@pramudyalyza From what I can tell, there isn't currently a built-in "run once per document" metadata extraction stage in the ingestion pipeline. The behavior you're seeing is consistent with how RAGFlow's ingestion pipeline is designed today.

Here's why:

• The ingestion pipeline is fundamentally built around processing the parsed output and eventually producing chunks for indexing.
• The Transformer component operates on whatever upstream objects it receives.
  • If its input is TokenChunker/chunks, it naturally runs once per chunk.
  • If its input is Parser/json, the parser output is still a collection of parsed blocks (for DOC/DOCX/HTML/etc.), not a single document object, so the transformer is invoked for each parsed element rather than once for the entire file. The parser normalizes documents into structured block lists rather than exposing a dedicated document-level payload. :contentReference[oaicite:0]{index=0}

So the log message:

Start to generate meta-data for every chunk ...

is expected for the current Auto Metadata implementation as well. The documentation describes automatic metadata generation during parsing, but I couldn't find anything indicating that it performs a single LLM call per document instead of operating during chunk processing. :contentReference[oaicite:1]{index=1}

If your metadata is truly document-level

Fields like:

• business_function
• country
• product

are generally properties of the entire document, not individual chunks.

I think your expectation makes sense because inconsistent chunk-level values can reduce the usefulness of metadata filtering.

Current options

From what I could find, your options today are roughly:

1. Use document metadata instead of Auto Metadata

   • RAGFlow supports attaching metadata directly to uploaded documents.
   • That metadata is associated with the document and is available during retrieval.
   • This is ideal if you already know the values externally (or can compute them before upload). :contentReference[oaicite:2]{index=2}

2. Precompute metadata outside RAGFlow

   • Run one LLM call over the full document yourself.
   • Produce something like:

   {
     "business_function": "Finance",
     "country": "Singapore",
     "product": "Product A"
   }

   • Then attach it as document metadata before or after ingestion.

3. Custom pipeline (limited)

   • I couldn't find documentation showing a pipeline component that receives the entire document as a single object and invokes an LLM exactly once.
   • So today, the custom ingestion pipeline doesn't appear to expose a true document-level transformer stage. :contentReference[oaicite:3]{index=3}

It may be a feature gap

Personally, I think a dedicated Document Transformer stage would be valuable:

Parser
    ↓
Document Transformer   <-- runs exactly once
    ↓
Chunker
    ↓
Chunk Transformer
    ↓
Indexer

That would allow document-level metadata to be generated once and propagated to every chunk during indexing.

I couldn't find documentation or an existing feature that implements this workflow today, so this may simply be a current limitation rather than an incorrect pipeline configuration.

One question that would help narrow down the best approach:

• Are you trying to use these metadata fields primarily for retrieval filtering, or do you mainly want them available as context for generation after retrieval?

The recommended solution may differ depending on that use case.

If this solves your problem, feel free to mark it as the accepted answer so others can find it easily.