1034 PDF Embedding Token Tracking Fix Ab2e8d78

PDF Embedding Token Tracking Fix

Issue

Version: 0.233.299
Date: December 19, 2025
Related Feature: Embedding Token Tracking (0.233.298)

Problem Description

After implementing embedding token tracking in version 0.233.298, PDF document uploads were showing 0 embedding tokens even though embeddings were being generated for all chunks.

Error Observed

Document c3076b27-e867-4081-a092-8ca1b2b46a1b (test.pdf) processed successfully with 7 chunks saved and 0 embedding tokens used.

Root Cause

The initial implementation only updated process_txt() to accumulate and return embedding token data. PDF files are processed through process_di_document() (Document Intelligence pathway), which was not updated to:

1. Capture token_usage from save_chunks() calls
2. Accumulate tokens across all chunks
3. Return token data as a tuple

Solution

Files Modified

functions_documents.py

1. Added token tracking initialization in process_di_document()

def process_di_document(...):
    # --- Token tracking initialization ---
    total_embedding_tokens = 0
    embedding_model_name = None

2. Updated save_chunks() call to capture token usage

# Before
save_chunks(**args)
total_final_chunks_processed += 1

# After
token_usage = save_chunks(**args)

# Accumulate embedding tokens
if token_usage:
    total_embedding_tokens += token_usage.get('total_tokens', 0)
    if not embedding_model_name:
        embedding_model_name = token_usage.get('model_deployment_name')

total_final_chunks_processed += 1

3. Updated return statement to include token data

# Before
return total_final_chunks_processed

# After
return total_final_chunks_processed, total_embedding_tokens, embedding_model_name

4. Updated process_document_upload_background() to handle tuple return

elif file_ext in di_supported_extensions:
    result = process_di_document(**args)
    # Handle tuple return (chunks, tokens, model_name)
    if isinstance(result, tuple) and len(result) == 3:
        total_chunks_saved, total_embedding_tokens, embedding_model_name = result
    else:
        total_chunks_saved = result

config.py

• Version incremented to 0.233.299

test_embedding_token_tracking.py

• Updated version to 0.233.299

Impact

Document Types Now Tracking Tokens

• ✅ PDF files (via Document Intelligence)
• ✅ DOCX files (via Document Intelligence)
• ✅ PPTX files (via Document Intelligence)
• ✅ Images (JPG, PNG, etc. via Document Intelligence)
• ✅ TXT files (direct processing)

Expected Behavior

When a PDF or other Document Intelligence-processed file is uploaded:

Document abc123 (test.pdf) processed successfully with 7 chunks saved and 1847 embedding tokens used.

The document metadata in Cosmos DB will now contain:

{
  "embedding_tokens": 1847,
  "embedding_model_deployment_name": "text-embedding-3-small"
}

Testing

Manual Testing

Upload a PDF file and verify:

1. Document processes successfully
2. Console shows non-zero embedding tokens
3. Cosmos DB document metadata contains embedding_tokens > 0
4. Cosmos DB document metadata contains embedding_model_deployment_name

Automated Testing

Run existing functional test:

python functional_tests\test_embedding_token_tracking.py

Related Issues

VectorizedQuery Serialization Warning

During testing, this warning may appear:

Error processing Hybrid search for document xxx: Unable to serialize value: [<azure.search.documents._generated.models._models_py3.VectorizedQuery object at 0x...>] as type: '[VectorQuery]'.

Status: This is a non-blocking warning in the metadata extraction phase. The document processing continues successfully, and this error is caught and handled gracefully. This is a separate issue related to the hybrid search functionality used for metadata extraction, not related to embedding token tracking.

Remaining Work

Other File Types to Update

The following process functions still need to be updated to track embedding tokens:

• process_xml()
• process_yaml()
• process_log()
• process_doc() (legacy .doc files)
• process_html()
• process_md()
• process_json()
• process_tabular() (CSV, XLSX, etc.)
• process_video_document()
• process_audio_document()

Group and Public Workspaces

Token tracking still needs to be implemented for:

• Group workspace documents
• Public workspace documents

Verification

After this fix, PDF uploads in personal workspaces should correctly track embedding tokens:

Before Fix:

embedding_tokens: 0
embedding_model_deployment_name: None

After Fix:

embedding_tokens: 1847
embedding_model_deployment_name: "text-embedding-3-small"

Related Documentation

• Embedding Token Tracking Feature
• Main implementation: application/single_app/functions_documents.py
• Test: functional_tests/test_embedding_token_tracking.py