feat: add csv output parser (#3483)

* Adds the CSV Output Parser

---------

Co-authored-by: autofix-ci[bot] <114827586+autofix-ci[bot]@users.noreply.github.com>
Co-authored-by: Gabriel Luiz Freitas Almeida <gabriel@langflow.org>

src/backend/base/langflow/base/models/model.py

@@ -4,6 +4,7 @@ from abc import abstractmethod
 
 from langchain_core.language_models.llms import LLM
 from langchain_core.messages import AIMessage, BaseMessage, HumanMessage, SystemMessage
+from langchain_core.output_parsers import BaseOutputParser
 
 from langflow.base.constants import STREAM_INFO_TEXT
 from langflow.custom import Component
@@ -19,6 +20,9 @@ class LCModelComponent(Component):
     description: str = "Model Description"
     trace_type = "llm"
 
+    # Optional output parser to pass to the runnable. Subclasses may allow the user to input an `output_parser`
+    output_parser: BaseOutputParser | None = None
+
     _base_inputs: list[InputTypes] = [
         MessageInput(name="input_value", display_name="Input"),
         MessageTextInput(
@@ -162,6 +166,9 @@ class LCModelComponent(Component):
             messages.append(SystemMessage(content=system_message))
         inputs: list | dict = messages or {}
         try:
+            if self.output_parser is not None:
+                runnable = runnable | self.output_parser
+
             runnable = runnable.with_config(  # type: ignore
                 {
                     "run_name": self.display_name,

src/backend/base/langflow/base/prompts/api_utils.py

@@ -220,20 +220,4 @@ def process_prompt_template(
     # Update the input variables field in the template
     update_input_variables_field(input_variables, frontend_node_template)
 
-    # Optional: cleanup fields based on specific conditions
-    cleanup_prompt_template_fields(input_variables, frontend_node_template)
-
     return input_variables
-
-
-def cleanup_prompt_template_fields(input_variables, template):
-    """Removes unused fields if the conditions are met in the template."""
-    prompt_fields = [
-        key for key, field in template.items() if isinstance(field, dict) and field.get("type") == "prompt"
-    ]
-
-    if len(prompt_fields) == 1:
-        for key in list(template.keys()):  # Use list to copy keys
-            field = template.get(key, {})
-            if isinstance(field, dict) and field.get("type") != "code" and key not in input_variables + prompt_fields:
-                del template[key]

src/backend/base/langflow/components/__init__.py

@@ -31,6 +31,7 @@ __all__ = [
     "link_extractors",
     "memories",
     "outputs",
+    "output_parsers",
     "retrievers",
     "textsplitters",
     "toolkits",

src/backend/base/langflow/components/models/AIMLModel.py

@@ -13,6 +13,7 @@ from langflow.inputs import (
     SecretStrInput,
     StrInput,
 )
+from langflow.inputs.inputs import HandleInput
 
 
 class AIMLModelComponent(LCModelComponent):
@@ -59,6 +60,13 @@ class AIMLModelComponent(LCModelComponent):
             advanced=True,
             value=1,
         ),
+        HandleInput(
+            name="output_parser",
+            display_name="Output Parser",
+            info="The parser to use to parse the output of the model",
+            advanced=True,
+            input_types=["OutputParser"],
+        ),
     ]
 
     def build_model(self) -> LanguageModel:  # type: ignore[type-var]

src/backend/base/langflow/components/models/AmazonBedrockModel.py

@@ -1,6 +1,7 @@
 from langflow.base.models.model import LCModelComponent
 from langflow.field_typing import LanguageModel
 from langflow.inputs import MessageTextInput, SecretStrInput
+from langflow.inputs.inputs import HandleInput
 from langflow.io import DictInput, DropdownInput
 
 
@@ -55,6 +56,13 @@ class AmazonBedrockComponent(LCModelComponent):
         MessageTextInput(name="region_name", display_name="Region Name", value="us-east-1"),
         DictInput(name="model_kwargs", display_name="Model Kwargs", advanced=True, is_list=True),
         MessageTextInput(name="endpoint_url", display_name="Endpoint URL", advanced=True),
+        HandleInput(
+            name="output_parser",
+            display_name="Output Parser",
+            info="The parser to use to parse the output of the model",
+            advanced=True,
+            input_types=["OutputParser"],
+        ),
     ]
 
     def build_model(self) -> LanguageModel:  # type: ignore[type-var]

src/backend/base/langflow/components/models/AnthropicModel.py

@@ -2,6 +2,7 @@ from pydantic.v1 import SecretStr
 
 from langflow.base.models.model import LCModelComponent
 from langflow.field_typing import LanguageModel
+from langflow.inputs.inputs import HandleInput
 from langflow.io import DropdownInput, FloatInput, IntInput, MessageTextInput, SecretStrInput
 
 
@@ -49,6 +50,13 @@ class AnthropicModelComponent(LCModelComponent):
             info="Prefill text to guide the model's response.",
             advanced=True,
         ),
+        HandleInput(
+            name="output_parser",
+            display_name="Output Parser",
+            info="The parser to use to parse the output of the model",
+            advanced=True,
+            input_types=["OutputParser"],
+        ),
     ]
 
     def build_model(self) -> LanguageModel:  # type: ignore[type-var]

src/backend/base/langflow/components/models/AzureOpenAIModel.py

@@ -3,6 +3,7 @@ from langchain_openai import AzureChatOpenAI
 from langflow.base.models.model import LCModelComponent
 from langflow.field_typing import LanguageModel
 from langflow.inputs import MessageTextInput
+from langflow.inputs.inputs import HandleInput
 from langflow.io import DropdownInput, FloatInput, IntInput, SecretStrInput
 
 
@@ -48,6 +49,13 @@ class AzureChatOpenAIComponent(LCModelComponent):
             advanced=True,
             info="The maximum number of tokens to generate. Set to 0 for unlimited tokens.",
         ),
+        HandleInput(
+            name="output_parser",
+            display_name="Output Parser",
+            info="The parser to use to parse the output of the model",
+            advanced=True,
+            input_types=["OutputParser"],
+        ),
     ]
 
     def build_model(self) -> LanguageModel:  # type: ignore[type-var]

src/backend/base/langflow/components/models/BaiduQianfanChatModel.py

@@ -3,6 +3,7 @@ from pydantic.v1 import SecretStr
 
 from langflow.base.models.model import LCModelComponent
 from langflow.field_typing.constants import LanguageModel
+from langflow.inputs.inputs import HandleInput
 from langflow.io import DropdownInput, FloatInput, MessageTextInput, SecretStrInput
 
 
@@ -67,6 +68,13 @@ class QianfanChatEndpointComponent(LCModelComponent):
             display_name="Endpoint",
             info="Endpoint of the Qianfan LLM, required if custom model used.",
         ),
+        HandleInput(
+            name="output_parser",
+            display_name="Output Parser",
+            info="The parser to use to parse the output of the model",
+            advanced=True,
+            input_types=["OutputParser"],
+        ),
     ]
 
     def build_model(self) -> LanguageModel:  # type: ignore[type-var]

src/backend/base/langflow/components/models/CohereModel.py

@@ -3,6 +3,7 @@ from pydantic.v1 import SecretStr
 
 from langflow.base.models.model import LCModelComponent
 from langflow.field_typing import LanguageModel
+from langflow.inputs.inputs import HandleInput
 from langflow.io import FloatInput, SecretStrInput
 
 
@@ -22,6 +23,13 @@ class CohereComponent(LCModelComponent):
             value="COHERE_API_KEY",
         ),
         FloatInput(name="temperature", display_name="Temperature", value=0.75),
+        HandleInput(
+            name="output_parser",
+            display_name="Output Parser",
+            info="The parser to use to parse the output of the model",
+            advanced=True,
+            input_types=["OutputParser"],
+        ),
     ]
 
     def build_model(self) -> LanguageModel:  # type: ignore[type-var]

src/backend/base/langflow/components/models/GoogleGenerativeAIModel.py

@@ -3,6 +3,7 @@ from pydantic.v1 import SecretStr
 from langflow.base.models.model import LCModelComponent
 from langflow.field_typing import LanguageModel
 from langflow.inputs import DropdownInput, FloatInput, IntInput, SecretStrInput
+from langflow.inputs.inputs import HandleInput
 
 
 class GoogleGenerativeAIComponent(LCModelComponent):
@@ -48,6 +49,13 @@ class GoogleGenerativeAIComponent(LCModelComponent):
             info="Decode using top-k sampling: consider the set of top_k most probable tokens. Must be positive.",
             advanced=True,
         ),
+        HandleInput(
+            name="output_parser",
+            display_name="Output Parser",
+            info="The parser to use to parse the output of the model",
+            advanced=True,
+            input_types=["OutputParser"],
+        ),
     ]
 
     def build_model(self) -> LanguageModel:  # type: ignore[type-var]

src/backend/base/langflow/components/models/GroqModel.py

@@ -4,6 +4,7 @@ from pydantic.v1 import SecretStr
 
 from langflow.base.models.model import LCModelComponent
 from langflow.field_typing import LanguageModel
+from langflow.inputs.inputs import HandleInput
 from langflow.io import DropdownInput, FloatInput, IntInput, MessageTextInput, SecretStrInput
 
 
@@ -51,6 +52,13 @@ class GroqModel(LCModelComponent):
             options=[],
             refresh_button=True,
         ),
+        HandleInput(
+            name="output_parser",
+            display_name="Output Parser",
+            info="The parser to use to parse the output of the model",
+            advanced=True,
+            input_types=["OutputParser"],
+        ),
     ]
 
     def get_models(self) -> list[str]:

src/backend/base/langflow/components/models/HuggingFaceModel.py

@@ -6,7 +6,8 @@ from tenacity import retry, stop_after_attempt, wait_fixed
 # TODO: langchain_community.llms.huggingface_endpoint is depreciated. Need to update to langchain_huggingface, but have dependency with langchain_core 0.3.0
 from langflow.base.models.model import LCModelComponent
 from langflow.field_typing import LanguageModel
-from langflow.io import DictInput, DropdownInput, FloatInput, IntInput, SecretStrInput, StrInput
+from langflow.inputs.inputs import HandleInput
+from langflow.io import DictInput, DropdownInput, IntInput, SecretStrInput, StrInput
 
 
 class HuggingFaceEndpointsComponent(LCModelComponent):
@@ -37,41 +38,12 @@ class HuggingFaceEndpointsComponent(LCModelComponent):
         SecretStrInput(name="huggingfacehub_api_token", display_name="API Token", password=True),
         DictInput(name="model_kwargs", display_name="Model Keyword Arguments", advanced=True),
         IntInput(name="retry_attempts", display_name="Retry Attempts", value=1, advanced=True),
-        IntInput(
-            name="max_new_tokens", display_name="Max New Tokens", value=512, info="Maximum number of generated tokens"
-        ),
-        IntInput(
-            name="top_k",
-            display_name="Top K",
+        HandleInput(
+            name="output_parser",
+            display_name="Output Parser",
+            info="The parser to use to parse the output of the model",
             advanced=True,
-            info="The number of highest probability vocabulary tokens to keep for top-k-filtering",
-        ),
-        FloatInput(
-            name="top_p",
-            display_name="Top P",
-            value=0.95,
-            advanced=True,
-            info="If set to < 1, only the smallest set of most probable tokens with probabilities that add up to `top_p` or higher are kept for generation",
-        ),
-        FloatInput(
-            name="typical_p",
-            display_name="Typical P",
-            value=0.95,
-            advanced=True,
-            info="Typical Decoding mass.",
-        ),
-        FloatInput(
-            name="temperature",
-            display_name="Temperature",
-            value=0.8,
-            advanced=True,
-            info="The value used to module the logits distribution",
-        ),
-        FloatInput(
-            name="repetition_penalty",
-            display_name="Repetition Penalty",
-            advanced=True,
-            info="The parameter for repetition penalty. 1.0 means no penalty.",
+            input_types=["OutputParser"],
         ),
     ]
 

src/backend/base/langflow/components/models/Maritalk.py

@@ -4,6 +4,7 @@ from langflow.base.models.model import LCModelComponent
 from langflow.field_typing import LanguageModel
 from langflow.field_typing.range_spec import RangeSpec
 from langflow.inputs import DropdownInput, FloatInput, IntInput, SecretStrInput
+from langflow.inputs.inputs import HandleInput
 
 
 class MaritalkModelComponent(LCModelComponent):
@@ -33,6 +34,13 @@ class MaritalkModelComponent(LCModelComponent):
             advanced=False,
         ),
         FloatInput(name="temperature", display_name="Temperature", value=0.1, range_spec=RangeSpec(min=0, max=1)),
+        HandleInput(
+            name="output_parser",
+            display_name="Output Parser",
+            info="The parser to use to parse the output of the model",
+            advanced=True,
+            input_types=["OutputParser"],
+        ),
     ]
 
     def build_model(self) -> LanguageModel:  # type: ignore[type-var]

src/backend/base/langflow/components/models/MistralModel.py

@@ -3,6 +3,7 @@ from pydantic.v1 import SecretStr
 
 from langflow.base.models.model import LCModelComponent
 from langflow.field_typing import LanguageModel
+from langflow.inputs.inputs import HandleInput
 from langflow.io import BoolInput, DropdownInput, FloatInput, IntInput, SecretStrInput, StrInput
 
 
@@ -55,6 +56,13 @@ class MistralAIModelComponent(LCModelComponent):
         FloatInput(name="top_p", display_name="Top P", advanced=True, value=1),
         IntInput(name="random_seed", display_name="Random Seed", value=1, advanced=True),
         BoolInput(name="safe_mode", display_name="Safe Mode", advanced=True),
+        HandleInput(
+            name="output_parser",
+            display_name="Output Parser",
+            info="The parser to use to parse the output of the model",
+            advanced=True,
+            input_types=["OutputParser"],
+        ),
     ]
 
     def build_model(self) -> LanguageModel:  # type: ignore[type-var]

src/backend/base/langflow/components/models/NvidiaModel.py

@@ -3,6 +3,7 @@ from typing import Any
 from langflow.base.models.model import LCModelComponent
 from langflow.field_typing import LanguageModel
 from langflow.inputs import DropdownInput, FloatInput, IntInput, SecretStrInput, StrInput
+from langflow.inputs.inputs import HandleInput
 from langflow.schema.dotdict import dotdict
 
 
@@ -47,6 +48,13 @@ class NVIDIAModelComponent(LCModelComponent):
             advanced=True,
             value=1,
         ),
+        HandleInput(
+            name="output_parser",
+            display_name="Output Parser",
+            info="The parser to use to parse the output of the model",
+            advanced=True,
+            input_types=["OutputParser"],
+        ),
     ]
 
     def update_build_config(self, build_config: dotdict, field_value: Any, field_name: str | None = None):

src/backend/base/langflow/components/models/OllamaModel.py

@@ -6,6 +6,7 @@ from langchain_community.chat_models import ChatOllama
 
 from langflow.base.models.model import LCModelComponent
 from langflow.field_typing import LanguageModel
+from langflow.inputs.inputs import HandleInput
 from langflow.io import BoolInput, DictInput, DropdownInput, FloatInput, IntInput, StrInput
 
 
@@ -204,6 +205,13 @@ class ChatOllamaComponent(LCModelComponent):
             info="Template to use for generating text.",
             advanced=True,
         ),
+        HandleInput(
+            name="output_parser",
+            display_name="Output Parser",
+            info="The parser to use to parse the output of the model",
+            advanced=True,
+            input_types=["OutputParser"],
+        ),
     ]
 
     def build_model(self) -> LanguageModel:  # type: ignore[type-var]

src/backend/base/langflow/components/models/OpenAIModel.py

@@ -17,6 +17,7 @@ from langflow.inputs import (
     SecretStrInput,
     StrInput,
 )
+from langflow.inputs.inputs import HandleInput
 
 
 class OpenAIModelComponent(LCModelComponent):
@@ -75,6 +76,13 @@ class OpenAIModelComponent(LCModelComponent):
             advanced=True,
             value=1,
         ),
+        HandleInput(
+            name="output_parser",
+            display_name="Output Parser",
+            info="The parser to use to parse the output of the model",
+            advanced=True,
+            input_types=["OutputParser"],
+        ),
     ]
 
     def build_model(self) -> LanguageModel:  # type: ignore[type-var]

src/backend/base/langflow/components/models/PerplexityModel.py

@@ -3,6 +3,7 @@ from pydantic.v1 import SecretStr
 
 from langflow.base.models.model import LCModelComponent
 from langflow.field_typing import LanguageModel
+from langflow.inputs.inputs import HandleInput
 from langflow.io import DropdownInput, FloatInput, IntInput, SecretStrInput
 
 
@@ -59,6 +60,13 @@ class PerplexityComponent(LCModelComponent):
             info="Decode using top-k sampling: consider the set of top_k most probable tokens. Must be positive.",
             advanced=True,
         ),
+        HandleInput(
+            name="output_parser",
+            display_name="Output Parser",
+            info="The parser to use to parse the output of the model",
+            advanced=True,
+            input_types=["OutputParser"],
+        ),
     ]
 
     def build_model(self) -> LanguageModel:  # type: ignore[type-var]

src/backend/base/langflow/components/models/VertexAiModel.py

@@ -3,6 +3,7 @@ from typing import cast
 from langflow.base.models.model import LCModelComponent
 from langflow.field_typing import LanguageModel
 from langflow.inputs import MessageTextInput
+from langflow.inputs.inputs import HandleInput
 from langflow.io import BoolInput, FileInput, FloatInput, IntInput, StrInput
 
 
@@ -28,6 +29,13 @@ class ChatVertexAIComponent(LCModelComponent):
         IntInput(name="top_k", display_name="Top K", advanced=True),
         FloatInput(name="top_p", display_name="Top P", value=0.95, advanced=True),
         BoolInput(name="verbose", display_name="Verbose", value=False, advanced=True),
+        HandleInput(
+            name="output_parser",
+            display_name="Output Parser",
+            info="The parser to use to parse the output of the model",
+            advanced=True,
+            input_types=["OutputParser"],
+        ),
     ]
 
     def build_model(self) -> LanguageModel:

src/backend/base/langflow/components/output_parsers/CSVOutputParser.py

@@ -0,0 +1,31 @@
+from langchain_core.output_parsers import CommaSeparatedListOutputParser
+
+from langflow.custom.custom_component.component import Component
+from langflow.field_typing.constants import OutputParser
+from langflow.io import Output
+from langflow.schema.message import Message
+
+
+class CSVOutputParserComponent(Component):
+    display_name = "CSV Output Parser"
+    description = "Use with an LLM to return a comma separated list."
+    icon = "type"
+    name = "CSVOutputParser"
+
+    inputs = []  # no inputs necessary
+
+    outputs = [
+        Output(
+            display_name="Format Instructions",
+            name="format_instructions",
+            info="Pass to a prompt template to include formatting instructions for LLM responses.",
+            method="format_instructions",
+        ),
+        Output(display_name="Output Parser", name="output_parser", method="build_parser"),
+    ]
+
+    def build_parser(self) -> OutputParser:
+        return CommaSeparatedListOutputParser()
+
+    def format_instructions(self) -> Message:
+        return Message(text=CommaSeparatedListOutputParser().get_format_instructions())

src/backend/base/langflow/components/output_parsers/__init__.py

@@ -0,0 +1,3 @@
+from .CSVOutputParser import CSVOutputParserComponent
+
+__all__ = ["CSVOutputParserComponent"]

src/backend/base/langflow/field_typing/constants.py

@@ -11,7 +11,7 @@ from langchain_core.embeddings import Embeddings
 from langchain_core.language_models import BaseLanguageModel, BaseLLM
 from langchain_core.language_models.chat_models import BaseChatModel
 from langchain_core.memory import BaseMemory
-from langchain_core.output_parsers import BaseOutputParser
+from langchain_core.output_parsers import BaseLLMOutputParser, BaseOutputParser
 from langchain_core.prompts import BasePromptTemplate, ChatPromptTemplate, PromptTemplate
 from langchain_core.retrievers import BaseRetriever
 from langchain_core.tools import BaseTool, Tool
@@ -29,6 +29,11 @@ Retriever = TypeVar(
     BaseRetriever,
     VectorStoreRetriever,
 )
+OutputParser = TypeVar(
+    "OutputParser",
+    BaseOutputParser,
+    BaseLLMOutputParser,
+)
 
 
 class Object:

src/backend/base/langflow/schema/message.py

@@ -235,7 +235,8 @@ class Message(Data):
         if contents:
             message = HumanMessage(content=[{"type": "text", "text": text}] + contents)
 
-        prompt_template = ChatPromptTemplate(messages=[message])  # type: ignore
+        prompt_template = ChatPromptTemplate.from_messages([message])  # type: ignore
+
         instance.prompt = jsonable_encoder(prompt_template.to_json())
         instance.messages = instance.prompt.get("kwargs", {}).get("messages", [])
         return instance

src/backend/base/langflow/template/template/base.py

@@ -56,9 +56,11 @@ class Template(BaseModel):
                     _input = Input(**value)
 
                 data["fields"].append(_input)
-        # Handles components with no inputs
+
+        # Necessary for components with no inputs(?)
         if "fields" not in data:
             data["fields"] = []
+
         return cls(**data)
 
     # For backwards compatibility

src/backend/tests/integration/components/output_parsers/test_csv_output_parser.py

@@ -0,0 +1,40 @@
+import os
+import pytest
+
+from langflow.components.models.OpenAIModel import OpenAIModelComponent
+from langflow.components.output_parsers.CSVOutputParser import CSVOutputParserComponent
+from langflow.components.prompts.Prompt import PromptComponent
+from tests.integration.utils import ComponentInputHandle, run_single_component
+
+
+@pytest.mark.asyncio
+@pytest.mark.api_key_required
+async def test_csv_output_parser_openai():
+    format_instructions = ComponentInputHandle(
+        clazz=CSVOutputParserComponent,
+        inputs={},
+        output_name="format_instructions",
+    )
+    output_parser_handle = ComponentInputHandle(
+        clazz=CSVOutputParserComponent,
+        inputs={},
+        output_name="output_parser",
+    )
+    prompt_handler = ComponentInputHandle(
+        clazz=PromptComponent,
+        inputs={
+            "template": "List the first five positive integers.\n\n{format_instructions}",
+            "format_instructions": format_instructions,
+        },
+        output_name="prompt",
+    )
+
+    outputs = await run_single_component(
+        OpenAIModelComponent,
+        inputs={
+            "api_key": os.environ["OPENAI_API_KEY"],
+            "output_parser": output_parser_handle,
+            "input_value": prompt_handler,
+        },
+    )
+    assert outputs["text_output"] == "1, 2, 3, 4, 5"