Merge branch 'main' into e-300

api/core/workflow/nodes/document_extractor/node.py

@@ -223,8 +223,8 @@ def _extract_text_from_doc(file_content: bytes) -> str:
     """
     from unstructured.partition.api import partition_via_api
 
-    if not (dify_config.UNSTRUCTURED_API_URL and dify_config.UNSTRUCTURED_API_KEY):
-        raise TextExtractionError("UNSTRUCTURED_API_URL and UNSTRUCTURED_API_KEY must be set")
+    if not dify_config.UNSTRUCTURED_API_URL:
+        raise TextExtractionError("UNSTRUCTURED_API_URL must be set")
 
     try:
         with tempfile.NamedTemporaryFile(suffix=".doc", delete=False) as temp_file:
@@ -235,7 +235,7 @@ def _extract_text_from_doc(file_content: bytes) -> str:
                     file=file,
                     metadata_filename=temp_file.name,
                     api_url=dify_config.UNSTRUCTURED_API_URL,
-                    api_key=dify_config.UNSTRUCTURED_API_KEY,
+                    api_key=dify_config.UNSTRUCTURED_API_KEY,  # type: ignore
                 )
             os.unlink(temp_file.name)
         return "\n".join([getattr(element, "text", "") for element in elements])

api/core/workflow/nodes/http_request/executor.py

@@ -262,7 +262,10 @@ class Executor:
                 headers[authorization.config.header] = f"Bearer {authorization.config.api_key}"
             elif self.auth.config.type == "basic":
                 credentials = authorization.config.api_key
-                encoded_credentials = base64.b64encode(credentials.encode("utf-8")).decode("utf-8")
+                if ":" in credentials:
+                    encoded_credentials = base64.b64encode(credentials.encode("utf-8")).decode("utf-8")
+                else:
+                    encoded_credentials = credentials
                 headers[authorization.config.header] = f"Basic {encoded_credentials}"
             elif self.auth.config.type == "custom":
                 headers[authorization.config.header] = authorization.config.api_key or ""

api/core/workflow/nodes/http_request/node.py

@@ -191,8 +191,9 @@ class HttpRequestNode(BaseNode[HttpRequestNodeData]):
         mime_type = (
             content_disposition_type or content_type or mimetypes.guess_type(filename)[0] or "application/octet-stream"
         )
+        tool_file_manager = ToolFileManager()
 
-        tool_file = ToolFileManager.create_file_by_raw(
+        tool_file = tool_file_manager.create_file_by_raw(
             user_id=self.user_id,
             tenant_id=self.tenant_id,
             conversation_id=None,

api/core/workflow/nodes/knowledge_retrieval/knowledge_retrieval_node.py

@@ -6,7 +6,7 @@ from collections import defaultdict
 from collections.abc import Mapping, Sequence
 from typing import Any, Optional, cast
 
-from sqlalchemy import Integer, and_, func, or_, text
+from sqlalchemy import Float, and_, func, or_, text
 from sqlalchemy import cast as sqlalchemy_cast
 
 from core.app.app_config.entities import DatasetRetrieveConfigEntity
@@ -32,11 +32,11 @@ from core.workflow.nodes.knowledge_retrieval.template_prompts import (
     METADATA_FILTER_COMPLETION_PROMPT,
     METADATA_FILTER_SYSTEM_PROMPT,
     METADATA_FILTER_USER_PROMPT_1,
+    METADATA_FILTER_USER_PROMPT_2,
     METADATA_FILTER_USER_PROMPT_3,
 )
 from core.workflow.nodes.llm.entities import LLMNodeChatModelMessage, LLMNodeCompletionModelPromptTemplate
 from core.workflow.nodes.llm.node import LLMNode
-from core.workflow.nodes.question_classifier.template_prompts import QUESTION_CLASSIFIER_USER_PROMPT_2
 from extensions.ext_database import db
 from extensions.ext_redis import redis_client
 from libs.json_in_md_parser import parse_and_check_json_markdown
@@ -493,24 +493,24 @@ class KnowledgeRetrievalNode(LLMNode):
                 if isinstance(value, str):
                     filters.append(Document.doc_metadata[metadata_name] == f'"{value}"')
                 else:
-                    filters.append(sqlalchemy_cast(Document.doc_metadata[metadata_name].astext, Integer) == value)
+                    filters.append(sqlalchemy_cast(Document.doc_metadata[metadata_name].astext, Float) == value)
             case "is not" | "≠":
                 if isinstance(value, str):
                     filters.append(Document.doc_metadata[metadata_name] != f'"{value}"')
                 else:
-                    filters.append(sqlalchemy_cast(Document.doc_metadata[metadata_name].astext, Integer) != value)
+                    filters.append(sqlalchemy_cast(Document.doc_metadata[metadata_name].astext, Float) != value)
             case "empty":
                 filters.append(Document.doc_metadata[metadata_name].is_(None))
             case "not empty":
                 filters.append(Document.doc_metadata[metadata_name].isnot(None))
             case "before" | "<":
-                filters.append(sqlalchemy_cast(Document.doc_metadata[metadata_name].astext, Integer) < value)
+                filters.append(sqlalchemy_cast(Document.doc_metadata[metadata_name].astext, Float) < value)
             case "after" | ">":
-                filters.append(sqlalchemy_cast(Document.doc_metadata[metadata_name].astext, Integer) > value)
-            case "≤" | ">=":
-                filters.append(sqlalchemy_cast(Document.doc_metadata[metadata_name].astext, Integer) <= value)
+                filters.append(sqlalchemy_cast(Document.doc_metadata[metadata_name].astext, Float) > value)
+            case "≤" | "<=":
+                filters.append(sqlalchemy_cast(Document.doc_metadata[metadata_name].astext, Float) <= value)
             case "≥" | ">=":
-                filters.append(sqlalchemy_cast(Document.doc_metadata[metadata_name].astext, Integer) >= value)
+                filters.append(sqlalchemy_cast(Document.doc_metadata[metadata_name].astext, Float) >= value)
             case _:
                 pass
         return filters
@@ -618,7 +618,7 @@ class KnowledgeRetrievalNode(LLMNode):
             )
             prompt_messages.append(assistant_prompt_message_1)
             user_prompt_message_2 = LLMNodeChatModelMessage(
-                role=PromptMessageRole.USER, text=QUESTION_CLASSIFIER_USER_PROMPT_2
+                role=PromptMessageRole.USER, text=METADATA_FILTER_USER_PROMPT_2
             )
             prompt_messages.append(user_prompt_message_2)
             assistant_prompt_message_2 = LLMNodeChatModelMessage(

api/core/workflow/nodes/knowledge_retrieval/template_prompts.py

@@ -2,7 +2,7 @@ METADATA_FILTER_SYSTEM_PROMPT = """
     ### Job Description',
     You are a text metadata extract engine that extract text's metadata based on user input and set the metadata value
     ### Task
-    Your task is to ONLY extract the metadatas that exist in the input text from the provided metadata list and Use the following operators ["=", "!=", ">", "<", ">=", "<="] to express logical relationships, then return result in JSON format with the key "metadata_fields" and value "metadata_field_value" and comparison operator "comparison_operator".
+    Your task is to ONLY extract the metadatas that exist in the input text from the provided metadata list and Use the following operators ["contains", "not contains", "start with", "end with", "is", "is not", "empty", "not empty", "=", "≠", ">", "<", "≥", "≤", "before", "after"] to express logical relationships, then return result in JSON format with the key "metadata_fields" and value "metadata_field_value" and comparison operator "comparison_operator".
     ### Format
     The input text is in the variable input_text. Metadata are specified as a list in the variable metadata_fields.
     ### Constraint
@@ -50,7 +50,7 @@ You are a text metadata extract engine that extract text's metadata based on use
 # Your task is to ONLY extract the metadatas that exist in the input text from the provided metadata list and Use the following operators ["=", "!=", ">", "<", ">=", "<="] to express logical relationships, then return result in JSON format with the key "metadata_fields" and value "metadata_field_value" and comparison operator "comparison_operator".
 ### Format
 The input text is in the variable input_text. Metadata are specified as a list in the variable metadata_fields.
-### Constraint 
+### Constraint
 DO NOT include anything other than the JSON array in your response.
 ### Example
 Here is the chat example between human and assistant, inside <example></example> XML tags.
@@ -59,7 +59,7 @@ User:{{"input_text": ["I want to know which company’s email address test@examp
 Assistant:{{"metadata_map": [{{"metadata_field_name": "email", "metadata_field_value": "test@example.com", "comparison_operator": "="}}]}}
 User:{{"input_text": "What are the movies with a score of more than 9 in 2024?", "metadata_fields": ["name", "year", "rating", "country"]}}
 Assistant:{{"metadata_map": [{{"metadata_field_name": "year", "metadata_field_value": "2024", "comparison_operator": "="}, {{"metadata_field_name": "rating", "metadata_field_value": "9", "comparison_operator": ">"}}]}}
-</example> 
+</example>
 ### User Input
 {{"input_text" : "{input_text}", "metadata_fields" : {metadata_fields}}}
 ### Assistant Output

api/core/workflow/nodes/llm/exc.py

@@ -38,3 +38,8 @@ class MemoryRolePrefixRequiredError(LLMNodeError):
 class FileTypeNotSupportError(LLMNodeError):
     def __init__(self, *, type_name: str):
         super().__init__(f"{type_name} type is not supported by this model")
+
+
+class UnsupportedPromptContentTypeError(LLMNodeError):
+    def __init__(self, *, type_name: str) -> None:
+        super().__init__(f"Prompt content type {type_name} is not supported.")

api/core/workflow/nodes/llm/file_saver.py

@@ -0,0 +1,160 @@
+import mimetypes
+import typing as tp
+
+from sqlalchemy import Engine
+
+from constants.mimetypes import DEFAULT_EXTENSION, DEFAULT_MIME_TYPE
+from core.file import File, FileTransferMethod, FileType
+from core.helper import ssrf_proxy
+from core.tools.signature import sign_tool_file
+from core.tools.tool_file_manager import ToolFileManager
+from models import db as global_db
+
+
+class LLMFileSaver(tp.Protocol):
+    """LLMFileSaver is responsible for save multimodal output returned by
+    LLM.
+    """
+
+    def save_binary_string(
+        self,
+        data: bytes,
+        mime_type: str,
+        file_type: FileType,
+        extension_override: str | None = None,
+    ) -> File:
+        """save_binary_string saves the inline file data returned by LLM.
+
+        Currently (2025-04-30), only some of Google Gemini models will return
+        multimodal output as inline data.
+
+        :param data: the contents of the file
+        :param mime_type: the media type of the file, specified by rfc6838
+            (https://datatracker.ietf.org/doc/html/rfc6838)
+        :param file_type: The file type of the inline file.
+        :param extension_override: Override the auto-detected file extension while saving this file.
+
+            The default value is `None`, which means do not override the file extension and guessing it
+            from the `mime_type` attribute while saving the file.
+
+            Setting it to values other than `None` means override the file's extension, and
+            will bypass the extension guessing saving the file.
+
+            Specially, setting it to empty string (`""`) will leave the file extension empty.
+
+            When it is not `None` or empty string (`""`), it should be a string beginning with a
+            dot (`.`). For example, `.py` and `.tar.gz` are both valid values, while `py`
+            and `tar.gz` are not.
+        """
+        pass
+
+    def save_remote_url(self, url: str, file_type: FileType) -> File:
+        """save_remote_url saves the file from a remote url returned by LLM.
+
+        Currently (2025-04-30), no model returns multimodel output as a url.
+
+        :param url: the url of the file.
+        :param file_type: the file type of the file, check `FileType` enum for reference.
+        """
+        pass
+
+
+EngineFactory: tp.TypeAlias = tp.Callable[[], Engine]
+
+
+class FileSaverImpl(LLMFileSaver):
+    _engine_factory: EngineFactory
+    _tenant_id: str
+    _user_id: str
+
+    def __init__(self, user_id: str, tenant_id: str, engine_factory: EngineFactory | None = None):
+        if engine_factory is None:
+
+            def _factory():
+                return global_db.engine
+
+            engine_factory = _factory
+        self._engine_factory = engine_factory
+        self._user_id = user_id
+        self._tenant_id = tenant_id
+
+    def _get_tool_file_manager(self):
+        return ToolFileManager(engine=self._engine_factory())
+
+    def save_remote_url(self, url: str, file_type: FileType) -> File:
+        http_response = ssrf_proxy.get(url)
+        http_response.raise_for_status()
+        data = http_response.content
+        mime_type_from_header = http_response.headers.get("Content-Type")
+        mime_type, extension = _extract_content_type_and_extension(url, mime_type_from_header)
+        return self.save_binary_string(data, mime_type, file_type, extension_override=extension)
+
+    def save_binary_string(
+        self,
+        data: bytes,
+        mime_type: str,
+        file_type: FileType,
+        extension_override: str | None = None,
+    ) -> File:
+        tool_file_manager = self._get_tool_file_manager()
+        tool_file = tool_file_manager.create_file_by_raw(
+            user_id=self._user_id,
+            tenant_id=self._tenant_id,
+            # TODO(QuantumGhost): what is conversation id?
+            conversation_id=None,
+            file_binary=data,
+            mimetype=mime_type,
+        )
+        extension_override = _validate_extension_override(extension_override)
+        extension = _get_extension(mime_type, extension_override)
+        url = sign_tool_file(tool_file.id, extension)
+
+        return File(
+            tenant_id=self._tenant_id,
+            type=file_type,
+            transfer_method=FileTransferMethod.TOOL_FILE,
+            filename=tool_file.name,
+            extension=extension,
+            mime_type=mime_type,
+            size=len(data),
+            related_id=tool_file.id,
+            url=url,
+            # TODO(QuantumGhost): how should I set the following key?
+            # What's the difference between `remote_url` and `url`?
+            # What's the purpose of `storage_key` and `dify_model_identity`?
+            storage_key=tool_file.file_key,
+        )
+
+
+def _get_extension(mime_type: str, extension_override: str | None = None) -> str:
+    """get_extension return the extension of file.
+
+    If the `extension_override` parameter is set, this function should honor it and
+    return its value.
+    """
+    if extension_override is not None:
+        return extension_override
+    return mimetypes.guess_extension(mime_type) or DEFAULT_EXTENSION
+
+
+def _extract_content_type_and_extension(url: str, content_type_header: str | None) -> tuple[str, str]:
+    """_extract_content_type_and_extension tries to
+    guess content type of file from url and `Content-Type` header in response.
+    """
+    if content_type_header:
+        extension = mimetypes.guess_extension(content_type_header) or DEFAULT_EXTENSION
+        return content_type_header, extension
+    content_type = mimetypes.guess_type(url)[0] or DEFAULT_MIME_TYPE
+    extension = mimetypes.guess_extension(content_type) or DEFAULT_EXTENSION
+    return content_type, extension
+
+
+def _validate_extension_override(extension_override: str | None) -> str | None:
+    # `extension_override` is allow to be `None or `""`.
+    if extension_override is None:
+        return None
+    if extension_override == "":
+        return ""
+    if not extension_override.startswith("."):
+        raise ValueError("extension_override should start with '.' if not None or empty.", extension_override)
+    return extension_override

api/core/workflow/nodes/llm/node.py

@@ -1,3 +1,5 @@
+import base64
+import io
 import json
 import logging
 from collections.abc import Generator, Mapping, Sequence
@@ -21,7 +23,7 @@ from core.model_runtime.entities import (
     PromptMessageContentType,
     TextPromptMessageContent,
 )
-from core.model_runtime.entities.llm_entities import LLMResult, LLMUsage
+from core.model_runtime.entities.llm_entities import LLMResult, LLMResultChunk, LLMUsage
 from core.model_runtime.entities.message_entities import (
     AssistantPromptMessage,
     PromptMessageContentUnionTypes,
@@ -38,7 +40,6 @@ from core.model_runtime.entities.model_entities import (
 )
 from core.model_runtime.model_providers.__base.large_language_model import LargeLanguageModel
 from core.model_runtime.utils.encoders import jsonable_encoder
-from core.model_runtime.utils.helper import convert_llm_result_chunk_to_str
 from core.plugin.entities.plugin import ModelProviderID
 from core.prompt.entities.advanced_prompt_entities import CompletionModelPromptTemplate, MemoryConfig
 from core.prompt.utils.prompt_message_util import PromptMessageUtil
@@ -95,9 +96,13 @@ from .exc import (
     TemplateTypeNotSupportError,
     VariableNotFoundError,
 )
+from .file_saver import FileSaverImpl, LLMFileSaver
 
 if TYPE_CHECKING:
     from core.file.models import File
+    from core.workflow.graph_engine.entities.graph import Graph
+    from core.workflow.graph_engine.entities.graph_init_params import GraphInitParams
+    from core.workflow.graph_engine.entities.graph_runtime_state import GraphRuntimeState
 
 logger = logging.getLogger(__name__)
 
@@ -106,6 +111,43 @@ class LLMNode(BaseNode[LLMNodeData]):
     _node_data_cls = LLMNodeData
     _node_type = NodeType.LLM
 
+    # Instance attributes specific to LLMNode.
+    # Output variable for file
+    _file_outputs: list["File"]
+
+    _llm_file_saver: LLMFileSaver
+
+    def __init__(
+        self,
+        id: str,
+        config: Mapping[str, Any],
+        graph_init_params: "GraphInitParams",
+        graph: "Graph",
+        graph_runtime_state: "GraphRuntimeState",
+        previous_node_id: Optional[str] = None,
+        thread_pool_id: Optional[str] = None,
+        *,
+        llm_file_saver: LLMFileSaver | None = None,
+    ) -> None:
+        super().__init__(
+            id=id,
+            config=config,
+            graph_init_params=graph_init_params,
+            graph=graph,
+            graph_runtime_state=graph_runtime_state,
+            previous_node_id=previous_node_id,
+            thread_pool_id=thread_pool_id,
+        )
+        # LLM file outputs, used for MultiModal outputs.
+        self._file_outputs: list[File] = []
+
+        if llm_file_saver is None:
+            llm_file_saver = FileSaverImpl(
+                user_id=graph_init_params.user_id,
+                tenant_id=graph_init_params.tenant_id,
+            )
+        self._llm_file_saver = llm_file_saver
+
     def _run(self) -> Generator[NodeEvent | InNodeEvent, None, None]:
         def process_structured_output(text: str) -> Optional[dict[str, Any] | list[Any]]:
             """Process structured output if enabled"""
@@ -215,6 +257,9 @@ class LLMNode(BaseNode[LLMNodeData]):
             structured_output = process_structured_output(result_text)
             if structured_output:
                 outputs["structured_output"] = structured_output
+            if self._file_outputs is not None:
+                outputs["files"] = self._file_outputs
+
             yield RunCompletedEvent(
                 run_result=NodeRunResult(
                     status=WorkflowNodeExecutionStatus.SUCCEEDED,
@@ -240,6 +285,7 @@ class LLMNode(BaseNode[LLMNodeData]):
                 )
             )
         except Exception as e:
+            logger.exception("error while executing llm node")
             yield RunCompletedEvent(
                 run_result=NodeRunResult(
                     status=WorkflowNodeExecutionStatus.FAILED,
@@ -268,44 +314,45 @@ class LLMNode(BaseNode[LLMNodeData]):
 
         return self._handle_invoke_result(invoke_result=invoke_result)
 
-    def _handle_invoke_result(self, invoke_result: LLMResult | Generator) -> Generator[NodeEvent, None, None]:
+    def _handle_invoke_result(
+        self, invoke_result: LLMResult | Generator[LLMResultChunk, None, None]
+    ) -> Generator[NodeEvent, None, None]:
+        # For blocking mode
         if isinstance(invoke_result, LLMResult):
-            message_text = convert_llm_result_chunk_to_str(invoke_result.message.content)
-
-            yield ModelInvokeCompletedEvent(
-                text=message_text,
-                usage=invoke_result.usage,
-                finish_reason=None,
-            )
+            event = self._handle_blocking_result(invoke_result=invoke_result)
+            yield event
             return
 
-        model = None
+        # For streaming mode
+        model = ""
         prompt_messages: list[PromptMessage] = []
-        full_text = ""
-        usage = None
+
+        usage = LLMUsage.empty_usage()
         finish_reason = None
+        full_text_buffer = io.StringIO()
         for result in invoke_result:
-            text = convert_llm_result_chunk_to_str(result.delta.message.content)
-            full_text += text
+            contents = result.delta.message.content
+            for text_part in self._save_multimodal_output_and_convert_result_to_markdown(contents):
+                full_text_buffer.write(text_part)
+                yield RunStreamChunkEvent(chunk_content=text_part, from_variable_selector=[self.node_id, "text"])
 
-            yield RunStreamChunkEvent(chunk_content=text, from_variable_selector=[self.node_id, "text"])
-
-            if not model:
+            # Update the whole metadata
+            if not model and result.model:
                 model = result.model
-
-            if not prompt_messages:
-                prompt_messages = result.prompt_messages
-
-            if not usage and result.delta.usage:
+            if len(prompt_messages) == 0:
+                # TODO(QuantumGhost): it seems that this update has no visable effect.
+                # What's the purpose of the line below?
+                prompt_messages = list(result.prompt_messages)
+            if usage.prompt_tokens == 0 and result.delta.usage:
                 usage = result.delta.usage
-
-            if not finish_reason and result.delta.finish_reason:
+            if finish_reason is None and result.delta.finish_reason:
                 finish_reason = result.delta.finish_reason
 
-        if not usage:
-            usage = LLMUsage.empty_usage()
+        yield ModelInvokeCompletedEvent(text=full_text_buffer.getvalue(), usage=usage, finish_reason=finish_reason)
 
-        yield ModelInvokeCompletedEvent(text=full_text, usage=usage, finish_reason=finish_reason)
+    def _image_file_to_markdown(self, file: "File", /):
+        text_chunk = f"![]({file.generate_url()})"
+        return text_chunk
 
     def _transform_chat_messages(
         self, messages: Sequence[LLMNodeChatModelMessage] | LLMNodeCompletionModelPromptTemplate, /
@@ -963,6 +1010,42 @@ class LLMNode(BaseNode[LLMNodeData]):
 
         return prompt_messages
 
+    def _handle_blocking_result(self, *, invoke_result: LLMResult) -> ModelInvokeCompletedEvent:
+        buffer = io.StringIO()
+        for text_part in self._save_multimodal_output_and_convert_result_to_markdown(invoke_result.message.content):
+            buffer.write(text_part)
+
+        return ModelInvokeCompletedEvent(
+            text=buffer.getvalue(),
+            usage=invoke_result.usage,
+            finish_reason=None,
+        )
+
+    def _save_multimodal_image_output(self, content: ImagePromptMessageContent) -> "File":
+        """_save_multimodal_output saves multi-modal contents generated by LLM plugins.
+
+        There are two kinds of multimodal outputs:
+
+          - Inlined data encoded in base64, which would be saved to storage directly.
+          - Remote files referenced by an url, which would be downloaded and then saved to storage.
+
+        Currently, only image files are supported.
+        """
+        # Inject the saver somehow...
+        _saver = self._llm_file_saver
+
+        # If this
+        if content.url != "":
+            saved_file = _saver.save_remote_url(content.url, FileType.IMAGE)
+        else:
+            saved_file = _saver.save_binary_string(
+                data=base64.b64decode(content.base64_data),
+                mime_type=content.mime_type,
+                file_type=FileType.IMAGE,
+            )
+        self._file_outputs.append(saved_file)
+        return saved_file
+
     def _handle_native_json_schema(self, model_parameters: dict, rules: list[ParameterRule]) -> dict:
         """
         Handle structured output for models with native JSON schema support.
@@ -1123,6 +1206,41 @@ class LLMNode(BaseNode[LLMNodeData]):
             else SupportStructuredOutputStatus.UNSUPPORTED
         )
 
+    def _save_multimodal_output_and_convert_result_to_markdown(
+        self,
+        contents: str | list[PromptMessageContentUnionTypes] | None,
+    ) -> Generator[str, None, None]:
+        """Convert intermediate prompt messages into strings and yield them to the caller.
+
+        If the messages contain non-textual content (e.g., multimedia like images or videos),
+        it will be saved separately, and the corresponding Markdown representation will
+        be yielded to the caller.
+        """
+
+        # NOTE(QuantumGhost): This function should yield results to the caller immediately
+        # whenever new content or partial content is available. Avoid any intermediate buffering
+        # of results. Additionally, do not yield empty strings; instead, yield from an empty list
+        # if necessary.
+        if contents is None:
+            yield from []
+            return
+        if isinstance(contents, str):
+            yield contents
+        elif isinstance(contents, list):
+            for item in contents:
+                if isinstance(item, TextPromptMessageContent):
+                    yield item.data
+                elif isinstance(item, ImagePromptMessageContent):
+                    file = self._save_multimodal_image_output(item)
+                    self._file_outputs.append(file)
+                    yield self._image_file_to_markdown(file)
+                else:
+                    logger.warning("unknown item type encountered, type=%s", type(item))
+                    yield str(item)
+        else:
+            logger.warning("unknown contents type encountered, type=%s", type(contents))
+            yield str(contents)
+
 
 def _combine_message_content_with_role(
     *, contents: Optional[str | list[PromptMessageContentUnionTypes]] = None, role: PromptMessageRole

api/core/workflow/nodes/parameter_extractor/prompts.py

@@ -17,7 +17,7 @@ Some additional information is provided below. Always adhere to these instructio
 </instruction>
 Steps:
 1. Review the chat history provided within the <histories> tags.
-2. Extract the relevant information based on the criteria given, output multiple values if there is multiple relevant information that match the criteria in the given text. 
+2. Extract the relevant information based on the criteria given, output multiple values if there is multiple relevant information that match the criteria in the given text.
 3. Generate a well-formatted output using the defined functions and arguments.
 4. Use the `extract_parameter` function to create structured outputs with appropriate parameters.
 5. Do not include any XML tags in your output.
@@ -89,13 +89,13 @@ Some extra information are provided below, I should always follow the instructio
 </instructions>
 
 ### Extract parameter Workflow
-I need to extract the following information from the input text. The <information to be extracted> tag specifies the 'type', 'description' and 'required' of the information to be extracted. 
+I need to extract the following information from the input text. The <information to be extracted> tag specifies the 'type', 'description' and 'required' of the information to be extracted.
 <information to be extracted>
 {{ structure }}
 </information to be extracted>
 
 Step 1: Carefully read the input and understand the structure of the expected output.
-Step 2: Extract relevant parameters from the provided text based on the name and description of object. 
+Step 2: Extract relevant parameters from the provided text based on the name and description of object.
 Step 3: Structure the extracted parameters to JSON object as specified in <structure>.
 Step 4: Ensure that the JSON object is properly formatted and valid. The output should not contain any XML tags. Only the JSON object should be outputted.
 
@@ -106,10 +106,10 @@ Here are the chat histories between human and assistant, inside <histories></his
 </histories>
 
 ### Structure
-Here is the structure of the expected output, I should always follow the output structure. 
+Here is the structure of the expected output, I should always follow the output structure.
 {{γγγ
-  'properties1': 'relevant text extracted from input', 
-  'properties2': 'relevant text extracted from input', 
+  'properties1': 'relevant text extracted from input',
+  'properties2': 'relevant text extracted from input',
 }}γγγ
 
 ### Input Text
@@ -119,7 +119,7 @@ Inside <text></text> XML tags, there is a text that I should extract parameters
 </text>
 
 ### Answer
-I should always output a valid JSON object. Output nothing other than the JSON object. 
+I should always output a valid JSON object. Output nothing other than the JSON object.
 ```JSON
 """  # noqa: E501
 

api/core/workflow/nodes/question_classifier/template_prompts.py

@@ -55,7 +55,7 @@ You are a text classification engine that analyzes text data and assigns categor
 Your task is to assign one categories ONLY to the input text and only one category may be assigned returned in the output. Additionally, you need to extract the key words from the text that are related to the classification.
 ### Format
 The input text is in the variable input_text. Categories are specified as a category list  with two filed category_id and category_name in the variable categories. Classification instructions may be included to improve the classification accuracy.
-### Constraint 
+### Constraint
 DO NOT include anything other than the JSON array in your response.
 ### Example
 Here is the chat example between human and assistant, inside <example></example> XML tags.
@@ -64,7 +64,7 @@ User:{{"input_text": ["I recently had a great experience with your company. The
 Assistant:{{"keywords": ["recently", "great experience", "company", "service", "prompt", "staff", "friendly"],"category_id": "f5660049-284f-41a7-b301-fd24176a711c","category_name": "Customer Service"}}
 User:{{"input_text": ["bad service, slow to bring the food"], "categories": [{{"category_id":"80fb86a0-4454-4bf5-924c-f253fdd83c02","category_name":"Food Quality"}},{{"category_id":"f6ff5bc3-aca0-4e4a-8627-e760d0aca78f","category_name":"Experience"}},{{"category_id":"cc771f63-74e7-4c61-882e-3eda9d8ba5d7","category_name":"Price"}}], "classification_instructions": []}}
 Assistant:{{"keywords": ["bad service", "slow", "food", "tip", "terrible", "waitresses"],"category_id": "f6ff5bc3-aca0-4e4a-8627-e760d0aca78f","category_name": "Experience"}}
-</example> 
+</example>
 ### Memory
 Here are the chat histories between human and assistant, inside <histories></histories> XML tags.
 <histories>

api/core/workflow/nodes/variable_assigner/v2/enums.py

@@ -11,6 +11,8 @@ class Operation(StrEnum):
     SUBTRACT = "-="
     MULTIPLY = "*="
     DIVIDE = "/="
+    REMOVE_FIRST = "remove-first"
+    REMOVE_LAST = "remove-last"
 
 
 class InputType(StrEnum):

api/core/workflow/nodes/variable_assigner/v2/helpers.py

@@ -23,6 +23,15 @@ def is_operation_supported(*, variable_type: SegmentType, operation: Operation):
                 SegmentType.ARRAY_NUMBER,
                 SegmentType.ARRAY_FILE,
             }
+        case Operation.REMOVE_FIRST | Operation.REMOVE_LAST:
+            # Only array variable can have elements removed
+            return variable_type in {
+                SegmentType.ARRAY_ANY,
+                SegmentType.ARRAY_OBJECT,
+                SegmentType.ARRAY_STRING,
+                SegmentType.ARRAY_NUMBER,
+                SegmentType.ARRAY_FILE,
+            }
         case _:
             return False
 
@@ -51,7 +60,7 @@ def is_constant_input_supported(*, variable_type: SegmentType, operation: Operat
 
 
 def is_input_value_valid(*, variable_type: SegmentType, operation: Operation, value: Any):
-    if operation == Operation.CLEAR:
+    if operation in {Operation.CLEAR, Operation.REMOVE_FIRST, Operation.REMOVE_LAST}:
         return True
     match variable_type:
         case SegmentType.STRING:

api/core/workflow/nodes/variable_assigner/v2/node.py

@@ -64,7 +64,7 @@ class VariableAssignerNode(BaseNode[VariableAssignerNodeData]):
                 # Get value from variable pool
                 if (
                     item.input_type == InputType.VARIABLE
-                    and item.operation != Operation.CLEAR
+                    and item.operation not in {Operation.CLEAR, Operation.REMOVE_FIRST, Operation.REMOVE_LAST}
                     and item.value is not None
                 ):
                     value = self.graph_runtime_state.variable_pool.get(item.value)
@@ -165,5 +165,15 @@ class VariableAssignerNode(BaseNode[VariableAssignerNodeData]):
                 return variable.value * value
             case Operation.DIVIDE:
                 return variable.value / value
+            case Operation.REMOVE_FIRST:
+                # If array is empty, do nothing
+                if not variable.value:
+                    return variable.value
+                return variable.value[1:]
+            case Operation.REMOVE_LAST:
+                # If array is empty, do nothing
+                if not variable.value:
+                    return variable.value
+                return variable.value[:-1]
             case _:
                 raise OperationNotSupportedError(operation=operation, variable_type=variable.value_type)