Skip to main content

Streaming Text Chunking Fix

Status: 🟢 In-use Category: Features & Enhancements Date: October 31, 2025

Problem​

When the agent streams responses, words were being split with extra spaces:

Before:

Looking up AWS Resources...I'll help you fin d the cost associated with the 'comn' EKS cluster.
Let me search for costs that include this cluster name in the usage details an d relate d expenses.
Let me try a different approach. I'll look for tag -based filtering that might include the cluster name:
Great ! I found evidence of the 'comn ' cluster through the Kubernetes cluster tag .
Now let me get costs associate d with this cluster using the tag filter :
Perfect! Now let me get a detaile d breakdown of the E K S- specific costs for the ' com n ' cluster :

Issues:

  • "fin d" should be "find"
  • "an d relate d" should be "and related"
  • "tag -based" should be "tag-based"
  • "associate d" should be "associated"
  • "detaile d" should be "detailed"
  • "E K S" should be "EKS"
  • "com n" should be "comn"

Root Cause​

The streaming text accumulation logic in AgentForgePage.tsx was adding spaces between chunks unnecessarily. The "smart spacing" logic was trying to be helpful but was actually breaking words:

// OLD CODE - INCORRECT
if (/[a-zA-Z0-9]/.test(lastChar) && /[a-zA-Z0-9]/.test(firstChar)) {
  if (!/[\s.,!?;:]/.test(firstChar)) {
    accumulatedText += ` ${cleanText}`;  // ❌ Adding space between chunks
  } else {
    accumulatedText += cleanText;
  }
} else {
  accumulatedText += cleanText;
}

When the server sends chunks like:

  1. "fin"
  2. "d the cost"

The old logic would detect:

  • Last char of "fin" = "n" (alphanumeric)
  • First char of "d the cost" = "d" (alphanumeric)
  • Result: Add space → "fin d the cost" ❌

Solution​

Removed the "smart spacing" logic and just concatenate chunks directly:

// NEW CODE - CORRECT
} else {
  // Append to existing text - direct concatenation
  // The server sends properly chunked text, just concatenate without adding spaces
  console.log('APPENDING to existing text (direct concat)');
  accumulatedText += cleanText;
}

After:

Looking up AWS Resources...I'll help you find the cost associated with the 'comn' EKS cluster.
Let me search for costs that include this cluster name in the usage details and related expenses.
Let me try a different approach. I'll look for tag-based filtering that might include the cluster name:
Great! I found evidence of the 'comn' cluster through the Kubernetes cluster tag.
Now let me get costs associated with this cluster using the tag filter:
Perfect! Now let me get a detailed breakdown of the EKS-specific costs for the 'comn' cluster:

Technical Details​

Location​

File: workspaces/agent-forge/plugins/agent-forge/src/components/AgentForgePage.tsx Lines: ~1907-1912

The Fix​

- } else {
-   // Append to existing text with smart spacing
-   console.log('APPENDING to existing text');
-
-   // Add spacing logic to prevent words from running together
-   if (accumulatedText && cleanText) {
-     const lastChar = accumulatedText.slice(-1);
-     const firstChar = cleanText.slice(0, 1);
-
-     // Add space if both are alphanumeric and no space exists
-     if (/[a-zA-Z0-9]/.test(lastChar) && /[a-zA-Z0-9]/.test(firstChar)) {
-       // Don't add space if the new text already starts with punctuation or whitespace
-       if (!/[\s.,!?;:]/.test(firstChar)) {
-         accumulatedText += ` ${cleanText}`;
-       } else {
-         accumulatedText += cleanText;
-       }
-     } else {
-       accumulatedText += cleanText;
-     }
-   } else {
-     accumulatedText += cleanText;
-   }
- }

+ } else {
+   // Append to existing text - direct concatenation
+   // The server sends properly chunked text, just concatenate without adding spaces
+   console.log('APPENDING to existing text (direct concat)');
+   accumulatedText += cleanText;
+ }

Why This Works​

  1. Server Responsibility: The server (agent) is responsible for sending properly formatted text chunks
  2. Chunk Boundaries: Text chunks may split at any character position, not just word boundaries
  3. Preserve Integrity: Client should preserve the exact text as received, not modify spacing
  4. Simple is Better: Direct concatenation is simpler and more reliable than trying to guess spacing

Testing​

To verify the fix:

  1. Ask the agent a question that requires multiple streaming chunks
  2. Watch for words that were previously split (like "find", "and", "related", "EKS")
  3. Verify text appears correctly without extra spaces mid-word
  4. Check that proper spacing between sentences is preserved
  • AgentForgePage.tsx - Main streaming logic (FIXED)
  • No other files needed changes

Impact​

  • ✅ Words no longer split mid-character
  • ✅ Natural reading flow restored
  • ✅ Proper spacing preserved
  • ✅ Simpler, more maintainable code
  • ✅ No impact on non-streaming responses

Notes​

The "smart spacing" logic was originally added to handle cases where chunks might not have proper spacing. However, in practice:

  • The SSE stream from the server already includes proper spacing
  • Text chunks can split at any UTF-8 character boundary
  • Adding spaces based on character type creates more problems than it solves
  • Direct concatenation is the correct approach for SSE text streaming