core[patch]: update dict prompt template (#30967)

Align with JS changes made in
https://github.com/langchain-ai/langchainjs/pull/8043

libs/core/langchain_core/load/mapping.py

@@ -540,6 +540,12 @@ SERIALIZABLE_MAPPING: dict[tuple[str, ...], tuple[str, ...]] = {
         "chat_models",
         "ChatSambaStudio",
     ),
+    ("langchain_core", "prompts", "message", "_DictMessagePromptTemplate"): (
+        "langchain_core",
+        "prompts",
+        "dict",
+        "DictPromptTemplate",
+    ),
 }
 
 # Needed for backwards compatibility for old versions of LangChain where things

libs/core/langchain_core/prompts/__init__.py

@@ -44,6 +44,7 @@ if TYPE_CHECKING:
         MessagesPlaceholder,
         SystemMessagePromptTemplate,
     )
+    from langchain_core.prompts.dict import DictPromptTemplate
     from langchain_core.prompts.few_shot import (
         FewShotChatMessagePromptTemplate,
         FewShotPromptTemplate,
@@ -68,6 +69,7 @@ __all__ = (
     "BasePromptTemplate",
     "ChatMessagePromptTemplate",
     "ChatPromptTemplate",
+    "DictPromptTemplate",
     "FewShotPromptTemplate",
     "FewShotPromptWithTemplates",
     "FewShotChatMessagePromptTemplate",
@@ -94,6 +96,7 @@ _dynamic_imports = {
     "BaseChatPromptTemplate": "chat",
     "ChatMessagePromptTemplate": "chat",
     "ChatPromptTemplate": "chat",
+    "DictPromptTemplate": "dict",
     "HumanMessagePromptTemplate": "chat",
     "MessagesPlaceholder": "chat",
     "SystemMessagePromptTemplate": "chat",

libs/core/langchain_core/prompts/chat.py

@@ -37,10 +37,10 @@ from langchain_core.messages import (
 from langchain_core.messages.base import get_msg_title_repr
 from langchain_core.prompt_values import ChatPromptValue, ImageURL, PromptValue
 from langchain_core.prompts.base import BasePromptTemplate
+from langchain_core.prompts.dict import DictPromptTemplate
 from langchain_core.prompts.image import ImagePromptTemplate
 from langchain_core.prompts.message import (
     BaseMessagePromptTemplate,
-    _DictMessagePromptTemplate,
 )
 from langchain_core.prompts.prompt import PromptTemplate
 from langchain_core.prompts.string import (
@@ -396,9 +396,7 @@ class _StringImageMessagePromptTemplate(BaseMessagePromptTemplate):
 
     prompt: Union[
         StringPromptTemplate,
-        list[
-            Union[StringPromptTemplate, ImagePromptTemplate, _DictMessagePromptTemplate]
-        ],
+        list[Union[StringPromptTemplate, ImagePromptTemplate, DictPromptTemplate]],
     ]
     """Prompt template."""
     additional_kwargs: dict = Field(default_factory=dict)
@@ -447,7 +445,12 @@ class _StringImageMessagePromptTemplate(BaseMessagePromptTemplate):
                 raise ValueError(msg)
             prompt = []
             for tmpl in template:
-                if isinstance(tmpl, str) or isinstance(tmpl, dict) and "text" in tmpl:
+                if (
+                    isinstance(tmpl, str)
+                    or isinstance(tmpl, dict)
+                    and "text" in tmpl
+                    and set(tmpl.keys()) <= {"type", "text"}
+                ):
                     if isinstance(tmpl, str):
                         text: str = tmpl
                     else:
@@ -457,7 +460,15 @@ class _StringImageMessagePromptTemplate(BaseMessagePromptTemplate):
                             text, template_format=template_format
                         )
                     )
-                elif isinstance(tmpl, dict) and "image_url" in tmpl:
+                elif (
+                    isinstance(tmpl, dict)
+                    and "image_url" in tmpl
+                    and set(tmpl.keys())
+                    <= {
+                        "type",
+                        "image_url",
+                    }
+                ):
                     img_template = cast("_ImageTemplateParam", tmpl)["image_url"]
                     input_variables = []
                     if isinstance(img_template, str):
@@ -503,7 +514,7 @@ class _StringImageMessagePromptTemplate(BaseMessagePromptTemplate):
                             "format."
                         )
                         raise ValueError(msg)
-                    data_template_obj = _DictMessagePromptTemplate(
+                    data_template_obj = DictPromptTemplate(
                         template=cast("dict[str, Any]", tmpl),
                         template_format=template_format,
                     )
@@ -592,7 +603,7 @@ class _StringImageMessagePromptTemplate(BaseMessagePromptTemplate):
             elif isinstance(prompt, ImagePromptTemplate):
                 formatted = prompt.format(**inputs)
                 content.append({"type": "image_url", "image_url": formatted})
-            elif isinstance(prompt, _DictMessagePromptTemplate):
+            elif isinstance(prompt, DictPromptTemplate):
                 formatted = prompt.format(**inputs)
                 content.append(formatted)
         return self._msg_class(
@@ -624,7 +635,7 @@ class _StringImageMessagePromptTemplate(BaseMessagePromptTemplate):
             elif isinstance(prompt, ImagePromptTemplate):
                 formatted = await prompt.aformat(**inputs)
                 content.append({"type": "image_url", "image_url": formatted})
-            elif isinstance(prompt, _DictMessagePromptTemplate):
+            elif isinstance(prompt, DictPromptTemplate):
                 formatted = prompt.format(**inputs)
                 content.append(formatted)
         return self._msg_class(

libs/core/langchain_core/prompts/dict.py

@@ -0,0 +1,137 @@
+"""Dict prompt template."""
+
+import warnings
+from functools import cached_property
+from typing import Any, Literal, Optional
+
+from langchain_core.load import dumpd
+from langchain_core.prompts.string import (
+    DEFAULT_FORMATTER_MAPPING,
+    get_template_variables,
+)
+from langchain_core.runnables import RunnableConfig, RunnableSerializable
+from langchain_core.runnables.config import ensure_config
+
+
+class DictPromptTemplate(RunnableSerializable[dict, dict]):
+    """Template represented by a dict.
+
+    Recognizes variables in f-string or mustache formatted string dict values. Does NOT
+    recognize variables in dict keys. Applies recursively.
+    """
+
+    template: dict[str, Any]
+    template_format: Literal["f-string", "mustache"]
+
+    @property
+    def input_variables(self) -> list[str]:
+        """Template input variables."""
+        return _get_input_variables(self.template, self.template_format)
+
+    def format(self, **kwargs: Any) -> dict[str, Any]:
+        """Format the prompt with the inputs."""
+        return _insert_input_variables(self.template, kwargs, self.template_format)
+
+    async def aformat(self, **kwargs: Any) -> dict[str, Any]:
+        """Format the prompt with the inputs."""
+        return self.format(**kwargs)
+
+    def invoke(
+        self, input: dict, config: Optional[RunnableConfig] = None, **kwargs: Any
+    ) -> dict:
+        """Invoke the prompt."""
+        return self._call_with_config(
+            lambda x: self.format(**x),
+            input,
+            ensure_config(config),
+            run_type="prompt",
+            serialized=self._serialized,
+            **kwargs,
+        )
+
+    @property
+    def _prompt_type(self) -> str:
+        return "dict-prompt"
+
+    @cached_property
+    def _serialized(self) -> dict[str, Any]:
+        return dumpd(self)
+
+    @classmethod
+    def is_lc_serializable(cls) -> bool:
+        """Return whether or not the class is serializable.
+
+        Returns: True.
+        """
+        return True
+
+    @classmethod
+    def get_lc_namespace(cls) -> list[str]:
+        """Serialization namespace."""
+        return ["langchain_core", "prompts", "dict"]
+
+    def pretty_repr(self, *, html: bool = False) -> str:
+        """Human-readable representation.
+
+        Args:
+            html: Whether to format as HTML. Defaults to False.
+
+        Returns:
+            Human-readable representation.
+        """
+        raise NotImplementedError
+
+
+def _get_input_variables(
+    template: dict, template_format: Literal["f-string", "mustache"]
+) -> list[str]:
+    input_variables = []
+    for v in template.values():
+        if isinstance(v, str):
+            input_variables += get_template_variables(v, template_format)
+        elif isinstance(v, dict):
+            input_variables += _get_input_variables(v, template_format)
+        elif isinstance(v, (list, tuple)):
+            for x in v:
+                if isinstance(x, str):
+                    input_variables += get_template_variables(x, template_format)
+                elif isinstance(x, dict):
+                    input_variables += _get_input_variables(x, template_format)
+        else:
+            pass
+    return list(set(input_variables))
+
+
+def _insert_input_variables(
+    template: dict[str, Any],
+    inputs: dict[str, Any],
+    template_format: Literal["f-string", "mustache"],
+) -> dict[str, Any]:
+    formatted = {}
+    formatter = DEFAULT_FORMATTER_MAPPING[template_format]
+    for k, v in template.items():
+        if isinstance(v, str):
+            formatted[k] = formatter(v, **inputs)
+        elif isinstance(v, dict):
+            if k == "image_url" and "path" in v:
+                msg = (
+                    "Specifying image inputs via file path in environments with "
+                    "user-input paths is a security vulnerability. Out of an abundance "
+                    "of caution, the utility has been removed to prevent possible "
+                    "misuse."
+                )
+                warnings.warn(msg, stacklevel=2)
+            formatted[k] = _insert_input_variables(v, inputs, template_format)
+        elif isinstance(v, (list, tuple)):
+            formatted_v = []
+            for x in v:
+                if isinstance(x, str):
+                    formatted_v.append(formatter(x, **inputs))
+                elif isinstance(x, dict):
+                    formatted_v.append(
+                        _insert_input_variables(x, inputs, template_format)
+                    )
+            formatted[k] = type(v)(formatted_v)
+        else:
+            formatted[k] = v
+    return formatted

libs/core/langchain_core/prompts/message.py

@@ -3,14 +3,10 @@
 from __future__ import annotations
 
 from abc import ABC, abstractmethod
-from typing import TYPE_CHECKING, Any, Literal
+from typing import TYPE_CHECKING, Any
 
 from langchain_core.load import Serializable
-from langchain_core.messages import BaseMessage, convert_to_messages
-from langchain_core.prompts.string import (
-    DEFAULT_FORMATTER_MAPPING,
-    get_template_variables,
-)
+from langchain_core.messages import BaseMessage
 from langchain_core.utils.interactive_env import is_interactive_env
 
 if TYPE_CHECKING:
@@ -98,89 +94,3 @@ class BaseMessagePromptTemplate(Serializable, ABC):
 
         prompt = ChatPromptTemplate(messages=[self])
         return prompt + other
-
-
-class _DictMessagePromptTemplate(BaseMessagePromptTemplate):
-    """Template represented by a dict that recursively fills input vars in string vals.
-
-    Special handling of image_url dicts to load local paths. These look like:
-    ``{"type": "image_url", "image_url": {"path": "..."}}``
-    """
-
-    template: dict[str, Any]
-    template_format: Literal["f-string", "mustache"]
-
-    def format_messages(self, **kwargs: Any) -> list[BaseMessage]:
-        msg_dict = _insert_input_variables(self.template, kwargs, self.template_format)
-        return convert_to_messages([msg_dict])
-
-    @property
-    def input_variables(self) -> list[str]:
-        return _get_input_variables(self.template, self.template_format)
-
-    @property
-    def _prompt_type(self) -> str:
-        return "message-dict-prompt"
-
-    @classmethod
-    def get_lc_namespace(cls) -> list[str]:
-        return ["langchain_core", "prompts", "message"]
-
-    def format(
-        self,
-        **kwargs: Any,
-    ) -> dict[str, Any]:
-        """Format the prompt with the inputs."""
-        return _insert_input_variables(self.template, kwargs, self.template_format)
-
-
-def _get_input_variables(
-    template: dict, template_format: Literal["f-string", "mustache"]
-) -> list[str]:
-    input_variables = []
-    for v in template.values():
-        if isinstance(v, str):
-            input_variables += get_template_variables(v, template_format)
-        elif isinstance(v, dict):
-            input_variables += _get_input_variables(v, template_format)
-        elif isinstance(v, (list, tuple)):
-            for x in v:
-                if isinstance(x, str):
-                    input_variables += get_template_variables(x, template_format)
-                elif isinstance(x, dict):
-                    input_variables += _get_input_variables(x, template_format)
-    return list(set(input_variables))
-
-
-def _insert_input_variables(
-    template: dict[str, Any],
-    inputs: dict[str, Any],
-    template_format: Literal["f-string", "mustache"],
-) -> dict[str, Any]:
-    formatted = {}
-    formatter = DEFAULT_FORMATTER_MAPPING[template_format]
-    for k, v in template.items():
-        if isinstance(v, str):
-            formatted[k] = formatter(v, **inputs)
-        elif isinstance(v, dict):
-            # No longer support loading local images.
-            if k == "image_url" and "path" in v:
-                msg = (
-                    "Specifying image inputs via file path in environments with "
-                    "user-input paths is a security vulnerability. Out of an abundance "
-                    "of caution, the utility has been removed to prevent possible "
-                    "misuse."
-                )
-                raise ValueError(msg)
-            formatted[k] = _insert_input_variables(v, inputs, template_format)
-        elif isinstance(v, (list, tuple)):
-            formatted_v = []
-            for x in v:
-                if isinstance(x, str):
-                    formatted_v.append(formatter(x, **inputs))
-                elif isinstance(x, dict):
-                    formatted_v.append(
-                        _insert_input_variables(x, inputs, template_format)
-                    )
-            formatted[k] = type(v)(formatted_v)
-    return formatted

libs/core/tests/unit_tests/prompts/__snapshots__/test_chat.ambr

@@ -3135,6 +3135,27 @@
                 'name': 'PromptTemplate',
                 'type': 'constructor',
               }),
+              dict({
+                'id': list([
+                  'langchain_core',
+                  'prompts',
+                  'dict',
+                  'DictPromptTemplate',
+                ]),
+                'kwargs': dict({
+                  'template': dict({
+                    'cache_control': dict({
+                      'type': '{foo}',
+                    }),
+                    'text': "What's in this image?",
+                    'type': 'text',
+                  }),
+                  'template_format': 'f-string',
+                }),
+                'lc': 1,
+                'name': 'DictPromptTemplate',
+                'type': 'constructor',
+              }),
               dict({
                 'id': list([
                   'langchain',

libs/core/tests/unit_tests/prompts/test_chat.py

@@ -973,6 +973,11 @@ def test_chat_tmpl_serdes(snapshot: SnapshotAssertion) -> None:
                         "hello",
                         {"text": "What's in this image?"},
                         {"type": "text", "text": "What's in this image?"},
+                        {
+                            "type": "text",
+                            "text": "What's in this image?",
+                            "cache_control": {"type": "{foo}"},
+                        },
                         {
                             "type": "image_url",
                             "image_url": "data:image/jpeg;base64,{my_image}",
@@ -1012,7 +1017,7 @@ def test_chat_tmpl_serdes(snapshot: SnapshotAssertion) -> None:
 @pytest.mark.xfail(
     reason=(
         "In a breaking release, we can update `_convert_to_message_template` to use "
-        "_DictMessagePromptTemplate for all `dict` inputs, allowing for templatization "
+        "DictPromptTemplate for all `dict` inputs, allowing for templatization "
         "of message attributes outside content blocks. That would enable the below "
         "test to pass."
     )

libs/core/tests/unit_tests/prompts/test_dict.py

@@ -0,0 +1,34 @@
+from langchain_core.load import load
+from langchain_core.prompts.dict import DictPromptTemplate
+
+
+def test__dict_message_prompt_template_fstring() -> None:
+    template = {
+        "type": "text",
+        "text": "{text1}",
+        "cache_control": {"type": "{cache_type}"},
+    }
+    prompt = DictPromptTemplate(template=template, template_format="f-string")
+    expected = {
+        "type": "text",
+        "text": "important message",
+        "cache_control": {"type": "ephemeral"},
+    }
+    actual = prompt.format(text1="important message", cache_type="ephemeral")
+    assert actual == expected
+
+
+def test_deserialize_legacy() -> None:
+    ser = {
+        "type": "constructor",
+        "lc": 1,
+        "id": ["langchain_core", "prompts", "message", "_DictMessagePromptTemplate"],
+        "kwargs": {
+            "template_format": "f-string",
+            "template": {"type": "audio", "audio": "{audio_data}"},
+        },
+    }
+    expected = DictPromptTemplate(
+        template={"type": "audio", "audio": "{audio_data}"}, template_format="f-string"
+    )
+    assert load(ser) == expected

libs/core/tests/unit_tests/prompts/test_imports.py

@@ -6,6 +6,7 @@ EXPECTED_ALL = [
     "BasePromptTemplate",
     "ChatMessagePromptTemplate",
     "ChatPromptTemplate",
+    "DictPromptTemplate",
     "FewShotPromptTemplate",
     "FewShotPromptWithTemplates",
     "FewShotChatMessagePromptTemplate",

libs/core/tests/unit_tests/prompts/test_message.py

@@ -1,61 +0,0 @@
-from pathlib import Path
-
-from langchain_core.messages import AIMessage, BaseMessage, ToolMessage
-from langchain_core.prompts.message import _DictMessagePromptTemplate
-
-CUR_DIR = Path(__file__).parent.absolute().resolve()
-
-
-def test__dict_message_prompt_template_fstring() -> None:
-    template = {
-        "role": "assistant",
-        "content": [
-            {"type": "text", "text": "{text1}", "cache_control": {"type": "ephemeral"}},
-        ],
-        "name": "{name1}",
-        "tool_calls": [
-            {
-                "name": "{tool_name1}",
-                "args": {"arg1": "{tool_arg1}"},
-                "id": "1",
-                "type": "tool_call",
-            }
-        ],
-    }
-    prompt = _DictMessagePromptTemplate(template=template, template_format="f-string")
-    expected: BaseMessage = AIMessage(
-        [
-            {
-                "type": "text",
-                "text": "important message",
-                "cache_control": {"type": "ephemeral"},
-            },
-        ],
-        name="foo",
-        tool_calls=[
-            {
-                "name": "do_stuff",
-                "args": {"arg1": "important arg1"},
-                "id": "1",
-                "type": "tool_call",
-            }
-        ],
-    )
-    actual = prompt.format_messages(
-        text1="important message",
-        name1="foo",
-        tool_arg1="important arg1",
-        tool_name1="do_stuff",
-    )[0]
-    assert actual == expected
-
-    template = {
-        "role": "tool",
-        "content": "{content1}",
-        "tool_call_id": "1",
-        "name": "{name1}",
-    }
-    prompt = _DictMessagePromptTemplate(template=template, template_format="f-string")
-    expected = ToolMessage("foo", name="bar", tool_call_id="1")
-    actual = prompt.format_messages(content1="foo", name1="bar")[0]
-    assert actual == expected