenity Message format.

    Handles both standard and streaming responses, extracting content, reasoning,
    tool calls, and token usage details.
    """

    async def convert_response(self, response: ModelResponse) -> Message:
        """
        Convert a LiteLLM ModelResponse to a Message.

        Args:
            response (ModelResponse): The LiteLLM model response object.

        Returns:
            Message: The converted message object.

        Raises:
            ImportError: If LiteLLM is not installed.
        """
        if not HAS_LITELLM:
            raise ImportError("litellm is not installed. Please install it to use this converter.")

        data = response.model_dump()

        usages_data = data.get("usage", {})

        usages = TokenUsages(
            completion_tokens=usages_data.get("completion_tokens", 0),
            prompt_tokens=usages_data.get("prompt_tokens", 0),
            total_tokens=usages_data.get("total_tokens", 0),
            cache_creation_input_tokens=usages_data.get("cache_creation_input_tokens", 0),
            cache_read_input_tokens=usages_data.get("cache_read_input_tokens", 0),
            reasoning_tokens=usages_data.get("prompt_tokens_details", {}).get(
                "reasoning_tokens", 0
            ),
        )

        created_date = data.get("created", datetime.now())

        # Extract tool calls from response
        tools_calls = data.get("choices", [{}])[0].get("message", {}).get("tool_calls", []) or []

        logger.debug("Creating message from model response with id: %s", response.id)
        content = data.get("choices", [{}])[0].get("message", {}).get("content", "") or ""
        reasoning_content = (
            data.get("choices", [{}])[0].get("message", {}).get("reasoning_content", "") or ""
        )

        blocks = []
        if content:
            blocks.append(TextBlock(text=content))
        if reasoning_content:
            blocks.append(ReasoningBlock(summary=reasoning_content))
        final_tool_calls = []
        for tool_call in tools_calls:
            tool_id = tool_call.get("id", None)
            args = tool_call.get("function", {}).get("arguments", None)
            name = tool_call.get("function", {}).get("name", None)

            if not tool_id or not args or not name:
                continue

            blocks.append(
                ToolCallBlock(
                    name=name,
                    args=json.loads(args),
                    id=tool_id,
                )
            )

            if hasattr(tool_call, "model_dump"):
                final_tool_calls.append(tool_call.model_dump())
            else:
                final_tool_calls.append(tool_call)

        return Message(
            message_id=generate_id(response.id),
            role="assistant",
            content=blocks,
            reasoning=reasoning_content,
            timestamp=created_date,
            metadata={
                "provider": "litellm",
                "model": data.get("model", ""),
                "finish_reason": data.get("choices", [{}])[0].get("finish_reason", "UNKNOWN"),
                "object": data.get("object", ""),
                "prompt_tokens_details": usages_data.get("prompt_tokens_details", {}),
                "completion_tokens_details": usages_data.get("completion_tokens_details", {}),
            },
            usages=usages,
            raw=data,
            tools_calls=final_tool_calls if final_tool_calls else None,
        )

    def _process_chunk(
        self,
        chunk: ModelResponseStream | None,
        seq: int,
        accumulated_content: str,
        accumulated_reasoning_content: str,
        tool_calls: list,
        tool_ids: set,
    ) -> tuple[str, str, list, int, Message | None]:
        """
        Process a single chunk from a LiteLLM streaming response.

        Args:
            chunk (ModelResponseStream | None): The current chunk from the stream.
            seq (int): Sequence number of the chunk.
            accumulated_content (str): Accumulated text content so far.
            accumulated_reasoning_content (str): Accumulated reasoning content so far.
            tool_calls (list): List of tool calls detected so far.
            tool_ids (set): Set of tool call IDs to avoid duplicates.

        Returns:
            tuple: Updated accumulated content, reasoning, tool calls, sequence,
                and Message (if any).
        """
        if not chunk:
            return accumulated_content, accumulated_reasoning_content, tool_calls, seq, None

        msg: ModelResponseStream = chunk  # type: ignore
        if msg is None:
            return accumulated_content, accumulated_reasoning_content, tool_calls, seq, None
        if msg.choices is None or len(msg.choices) == 0:
            return accumulated_content, accumulated_reasoning_content, tool_calls, seq, None
        delta = msg.choices[0].delta
        if delta is None:
            return accumulated_content, accumulated_reasoning_content, tool_calls, seq, None

        # update text delta
        text_part = delta.content or ""
        content_blocks = []
        if text_part:
            content_blocks.append(TextBlock(text=text_part))
        reasoning_part = getattr(delta, "reasoning_content", "") or ""
        if reasoning_part:
            content_blocks.append(ReasoningBlock(summary=reasoning_part))
        accumulated_content += text_part
        accumulated_reasoning_content += reasoning_part
        # handle tool calls if present
        if getattr(delta, "tool_calls", None):
            for tc in delta.tool_calls:  # type: ignore[attr-defined]
                if not tc:
                    continue
                if tc.id in tool_ids:
                    continue
                tool_ids.add(tc.id)
                tool_calls.append(tc.model_dump())
                content_blocks.append(
                    ToolCallBlock(
                        name=tc.function.name,  # type: ignore
                        args=json.loads(tc.function.arguments),  # type: ignore
                        id=tc.id,  # type: ignore
                    )
                )

        output_message = Message(
            message_id=generate_id(msg.id),
            role="assistant",
            content=content_blocks,
            reasoning=accumulated_reasoning_content,
            tools_calls=tool_calls,
            delta=True,
        )

        return accumulated_content, accumulated_reasoning_content, tool_calls, seq, output_message

    async def _handle_stream(
        self,
        config: dict,
        node_name: str,
        stream: CustomStreamWrapper,
        meta: dict | None = None,
    ) -> AsyncGenerator[Message]:
        """
        Handle a LiteLLM streaming response and yield Message objects for each chunk.

        Args:
            config (dict): Node configuration parameters.
            node_name (str): Name of the node processing the response.
            stream (CustomStreamWrapper): The LiteLLM streaming response object.
            meta (dict | None): Optional metadata for conversion.

        Yields:
            Message: Converted message chunk from the stream.
        """
        accumulated_content = ""
        tool_calls = []
        tool_ids = set()
        accumulated_reasoning_content = ""
        seq = 0

        is_awaitable = inspect.isawaitable(stream)

        # Await stream if necessary
        if is_awaitable:
            stream = await stream

        # Try async iteration (acompletion)
        try:
            async for chunk in stream:
                accumulated_content, accumulated_reasoning_content, tool_calls, seq, message = (
                    self._process_chunk(
                        chunk,
                        seq,
                        accumulated_content,
                        accumulated_reasoning_content,
                        tool_calls,
                        tool_ids,
                    )
                )

                if message:
                    yield message
        except Exception:  # noqa: S110 # nosec B110
            pass

        # Try sync iteration (completion)
        try:
            for chunk in stream:
                accumulated_content, accumulated_reasoning_content, tool_calls, seq, message = (
                    self._process_chunk(
                        chunk,
                        seq,
                        accumulated_content,
                        accumulated_reasoning_content,
                        tool_calls,
                        tool_ids,
                    )
                )

                if message:
                    yield message
        except Exception:  # noqa: S110 # nosec B110
            pass

        # After streaming, yield final message
        metadata = meta or {}
        metadata["provider"] = "litellm"
        metadata["node_name"] = node_name
        metadata["thread_id"] = config.get("thread_id")

        blocks = []
        if accumulated_content:
            blocks.append(TextBlock(text=accumulated_content))
        if accumulated_reasoning_content:
            blocks.append(ReasoningBlock(summary=accumulated_reasoning_content))
        if tool_calls:
            for tc in tool_calls:
                blocks.append(
                    ToolCallBlock(
                        name=tc.get("function", {}).get("name", ""),
                        args=json.loads(tc.get("function", {}).get("arguments", "{}")),
                        id=tc.get("id", ""),
                    )
                )

        logger.debug(
            "Loop done Content: %s  Reasoning: %s Tool Calls: %s",
            accumulated_content,
            accumulated_reasoning_content,
            len(tool_calls),
        )
        message = Message(
            role="assistant",
            message_id=generate_id(None),
            content=blocks,
            delta=False,
            reasoning=accumulated_reasoning_content,
            tools_calls=tool_calls,
            metadata=metadata,
        )
        yield message

    async def convert_streaming_response(  # type: ignore
        self,
        config: dict,
        node_name: str,
        response: Any,
        meta: dict | None = None,
    ) -> AsyncGenerator[Message]:
        """
        Convert a LiteLLM streaming or standard response to Message(s).

        Args:
            config (dict): Node configuration parameters.
            node_name (str): Name of the node processing the response.
            response (Any): The LiteLLM response object (stream or standard).
            meta (dict | None): Optional metadata for conversion.

        Yields:
            Message: Converted message(s) from the response.

        Raises:
            ImportError: If LiteLLM is not installed.
            Exception: If response type is unsupported.
        """
        if not HAS_LITELLM:
            raise ImportError("litellm is not installed. Please install it to use this converter.")

        if isinstance(response, CustomStreamWrapper):  # type: ignore[possibly-unbound]
            stream = cast(CustomStreamWrapper, response)
            async for event in self._handle_stream(
                config or {},
                node_name or "",
                stream,
                meta,
            ):
                yield event
        elif isinstance(response, ModelResponse):  # type: ignore[possibly-unbound]
            message = await self.convert_response(cast(ModelResponse, response))
            yield message
        else:
            raise Exception("Unsupported response type for LiteLLMConverter")
Attributes
state instance-attribute 
state = state
Functions
__init__ 
__init__(state=None)

Initialize the converter.

Parameters:

Name Type Description Default
state AgentState | None

Optional agent state for context during conversion.

 None
Source code in pyagenity/adapters/llm/base_converter.py 
32
33
34
35
36
37
38
39
def __init__(self, state: AgentState | None = None) -> None:
    """
    Initialize the converter.

    Args:
        state (AgentState | None): Optional agent state for context during conversion.
    """
    self.state = state
convert_response async 
convert_response(response)

Convert a LiteLLM ModelResponse to a Message.

Parameters:

Name Type Description Default
response ModelResponse

The LiteLLM model response object.

 required

Returns:

Name Type Description
Message Message

The converted message object.

Raises:

Type Description
ImportError

If LiteLLM is not installed.
Source code in pyagenity/adapters/llm/litellm_converter.py 
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
async def convert_response(self, response: ModelResponse) -> Message:
    """
    Convert a LiteLLM ModelResponse to a Message.

    Args:
        response (ModelResponse): The LiteLLM model response object.

    Returns:
        Message: The converted message object.

    Raises:
        ImportError: If LiteLLM is not installed.
    """
    if not HAS_LITELLM:
        raise ImportError("litellm is not installed. Please install it to use this converter.")

    data = response.model_dump()

    usages_data = data.get("usage", {})

    usages = TokenUsages(
        completion_tokens=usages_data.get("completion_tokens", 0),
        prompt_tokens=usages_data.get("prompt_tokens", 0),
        total_tokens=usages_data.get("total_tokens", 0),
        cache_creation_input_tokens=usages_data.get("cache_creation_input_tokens", 0),
        cache_read_input_tokens=usages_data.get("cache_read_input_tokens", 0),
        reasoning_tokens=usages_data.get("prompt_tokens_details", {}).get(
            "reasoning_tokens", 0
        ),
    )

    created_date = data.get("created", datetime.now())

    # Extract tool calls from response
    tools_calls = data.get("choices", [{}])[0].get("message", {}).get("tool_calls", []) or []

    logger.debug("Creating message from model response with id: %s", response.id)
    content = data.get("choices", [{}])[0].get("message", {}).get("content", "") or ""
    reasoning_content = (
        data.get("choices", [{}])[0].get("message", {}).get("reasoning_content", "") or ""
    )

    blocks = []
    if content:
        blocks.append(TextBlock(text=content))
    if reasoning_content:
        blocks.append(ReasoningBlock(summary=reasoning_content))
    final_tool_calls = []
    for tool_call in tools_calls:
        tool_id = tool_call.get("id", None)
        args = tool_call.get("function", {}).get("arguments", None)
        name = tool_call.get("function", {}).get("name", None)

        if not tool_id or not args or not name:
            continue

        blocks.append(
            ToolCallBlock(
                name=name,
                args=json.loads(args),
                id=tool_id,
            )
        )

        if hasattr(tool_call, "model_dump"):
            final_tool_calls.append(tool_call.model_dump())
        else:
            final_tool_calls.append(tool_call)

    return Message(
        message_id=generate_id(response.id),
        role="assistant",
        content=blocks,
        reasoning=reasoning_content,
        timestamp=created_date,
        metadata={
            "provider": "litellm",
            "model": data.get("model", ""),
            "finish_reason": data.get("choices", [{}])[0].get("finish_reason", "UNKNOWN"),
            "object": data.get("object", ""),
            "prompt_tokens_details": usages_data.get("prompt_tokens_details", {}),
            "completion_tokens_details": usages_data.get("completion_tokens_details", {}),
        },
        usages=usages,
        raw=data,
        tools_calls=final_tool_calls if final_tool_calls else None,
    )
convert_streaming_response async 
convert_streaming_response(config, node_name, response, meta=None)

Convert a LiteLLM streaming or standard response to Message(s).

Parameters:

Name Type Description Default
config dict

Node configuration parameters.

 required
node_name str

Name of the node processing the response.

 required
response Any

The LiteLLM response object (stream or standard).

 required
meta dict | None

Optional metadata for conversion.

 None

Yields:

Name Type Description
Message AsyncGenerator[Message]

Converted message(s) from the response.

Raises:

Type Description
ImportError

If LiteLLM is not installed.
Exception

If response type is unsupported.
Source code in pyagenity/adapters/llm/litellm_converter.py 
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
async def convert_streaming_response(  # type: ignore
    self,
    config: dict,
    node_name: str,
    response: Any,
    meta: dict | None = None,
) -> AsyncGenerator[Message]:
    """
    Convert a LiteLLM streaming or standard response to Message(s).

    Args:
        config (dict): Node configuration parameters.
        node_name (str): Name of the node processing the response.
        response (Any): The LiteLLM response object (stream or standard).
        meta (dict | None): Optional metadata for conversion.

    Yields:
        Message: Converted message(s) from the response.

    Raises:
        ImportError: If LiteLLM is not installed.
        Exception: If response type is unsupported.
    """
    if not HAS_LITELLM:
        raise ImportError("litellm is not installed. Please install it to use this converter.")

    if isinstance(response, CustomStreamWrapper):  # type: ignore[possibly-unbound]
        stream = cast(CustomStreamWrapper, response)
        async for event in self._handle_stream(
            config or {},
            node_name or "",
            stream,
            meta,
        ):
            yield event
    elif isinstance(response, ModelResponse):  # type: ignore[possibly-unbound]
        message = await self.convert_response(cast(ModelResponse, response))
        yield message
    else:
        raise Exception("Unsupported response type for LiteLLMConverter")

Functions

model_response_converter

Classes:

Name Description
ModelResponseConverter

Wrap an LLM SDK call and normalize its output via a converter.

Classes

ModelResponseConverter

Wrap an LLM SDK call and normalize its output via a converter.

Supports sync/async invocation and streaming. Use invoke() for non-streaming calls and stream() for streaming calls.

Methods:

Name Description
__init__

Initialize ModelResponseConverter.
invoke

Call the underlying function and convert a non-streaming response to Message.
stream

Call the underlying function and yield normalized streaming events and final Message.

Attributes:

Name Type Description
converter
response
Source code in pyagenity/adapters/llm/model_response_converter.py 
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
class ModelResponseConverter:
    """Wrap an LLM SDK call and normalize its output via a converter.

    Supports sync/async invocation and streaming. Use `invoke()` for
    non-streaming calls and `stream()` for streaming calls.
    """

    def __init__(
        self,
        response: Any | Callable[..., Any],
        converter: BaseConverter | str,
    ) -> None:
        """
        Initialize ModelResponseConverter.

        Args:
            response (Any | Callable[..., Any]): The LLM response or a callable returning
                a response.
            converter (BaseConverter | str): Converter instance or string identifier
                (e.g., "litellm").

        Raises:
            ValueError: If the converter is not supported.
        """
        self.response = response

        if isinstance(converter, str) and converter == "litellm":
            from .litellm_converter import LiteLLMConverter

            self.converter = LiteLLMConverter()
        elif isinstance(converter, BaseConverter):
            self.converter = converter
        else:
            raise ValueError(f"Unsupported converter: {converter}")

    async def invoke(self) -> Message:
        """
        Call the underlying function and convert a non-streaming response to Message.

        Returns:
            Message: The normalized message from the LLM response.

        Raises:
            Exception: If the underlying function or converter fails.
        """
        if callable(self.response):
            if inspect.iscoroutinefunction(self.response):
                response = await self.response()
            else:
                response = self.response()
        else:
            response = self.response

        return await self.converter.convert_response(response)  # type: ignore

    async def stream(
        self,
        config: dict,
        node_name: str,
        meta: dict | None = None,
    ) -> AsyncGenerator[Message]:
        """
        Call the underlying function and yield normalized streaming events and final Message.

        Args:
            config (dict): Node configuration parameters for streaming.
            node_name (str): Name of the node processing the response.
            meta (dict | None): Optional metadata for conversion.

        Yields:
            Message: Normalized streaming message events from the LLM response.

        Raises:
            ValueError: If config is not provided.
            Exception: If the underlying function or converter fails.
        """
        if not config:
            raise ValueError("Config must be provided for streaming conversion")

        if callable(self.response):
            if inspect.iscoroutinefunction(self.response):
                response = await self.response()
            else:
                response = self.response()
        else:
            response = self.response

        async for item in self.converter.convert_streaming_response(  # type: ignore
            config,
            node_name=node_name,
            response=response,
            meta=meta,
        ):
            yield item
Attributes
converter instance-attribute 
converter = LiteLLMConverter()
response instance-attribute 
response = response
Functions
__init__ 
__init__(response, converter)

Initialize ModelResponseConverter.

Parameters:

Name Type Description Default
response Any | Callable[..., Any]

The LLM response or a callable returning a response.

 required
converter BaseConverter | str

Converter instance or string identifier (e.g., "litellm").

 required

Raises:

Type Description
ValueError

If the converter is not supported.
Source code in pyagenity/adapters/llm/model_response_converter.py 
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
def __init__(
    self,
    response: Any | Callable[..., Any],
    converter: BaseConverter | str,
) -> None:
    """
    Initialize ModelResponseConverter.

    Args:
        response (Any | Callable[..., Any]): The LLM response or a callable returning
            a response.
        converter (BaseConverter | str): Converter instance or string identifier
            (e.g., "litellm").

    Raises:
        ValueError: If the converter is not supported.
    """
    self.response = response

    if isinstance(converter, str) and converter == "litellm":
        from .litellm_converter import LiteLLMConverter

        self.converter = LiteLLMConverter()
    elif isinstance(converter, BaseConverter):
        self.converter = converter
    else:
        raise ValueError(f"Unsupported converter: {converter}")
invoke async 
invoke()

Call the underlying function and convert a non-streaming response to Message.

Returns:

Name Type Description
Message Message

The normalized message from the LLM response.

Raises:

Type Description
Exception

If the underlying function or converter fails.
Source code in pyagenity/adapters/llm/model_response_converter.py 
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
async def invoke(self) -> Message:
    """
    Call the underlying function and convert a non-streaming response to Message.

    Returns:
        Message: The normalized message from the LLM response.

    Raises:
        Exception: If the underlying function or converter fails.
    """
    if callable(self.response):
        if inspect.iscoroutinefunction(self.response):
            response = await self.response()
        else:
            response = self.response()
    else:
        response = self.response

    return await self.converter.convert_response(response)  # type: ignore
stream async 
stream(config, node_name, meta=None)

Call the underlying function and yield normalized streaming events and final Message.

Parameters:

Name Type Description Default
config dict

Node configuration parameters for streaming.

 required
node_name str

Name of the node processing the response.

 required
meta dict | None

Optional metadata for conversion.

 None

Yields:

Name Type Description
Message AsyncGenerator[Message]

Normalized streaming message events from the LLM response.

Raises:

Type Description
ValueError

If config is not provided.
Exception

If the underlying function or converter fails.
Source code in pyagenity/adapters/llm/model_response_converter.py 
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
async def stream(
    self,
    config: dict,
    node_name: str,
    meta: dict | None = None,
) -> AsyncGenerator[Message]:
    """
    Call the underlying function and yield normalized streaming events and final Message.

    Args:
        config (dict): Node configuration parameters for streaming.
        node_name (str): Name of the node processing the response.
        meta (dict | None): Optional metadata for conversion.

    Yields:
        Message: Normalized streaming message events from the LLM response.

    Raises:
        ValueError: If config is not provided.
        Exception: If the underlying function or converter fails.
    """
    if not config:
        raise ValueError("Config must be provided for streaming conversion")

    if callable(self.response):
        if inspect.iscoroutinefunction(self.response):
            response = await self.response()
        else:
            response = self.response()
    else:
        response = self.response

    async for item in self.converter.convert_streaming_response(  # type: ignore
        config,
        node_name=node_name,
        response=response,
        meta=meta,
    ):
        yield item