Potential bug in multimodal vector retrieval: missing file attachments in segment_attachment_bindings table when using pipeline knowledge base

[shone5499]

Self Checks

• I have searched for existing issues search for existing issues, including closed ones.
• I confirm that I am using English to submit this report (我已阅读并同意 Language Policy).
• [FOR CHINESE USERS] 请务必使用英文提交 Issue，否则会被关闭。谢谢！:)
• Please do not modify this template :) and fill in all the required fields.

1. Is this request related to a challenge you're experiencing? Tell me about your story.

Hello Dify team and community!
I am testing the multimodal vectorization retrieval feature in the pipeline knowledge base of Dify version 1.14.2. The index mode in the pipeline knowledge base is set to "High Quality" mode, the Embedding model is configured as multimodal_embedding_v1, and the retrieval setting uses vector retrieval.

After successfully vectorizing an uploaded file that contains images, I performed a recall test by uploading an image, but it failed to match the corresponding paragraphs.

Upon investigation, I found that no corresponding file records were created in the segment_attachment_bindings table. Further debugging revealed that in the index function within paragraph_index_processor.py, under the line if isinstance(chunks, list):, the call self._get_content_files(doc) does not pass the current_user parameter. This causes the code inside the if block within the _get_content_files function to be skipped when the file path is in the format /files/tools/xxxx. c
code:
paragraph_index_processor.py

@override
    def index(self, dataset: Dataset, document: DatasetDocument, chunks: Any) -> None:
        documents: list[Any] = []
        all_multimodal_documents: list[Any] = []
        if isinstance(chunks, list):
            for content in chunks:
                metadata = {
                    "dataset_id": dataset.id,
                    "document_id": document.id,
                    "doc_id": str(uuid.uuid4()),
                    "doc_hash": helper.generate_text_hash(content),
                }
                doc = Document(page_content=content, metadata=metadata)
                attachments = self._get_content_files(doc)
                if attachments:
                    doc.attachments = attachments
                    all_multimodal_documents.extend(attachments)
                documents.append(doc)
        else:
            multimodal_general_structure = MultimodalGeneralStructureChunk.model_validate(chunks)
            for general_chunk in multimodal_general_structure.general_chunks:

index_process_base.py

def _get_content_files(self, document: Document, current_user: Account | None = None) -> list[AttachmentDocument]:
 # For tools directory - direct file formats (e.g., .png, .jpg, etc.)
            # Match URL including any query parameters up to common URL boundaries (space, parenthesis, quotes)
            pattern = r"/files/tools/([a-f0-9\-]+)\.([a-zA-Z0-9]+)(?:\?[^\s\)\"\']*)?"
            match = re.search(pattern, image)
            if match:
                if current_user:
                    tool_file_id = match.group(1)
                    upload_file_id = self._download_tool_file(tool_file_id, current_user)
                    if upload_file_id:
                        upload_file_id_list.append(upload_file_id)
                continue
            if current_user:
                upload_file_id = self._download_image(image.split(" ")[0], current_user)
                if upload_file_id:
                    upload_file_id_list.append(upload_file_id)

Is this a bug, or is it intentionally implemented this way?

2. Additional context or comments

In parent_child_index_processor.py，there has the same code too! code :self._get_content_files(document) no current_user parameter!

[dosubot[bot]]

Hi @shone5499! I'm Dosu and I'm helping the Dify team.

This does appear to be a legitimate bug. I've confirmed in the source code that the index method in paragraph_index_processor.py has two different code paths with inconsistent parameter passing:

• List-based chunks path (line 197): Calls self._get_content_files(doc) without current_user — this is the bug you identified [1].
• Multimodal general structure path (line 232): Correctly loads the account and passes it: self._get_content_files(doc, current_user=account) [2].

The same inconsistency exists in parent_child_index_processor.py — the FULL_DOC mode path at line 111 also omits current_user [3], while other paths in the same file pass it correctly.

When current_user is None, the _get_content_files method in index_processor_base.py silently skips downloading tool files (/files/tools/ paths) and external URLs via continue statements [4]. This explains why no records appear in the segment_attachment_bindings table for your pipeline knowledge base documents.

Notably, there's also a related issue #38079 that reports segment attachment images being indexed as text instead of multimodal vectors — a different but related multimodal indexing bug [5].

I didn't find an existing issue or PR that specifically addresses this missing current_user parameter. The fix would likely involve loading the account (similar to what the multimodal structure path already does) and passing it to _get_content_files in both affected locations. You could consider opening an issue or PR for this.

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.