Skip to content

armgit / dify

Public Imported from GitHub
0

Build Agentic workflows, RAG pipelines, with rich AI model and tool support on one collaborative workspace. Deploy on cloud, VPC, or self-hosted, so teams move from prototype to production without rebuilding the stack.

19.1 KB
Raw Blame
"""
Service for generating snippet workflow executions.

Uses an adapter pattern to bridge CustomizedSnippet with the App-based
WorkflowAppGenerator. The adapter (_SnippetAsApp) provides the minimal App-like
interface needed by the generator, avoiding modifications to core workflow
infrastructure.

Key invariants:
- Snippets always run as WORKFLOW mode (not CHAT or ADVANCED_CHAT).
- The adapter maps snippet.id to app_id in workflow execution records.
- Snippet debugging has no rate limiting (max_active_requests = 0).

Supported execution modes:
- Full workflow run (generate): Runs the entire draft workflow as SSE stream.
- Single node run (run_draft_node): Synchronous single-step debugging for regular nodes.
- Single iteration run (generate_single_iteration): SSE stream for iteration container nodes.
- Single loop run (generate_single_loop): SSE stream for loop container nodes.
"""

import json
import logging
from collections.abc import Generator, Mapping, Sequence
from typing import Any, cast

from sqlalchemy.orm import Session, make_transient, sessionmaker

from core.app.app_config.features.file_upload.manager import FileUploadConfigManager
from core.app.apps.workflow.app_generator import WorkflowAppGenerator
from core.app.entities.app_invoke_entities import InvokeFrom
from core.app.file_access import DatabaseFileAccessController
from core.workflow.snippet_start import SNIPPET_VIRTUAL_START_NODE_ID
from factories import file_factory
from graphon.file.models import File
from graphon.graph.scoping import resolve_container_id
from models import Account
from models.model import App, AppMode, EndUser
from models.snippet import CustomizedSnippet
from models.workflow import Workflow, WorkflowNodeExecutionModel
from services.snippet_service import SnippetService
from services.workflow_execution_identity import build_workflow_execution_identity
from services.workflow_run_agg import WorkflowRunAgg
from services.workflow_service import WorkflowService

logger = logging.getLogger(__name__)
_file_access_controller = DatabaseFileAccessController()


class _SnippetAsApp:
    """
    Minimal adapter that wraps a CustomizedSnippet to satisfy the App-like
    interface required by WorkflowAppGenerator, WorkflowAppConfigManager,
    and WorkflowService.run_draft_workflow_node.

    Used properties:
    - id: maps to snippet.id (stored as app_id in workflows table)
    - tenant_id: maps to snippet.tenant_id
    - mode: hardcoded to AppMode.WORKFLOW since snippets always run as workflows
    - max_active_requests: defaults to 0 (no limit) for snippet debugging
    - app_model_config_id: None (snippets don't have app model configs)
    """

    id: str
    tenant_id: str
    mode: str
    max_active_requests: int
    app_model_config_id: str | None

    def __init__(self, snippet: CustomizedSnippet) -> None:
        self.id = snippet.id
        self.tenant_id = snippet.tenant_id
        self.mode = AppMode.WORKFLOW.value
        self.max_active_requests = 0
        self.app_model_config_id = None


class SnippetGenerateService:
    """
    Service for running snippet workflow executions.

    Adapts CustomizedSnippet to work with the existing App-based
    WorkflowAppGenerator infrastructure, avoiding duplication of the
    complex workflow execution pipeline.
    """

    # Specific ID for the injected virtual Start node so it can be recognised
    _VIRTUAL_START_NODE_ID = SNIPPET_VIRTUAL_START_NODE_ID

    @classmethod
    def _is_virtual_start_event(cls, message: Mapping[str, Any] | str) -> bool:
        """
        Return True when *message* is a snippet-only virtual Start node event.

        The virtual Start node is injected purely for snippet execution and is
        not part of the persisted draft graph. Filter its node lifecycle events
        out of the SSE stream so the frontend only receives nodes that exist on
        the canvas.
        """
        if not isinstance(message, Mapping):
            return False

        if message.get("event") not in {"node_started", "node_finished"}:
            return False

        data = message.get("data")
        if not isinstance(data, Mapping):
            return False

        return data.get("node_id") == cls._VIRTUAL_START_NODE_ID

    @classmethod
    def _filter_virtual_start_events(
        cls,
        response: Mapping[str, Any] | Generator[Mapping[str, Any] | str],
    ) -> Mapping[str, Any] | Generator[Mapping[str, Any] | str]:
        """
        Drop snippet virtual Start node lifecycle events from stream responses.

        Blocking responses are returned unchanged because they never expose the
        injected node as a standalone event payload.
        """
        if isinstance(response, Mapping):
            return response

        def _stream() -> Generator[Mapping[str, Any] | str]:
            for message in response:
                if cls._is_virtual_start_event(message):
                    continue
                yield message

        return _stream()

    @classmethod
    def generate(
        cls,
        snippet: CustomizedSnippet,
        user: Account | EndUser,
        args: Mapping[str, Any],
        invoke_from: InvokeFrom,
        streaming: bool = True,
        session_maker: sessionmaker[Session] | None = None,
    ) -> Mapping[str, Any] | Generator[str]:
        """
        Run a snippet's draft workflow.

        Retrieves the draft workflow, adapts the snippet to an App-like proxy,
        then delegates execution to WorkflowAppGenerator.

        If the workflow graph has no Start node, a virtual Start node is injected
        in-memory so that:
        1. Graph validation passes (root node must have execution_type=ROOT).
        2. User inputs are processed into the variable pool by the StartNode logic.

        :param snippet: CustomizedSnippet instance
        :param user: Account or EndUser initiating the run
        :param args: Workflow inputs (must include "inputs" key)
        :param invoke_from: Source of invocation (typically DEBUGGER)
        :param streaming: Whether to stream the response
        :return: Blocking response mapping or SSE streaming generator
        :raises ValueError: If the snippet has no draft workflow
        """
        snippet_service = SnippetService(session_maker)
        workflow = snippet_service.get_draft_workflow(snippet=snippet)
        if not workflow:
            raise ValueError("Workflow not initialized")

        # Inject a virtual Start node when the graph doesn't have one.
        workflow = cls._ensure_start_node(workflow, snippet)

        # Adapt snippet to App-like interface for WorkflowAppGenerator
        app_proxy = cast(App, _SnippetAsApp(snippet))

        response = WorkflowAppGenerator(execution_driver=WorkflowRunAgg.run).generate(
            app_model=app_proxy,
            workflow=workflow,
            user=user,
            execution_identity=build_workflow_execution_identity(user),
            args=args,
            invoke_from=invoke_from,
            streaming=streaming,
            call_depth=0,
        )

        return WorkflowAppGenerator.convert_to_event_stream(cls._filter_virtual_start_events(response))

    @classmethod
    def run_published(
        cls,
        snippet: CustomizedSnippet,
        user: Account | EndUser,
        args: Mapping[str, Any],
        invoke_from: InvokeFrom,
        session_maker: sessionmaker[Session] | None = None,
    ) -> Mapping[str, Any]:
        """
        Run a snippet's published workflow in non-streaming (blocking) mode.

        Similar to :meth:`generate` but targets the published workflow instead
        of the draft, and returns the raw blocking response without SSE
        wrapping. Designed for programmatic callers that need direct workflow outputs.

        :param snippet: CustomizedSnippet instance (must be published)
        :param user: Account or EndUser initiating the run
        :param args: Workflow inputs (must include "inputs" key)
        :param invoke_from: Source of invocation
        :return: Blocking response mapping with workflow outputs
        :raises ValueError: If the snippet has no published workflow
        """
        snippet_service = SnippetService(session_maker)
        workflow = snippet_service.get_published_workflow(snippet)
        if not workflow:
            raise ValueError("No published workflow found for snippet")

        # Inject a virtual Start node when the graph doesn't have one.
        workflow = cls._ensure_start_node(workflow, snippet)

        app_proxy = cast(App, _SnippetAsApp(snippet))

        response = WorkflowAppGenerator(execution_driver=WorkflowRunAgg.run).generate(
            app_model=app_proxy,
            workflow=workflow,
            user=user,
            execution_identity=build_workflow_execution_identity(user),
            args=args,
            invoke_from=invoke_from,
            streaming=False,
            call_depth=0,
        )
        return response

    @classmethod
    def ensure_start_node_for_worker(cls, workflow: Workflow, snippet: CustomizedSnippet) -> Workflow:
        """Public wrapper for worker-thread start-node injection."""
        return cls._ensure_start_node(workflow, snippet)

    @classmethod
    def _ensure_start_node(cls, workflow: Workflow, snippet: CustomizedSnippet) -> Workflow:
        """
        Return *workflow* with a Start node.

        If the graph already contains a Start node, the original workflow is
        returned unchanged.  Otherwise a virtual Start node is injected and the
        workflow object is detached from the SQLAlchemy session so the in-memory
        change is never flushed to the database.
        """
        graph_dict = workflow.graph_dict
        nodes: list[dict[str, Any]] = graph_dict.get("nodes", [])

        has_start = any(node.get("data", {}).get("type") == "start" for node in nodes)
        if has_start:
            return workflow

        modified_graph = cls._inject_virtual_start_node(
            graph_dict=graph_dict,
            input_fields=snippet.input_fields_list,
        )

        # Detach from session to prevent accidental DB persistence of the
        # modified graph.  All attributes remain accessible for read.
        make_transient(workflow)
        workflow.graph = json.dumps(modified_graph)
        return workflow

    @classmethod
    def _inject_virtual_start_node(
        cls,
        graph_dict: Mapping[str, Any],
        input_fields: list[dict[str, Any]],
    ) -> dict[str, Any]:
        """
        Build a new graph dict with a virtual Start node prepended.

        The virtual Start node is wired to top-level nodes with no incoming
        edges. Container children retain their own entry points.

        :param graph_dict: Original graph configuration.
        :param input_fields: Snippet input field definitions from
            ``CustomizedSnippet.input_fields_list``.
        :return: New graph dict containing the virtual Start node and edges.
        """
        nodes: list[dict[str, Any]] = list(graph_dict.get("nodes", []))
        edges: list[dict[str, Any]] = list(graph_dict.get("edges", []))

        nodes_by_id = {node["id"]: node for node in nodes}
        # A container's entry node has no incoming edges, but is not a graph root.
        nodes_with_incoming: set[str] = set()
        for edge in edges:
            target = edge.get("target")
            if isinstance(target, str):
                nodes_with_incoming.add(target)
        root_candidate_ids = [
            node["id"]
            for node in nodes
            if node["id"] not in nodes_with_incoming and not resolve_container_id(node, nodes_by_id=nodes_by_id)
        ]

        # Build Start node ``variables`` from snippet input fields.
        start_variables: list[dict[str, Any]] = []
        for field in input_fields:
            var: dict[str, Any] = {
                "variable": field.get("variable", ""),
                "label": field.get("label", field.get("variable", "")),
                "type": field.get("type", "text-input"),
                "required": field.get("required", False),
                "options": field.get("options", []),
            }
            if field.get("max_length") is not None:
                var["max_length"] = field["max_length"]
            start_variables.append(var)

        virtual_start_node: dict[str, Any] = {
            "id": cls._VIRTUAL_START_NODE_ID,
            "data": {
                "type": "start",
                "title": "Start",
                "variables": start_variables,
            },
        }

        # Create edges from virtual Start to each root candidate.
        new_edges: list[dict[str, Any]] = [
            {
                "source": cls._VIRTUAL_START_NODE_ID,
                "sourceHandle": "source",
                "target": root_id,
                "targetHandle": "target",
            }
            for root_id in root_candidate_ids
        ]

        return {
            **graph_dict,
            "nodes": [virtual_start_node, *nodes],
            "edges": [*edges, *new_edges],
        }

    @classmethod
    def run_draft_node(
        cls,
        snippet: CustomizedSnippet,
        node_id: str,
        user_inputs: Mapping[str, Any],
        account: Account,
        query: str = "",
        files: Sequence[File] | None = None,
        session_maker: sessionmaker[Session] | None = None,
    ) -> WorkflowNodeExecutionModel:
        """
        Run a single node in a snippet's draft workflow (single-step debugging).

        Retrieves the draft workflow, adapts the snippet to an App-like proxy,
        parses file inputs, then delegates to WorkflowService.run_draft_workflow_node.

        :param snippet: CustomizedSnippet instance
        :param node_id: ID of the node to run
        :param user_inputs: User input values for the node
        :param account: Account initiating the run
        :param query: Optional query string
        :param files: Optional parsed file objects
        :return: WorkflowNodeExecutionModel with execution results
        :raises ValueError: If the snippet has no draft workflow
        """
        snippet_service = SnippetService(session_maker)
        draft_workflow = snippet_service.get_draft_workflow(snippet=snippet)
        if not draft_workflow:
            raise ValueError("Workflow not initialized")

        app_proxy = cast(App, _SnippetAsApp(snippet))

        workflow_service = WorkflowService()
        return workflow_service.run_draft_workflow_node(
            app_model=app_proxy,
            draft_workflow=draft_workflow,
            node_id=node_id,
            user_inputs=user_inputs,
            account=account,
            query=query,
            files=files,
        )

    @classmethod
    def generate_single_iteration(
        cls,
        snippet: CustomizedSnippet,
        user: Account | EndUser,
        node_id: str,
        args: Mapping[str, Any],
        streaming: bool = True,
        *,
        session_maker: sessionmaker[Session],
    ) -> Mapping[str, Any] | Generator[str]:
        """
        Run a single iteration node in a snippet's draft workflow.

        Iteration nodes are container nodes that execute their sub-graph multiple
        times, producing many events. Therefore, this uses the full WorkflowAppGenerator
        pipeline with SSE streaming (unlike regular single-step node run).

        :param snippet: CustomizedSnippet instance
        :param user: Account or EndUser initiating the run
        :param node_id: ID of the iteration node to run
        :param args: Dict containing 'inputs' key with iteration input data
        :param streaming: Whether to stream the response (should be True)
        :param session_maker: Factory for the synchronous database work before worker startup
        :return: SSE streaming generator
        :raises ValueError: If the snippet has no draft workflow
        """
        snippet_service = SnippetService(session_maker)
        workflow = snippet_service.get_draft_workflow(snippet=snippet)
        if not workflow:
            raise ValueError("Workflow not initialized")

        app_proxy = cast(App, _SnippetAsApp(snippet))

        with session_maker() as session:
            return WorkflowAppGenerator.convert_to_event_stream(
                WorkflowAppGenerator(execution_driver=WorkflowRunAgg.run).single_iteration_generate(
                    app_model=app_proxy,
                    workflow=workflow,
                    node_id=node_id,
                    user=user,
                    execution_identity=build_workflow_execution_identity(user),
                    args=args,
                    streaming=streaming,
                    session=session,
                )
            )

    @classmethod
    def generate_single_loop(
        cls,
        snippet: CustomizedSnippet,
        user: Account | EndUser,
        node_id: str,
        args: Any,
        streaming: bool = True,
        *,
        session_maker: sessionmaker[Session],
    ) -> Mapping[str, Any] | Generator[str]:
        """
        Run a single loop node in a snippet's draft workflow.

        Loop nodes are container nodes that execute their sub-graph repeatedly,
        producing many events. Therefore, this uses the full WorkflowAppGenerator
        pipeline with SSE streaming (unlike regular single-step node run).

        :param snippet: CustomizedSnippet instance
        :param user: Account or EndUser initiating the run
        :param node_id: ID of the loop node to run
        :param args: Pydantic model with 'inputs' attribute containing loop input data
        :param streaming: Whether to stream the response (should be True)
        :param session_maker: Factory for the synchronous database work before worker startup
        :return: SSE streaming generator
        :raises ValueError: If the snippet has no draft workflow
        """
        snippet_service = SnippetService(session_maker)
        workflow = snippet_service.get_draft_workflow(snippet=snippet)
        if not workflow:
            raise ValueError("Workflow not initialized")

        app_proxy = cast(App, _SnippetAsApp(snippet))

        with session_maker() as session:
            return WorkflowAppGenerator.convert_to_event_stream(
                WorkflowAppGenerator(execution_driver=WorkflowRunAgg.run).single_loop_generate(
                    app_model=app_proxy,
                    workflow=workflow,
                    node_id=node_id,
                    user=user,
                    execution_identity=build_workflow_execution_identity(user),
                    args=args,  # type: ignore[arg-type]
                    streaming=streaming,
                    session=session,
                )
            )

    @staticmethod
    def parse_files(workflow: Workflow, files: list[dict] | None = None) -> Sequence[File]:
        """
        Parse file mappings into File objects based on workflow configuration.

        :param workflow: Workflow instance for file upload config
        :param files: Raw file mapping dicts
        :return: Parsed File objects
        """
        files = files or []
        file_extra_config = FileUploadConfigManager.convert(workflow.features_dict, is_vision=False)
        if file_extra_config is None:
            return []
        return file_factory.build_from_mappings(
            mappings=files,
            tenant_id=workflow.tenant_id,
            config=file_extra_config,
            access_controller=_file_access_controller,
        )