Flowise AgentFlow V2 Not Returning sourceDocuments from Chatflow

[abdelhafid-cloud]

Describe the bug

When a Chatflow that uses RAG is executed directly in Flowise, the response correctly includes the retrieved document sources (sourceDocuments). However, when the same Chatflow is called inside an AgentFlow V2 using the Execute Flow node, the sources are no longer returned.

The Chatflow itself works correctly:

• It uses Cohere Embeddings and Qdrant as the vector database.
• RAG retrieval works as expected.
• When tested directly, the response includes the associated sourceDocuments.
• But when the Chatflow is executed through AgentFlow V2 → Execute Flow:
• The Chatflow runs successfully.
• The text response is returned correctly.
• The sourceDocuments field is missing from the response.

As a result, the RAG sources are lost when the Chatflow is executed from AgentFlow V2, even though they are available when running the Chatflow directly.

To Reproduce

1. Create a Chatflow in Flowise that uses:

• Cohere Embeddings

• Qdrant Vector Database

• A RAG pipeline that retrieves documents.

2. Test the Chatflow directly in the Flowise chat interface.

3. Ask a question related to the indexed documents.

4. Observe that the response includes the retrieved sourceDocuments.

5. Create an AgentFlow V2.

6. Add an Execute Flow node and configure it to call the previously created Chatflow.

7. Run the AgentFlow and ask the same question.

8. Observe the response returned by the agent.

Result:
The response text is returned correctly, but the sourceDocuments field is missing.

Expected behavior

When a Chatflow that performs RAG retrieval is executed inside AgentFlow V2 using the Execute Flow node, the response should include the same sourceDocuments as when the Chatflow is executed directly.

The AgentFlow should propagate the full response from the Chatflow, including:

• The generated text response
• The retrieved sourceDocuments (RAG sources)

This would allow the agent to preserve and use document sources even when the Chatflow is called as part of an AgentFlow.

Screenshots

No response

Flow

No response

Use Method

None

Flowise Version

3.0.4

Operating System

None

Browser

None

Additional context

No response

[BadrLaajali]

Confirming the Issue with Technical Details

I can confirm this bug exists and have analyzed the root cause in the codebase. The issue is in the ExecuteFlow_Agentflow class located at /packages/components/nodes/agentflow/ExecuteFlow/ExecuteFlow.ts.

Technical Root Cause

The problem occurs in lines 207-210 of the Execute Flow node implementation:

let resultText = ''
if (response.data.text) resultText = response.data.text
else if (response.data.json) resultText = '```json\n' + JSON.stringify(response.data.json, null, 2)
else resultText = JSON.stringify(response.data, null, 2)

The Execute Flow node only extracts and preserves the text content from the Chatflow response, but discards all metadata including:

• sourceDocuments (RAG sources)
• usedTools
• artifacts
• fileAnnotations
• Any other response metadata

Why This Happens

1. Design Simplification: The Execute Flow node was designed to return only the text content for simplicity
2. Metadata Loss: When processing the API response, only response.data.text is preserved, while response.data.sourceDocuments and other fields are ignored
3. Return Object Structure: The current return object only includes content in the output:

const returnOutput = {
    // ... other fields
    output: {
        content: resultText  // ← Only text, no metadata
    },
    // ...
}

Proposed Solution

Here's the fix needed in /packages/components/nodes/agentflow/ExecuteFlow/ExecuteFlow.ts:

1. Extract and Preserve Metadata

Replace lines 207-210 with:

let resultText = ''
let sourceDocuments = null
let usedTools = null
let artifacts = null
let fileAnnotations = null

if (response.data.text) {
    resultText = response.data.text
    sourceDocuments = response.data.sourceDocuments || null
    usedTools = response.data.usedTools || null
    artifacts = response.data.artifacts || null
    fileAnnotations = response.data.fileAnnotations || null
} else if (response.data.json) {
    resultText = '```json\n' + JSON.stringify(response.data.json, null, 2)
} else {
    resultText = JSON.stringify(response.data, null, 2)
}

2. Include Metadata in Return Object

Modify the returnOutput object (lines 236-259):

const returnOutput = {
    id: nodeData.id,
    name: this.name,
    input: {
        messages: [
            {
                role: 'user',
                content: flowInput
            }
        ]
    },
    output: {
        content: resultText,
        sourceDocuments,      // ← Added
        usedTools,            // ← Added
        artifacts,            // ← Added
        fileAnnotations       // ← Added
    },
    state: newState,
    chatHistory: [
        ...inputMessages,
        {
            role: returnRole,
            content: resultText,
            name: nodeData?.label ? nodeData?.label.toLowerCase().replace(/\s/g, '_').trim() : nodeData?.id
        }
    ]
}

Additional Considerations

1. Backward Compatibility: This change maintains backward compatibility as existing code using only output.content will continue to work
2. Sequential Agents: The same issue likely exists in the Sequential Agents Execute Flow node at /packages/components/nodes/sequentialagents/ExecuteFlow/ExecuteFlow.ts
3. Error Handling: Similar metadata preservation should be added to the error response section

Impact

This fix would ensure that:

• RAG sources are preserved when Chatflows are called from AgentFlows
• All response metadata is properly propagated
• The AgentFlow has access to the same information as direct Chatflow execution
• Downstream nodes in the AgentFlow can access and use the source documents

This is a critical fix for anyone using RAG systems within AgentFlows, as it preserves the provenance and traceability of retrieved documents.

[JiwaniZakir]

The issue is in the Execute Flow node — when it calls the child chatflow, it's only propagating the text response and dropping the sourceDocuments field from the result. I'll trace through packages/components/nodes/agentflow/ExecuteFlow to make sure the full response object (including sourceDocuments and usedTools) gets passed back to the parent AgentFlow instead of just the text content. Should be a straightforward fix in how the node handles the chatflow's return value.

[JiwaniZakir]

Hey, stepping back from this one. Ran into some tooling issues trying to trace the sourceDocuments flow through the AgentFlow V2 path and couldn't make solid progress, so freeing it up for someone else.