PowerOn/gateway
2
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
gateway/notes/doc_statemachine_backend.md
ValueOn AG 64d1c083e0 mvp 1.2 ready for single test
2025-04-26 02:13:22 +02:00

14 KiB
Raw Blame History

State Machine Documentation for Backend Chat Workflow

Overview

The Chat Workflow system implements a state machine that processes user inputs through a sequence of well-defined steps. The system orchestrates interactions between users, project managers, and specialized agents to produce final outputs.

Core Objects

Workflow Object

{
  "id": "uuid-string",
  "mandateId": int,
  "userId": int,
  "name": "Workflow name",
  "startedAt": "ISO-datetime",
  "messages": [], // References to messages
  "messageIds": [], // List of message IDs
  "logs": [], // Log entries
  "dataStats": {}, // Performance metrics
  "currentRound": int, // Increments with each interaction
  "status": "string", // running, completed, failed, stopped
  "lastActivity": "ISO-datetime"
}

Message Object

{
  "id": "msg_uuid-string",
  "workflowId": "workflow-uuid",
  "role": "string", // user, assistant
  "agentName": "string", // Empty for user, agent name for assistant
  "content": "string", // The message text
  "documents": [], // List of document objects
  "timestamp": "ISO-datetime",
  "sequenceNo": int, // Position in conversation
  "status": "string" // first, step, last
}

Log Entry Object

{
  "id": "log_uuid-string",
  "workflowId": "workflow-uuid",
  "message": "string",
  "progress": int, // Optional, 0-100
  "type": "string", // info, warning, error
  "timestamp": "ISO-datetime",
  "agentName": "string", // Name of the agent that generated the log
  "status": "string" // current workflow status (running, completed, failed, stopped)
}

Document Object

{
  "id": "doc_uuid-string",
  "fileId": int,
  "name": "string", // Filename without extension
  "ext": "string", // File extension
  "data": "base64-encoded-string", // File contents
  "contents": [] // Extracted content items in text format
}

Content Item Object

{
  "sequenceNr": int, // Sequence in the document
  "name": "string",
  "ext": "string",
  "contentType": "string", // mime type
  "data": "string|base64", // Original content
  "dataExtracted": "string", // Optional AI-processed content based on extraction requirement
  "metadata": {
    "isText": boolean,
    "base64Encoded": boolean,
    "aiProcessed": boolean,
    // Optional metadata specific to content type
  },
  "summary": "string" // AI-generated static summary of the content
}

State Machine Workflow

1. Workflow Initialization

  • Trigger: User message received via /api/workflows/start OR /api/workflows/start?id=string
  • Input: UserInputRequest with prompt and optional listFileId
  • Process:
    • If id existing and workflow exists for id==workflowId: Load workflow, increment currentRound, set status "running"
    • Else: Create new workflow with "currentRound"=1, status "running"
  • Logs: "Workflow initialized" or "Running workflow", progress 0%
  • API Responses:
    • Success: 200 OK with workflow ID
    • Error: 400 Bad Request if input invalid, 404 Not Found if workflow ID not found

2. Workflow Exception

  • Trigger:
    • User stopped workflow via API
    • An exception happened
  • Process:
    • If status=="stopped": Set workflow status to "stopped", add message with status "last", update lastActivity, stop execution immediately
    • If status=="failed": Set workflow status to "failed", add message with status "last", update lastActivity, stop execution immediately
    • Else: Continue normally
  • Logs: "Workflow failure reported", progress 100%
  • API Responses:
    • For stop request: 200 OK when workflow successfully stopped
    • For exceptions: 500 Internal Server Error with error details

3. User Message Processing

  • Process:
    • Transform user input into message object with documents, message status "first"
    • Extract contents from files using getDocumentContents()
    • Generate static summaries for each content item
  • State Changes:
    • Add user message to workflow.messages array
    • Add message ID to workflow.messageIds array
    • Update workflow.lastActivity
  • Logs: "Workflow processing started", progress 0%

4. Project Manager Analysis

  • Process:
    • Generate prompt for project manager AI
    • Project manager analyzes request and documents
    • Project manager generates work plan and response
  • Outputs:
    • objFinalDocuments: List of str "filename.ext" for expected final output documents
    • objWorkplan: List of agent tasks
    • objUserResponse: Text response to user
    • userLanguage: Detected language code (e.g. en)
  • State Changes:
    • Add assistant message with project manager response, status "step"
    • Set user language in mydom interface
  • Logs: "Analyzing request and planning work" (10%), "Planned outputs" (20%), "Work plan created" (25%)

5. Agent Execution

  • Process (For each task in workplan):
    • Prepare input documents for agent
    • Execute agent with standardized task object
    • Save produced documents
    • Create assistant message with agent response, status "step"
  • Agent Task Object:
{
  "taskId": "uuid-string",
  "workflowId": "workflow-uuid",
  "prompt": "string",
  "inputDocuments": [], // list of documents including original document data and all content items data with original (attribute "data") and based on prompt (attribute "dataExtracted")
  "outputSpecifications": [
    {
      "label": "filename.ext",
      "description": "string"
    }
  ],
  "context": {
    "workflowRound": int,
    "agentType": "string",
    "timestamp": "ISO-datetime",
    "language": "language-code"
  }
}
  • Agent Result Object:
{
  "feedback": "string", // Text describing what the agent did
  "documents": [
    {
      "label": "filename.ext",
      "content": "string|binary" // Document contents
    }
  ]
}
  • State Changes: Add assistant message for each agent with agentName set, status "step"
  • Logs: "Running task X/Y: agentName" with progress updates from 30% to 90%

6. Final Response Generation

  • Process:
    • Create final message reviewing promised and delivered documents
    • Add documents to workflow
  • State Changes: Add final assistant message from projectManager, status "last"
  • Logs: "Creating final response" (90%)

7. Workflow Completion

  • Process:
    • Finalize workflow and update status
  • State Changes:
    • Set workflow status to "completed"
    • Update workflow.lastActivity
  • Logs: "Workflow completed successfully" with progress 100%
  • API Responses:
    • A message with status "last" is included in the response
    • Status endpoint will return "completed"

8. Workflow Stopped

  • Trigger: /api/workflows/{workflowId}/stop endpoint called
  • Process:
    • Immediately interrupt workflow execution
    • Save current state and mark as stopped
  • State Changes:
    • Set workflow status to "stopped"
    • Update lastActivity timestamp
  • Logs: "Workflow stopped by user" with progress 100%
  • API Responses:
    • 200 OK with confirmation message

9. Workflow Failed

  • Trigger: Exception during workflow execution
  • Process:
    • Log error details
    • Set workflow status to "failed"
  • State Changes:
    • Set workflow status to "failed"
    • Update lastActivity timestamp
  • Logs: Detailed error message with progress 100%
  • API Responses:
    • Status endpoint will return "failed" with error context

10. Workflow Resumption

  • Trigger: /api/workflows/start?id={workflowId} endpoint called with existing workflow ID
  • Process:
    • Load existing workflow
    • Increment currentRound counter
    • Start processing from user message
  • State Changes:
    • Set status to "running"
    • Increment currentRound
    • Add new user message
  • Logs: "Resuming workflow, round {currentRound}" with progress 0%
  • API Responses:
    • Same as workflow initialization

11. Workflow Reset/Deletion

  • Trigger: /api/workflows/{workflowId} DELETE endpoint called
  • Process:
    • Remove all workflow data from storage
  • State Changes:
    • Workflow no longer exists in the system
  • Logs: Log in system log that workflow was deleted
  • API Responses:
    • 200 OK if successful
    • 404 Not Found if workflow didn't exist

API Endpoints and Polling Support

Main Workflow Endpoints

  • POST /api/workflows/start?id=string: Submit user input to start a new workflow, optional with workflow id to continue existing workflow
  • POST /api/workflows/{workflowId}/stop: Stop a running workflow: Immediately to set workflow status to "stopped"
  • DELETE /api/workflows/{workflowId}: Delete a workflow
  • GET /api/workflows/{workflowId}/status: Get workflow status (running, completed, failed, stopped)
  • GET /api/workflows/{workflowId}/logs?id=string: Get workflow logs, optional with log id to get only logs produced after and including log with log id
  • GET /api/workflows/{workflowId}/messages?id=string: Get workflow messages, optional with message id to get only messages produced after and including message with log id

Document Management

  • DELETE /api/workflows/{workflowId}/messages/{messageId}: Delete a message
  • DELETE /api/workflows/{workflowId}/messages/{messageId}/files/{fileId}: Remove file from message

Backend Support for Frontend Polling

The backend implements efficient support for frontend polling mechanisms:

  1. Selective Data Transfer:

    • Both /logs and /messages endpoints accept an optional id parameter
    • When provided, only records with IDs equal to or newer than the specified ID are returned
    • This minimizes data transfer and improves performance

  2. Log Storage:

    • Each log entry includes timestamp, progress indicators, and status
    • Frontend can accurately track workflow progress and update UI accordingly
    • Logs are stored in chronological order with monotonically increasing IDs

  3. Message Handling:

    • Messages include a status field ("first", "step", "last")
    • The "last" status indicates completion of the current workflow round
    • Frontend uses this to determine when to enable user input

  4. Status Endpoint:

    • Lightweight endpoint that returns only the current workflow status
    • Used by frontend to detect state changes without transferring all data
    • Also includes lastActivity timestamp to detect stalled workflows

  5. Caching Layer:

    • Backend implements caching for frequent polling requests
    • Reduces database load and improves response times
    • Cache invalidation occurs when workflow status changes

  6. Batch Processing:

    • Large log or message sets are paginated automatically
    • Frontend receives data in manageable chunks
    • Prevents memory issues with long-running workflows

Document Object Structure Clarification

The Document Object contains both raw data and processed contents:

  • data: Contains the base64-encoded binary representation of the entire original file
  • contents: Contains an array of structured Content Item objects extracted from the original file

The relationship works as follows:

  1. When a file is uploaded, its binary data is stored in the data field
  2. The original file's complete data is always preserved in the document's data field
  3. The file is then processed by content extractors based on file type (PDF, image, text, etc.)
  4. Each logically separate piece of content is added to the contents array
  5. For text files, there might be just one content item
  6. For PDFs, there might be multiple content items (one per page or per embedded image)
  7. For complex documents, content items might represent different sections or formats
  8. Each content item contains its own data field with the specific extracted content; for agents convenience it contains the additional field dataExtracted with extracted data based on agents task prompt

This dual structure allows agents to:

  • Access the complete original file when needed
  • Work with pre-processed, extracted content for efficiency
  • Process specific sections of a document without loading the entire file

State Transitions

[null] → [running]               // New workflow created
[running] → [completed]          // Workflow completes successfully
[running] → [stopped]            // User manually stops workflow
[running] → [failed]             // Error occurs during workflow
[completed] → [running]          // User continues workflow with new input (new round)
[stopped] → [running]            // User continues after manual stop (new round)
[failed] → [running]             // User retries workflow despite error (new round)
[any] → [null]                   // Workflow deleted

Exception Handling

Rules

  • Workflow status changes to "failed" on exceptions, all message and workflow generation exceptions to handle to ensure data consistency in the database
  • Errors are logged in workflow logs with type "error"
  • Produced project manager analysis output, inputs to agents, output from agents, workflow items, message items are all logged for debugging in the logger with type "debug"
  • HTTP exceptions are returned to the client with appropriate status codes
  • Failed agent tasks are recorded but don't stop the workflow

Workflow Stop Conditions

  • User explicitly cancels the workflow via the stop endpoint

Action to take:

  • workflow to set to "stopped" status

Workflow Failure Conditions

  • Unhandled exceptions in the main workflow execution path
  • Project manager analysis fails to generate a valid workplan
  • More than 50% of the agent tasks in the workplan fail to complete
  • Timeout exceeded (workflow runs longer than the configured maximum duration)
  • System resource limits exceeded (memory, CPU, etc.)

Action to take, when a workflow fails:

  • The last log entry will contain details about the failure reason
  • workflow to set to "failed" status

Workflow Exception Checkpoints

At the following points in the code the Workflow Execution routine is called:

  • Before adding or updating a message to the workflow
  • Before doing an API call

Special Notes

  1. Document Processing: Files uploaded by users are processed with content extraction to make them accessible to agents.
  2. AI Language Support: The system detects and adapts to the user's language.
  3. Round Counting: Each interaction increments the currentRound counter.
  4. Agent Registry: Agents are loaded dynamically and registered in the AgentRegistry.
  5. Standardized Task Processing: All agents implement the same task processing interface.