[BUG] Incompatible metadata extraction for Qdrant Retriever with Extract Metadata Retriever

[kirisame-wang]

Describe the bug
Qdrant Retriever is incompatible with the newly added Extract Metadata Retriever feature from #3579.

To Reproduce
Steps to reproduce the behavior:
Set up a tool agent with the retriever tool chained with Qdrant as the vector store and Extract Metadata Retriever as the retriever.

Expected behavior
The Extract Metadata Retriever should extract metadata in the form of key-value pairs within a JSON-structured output. This metadata should then be used to retrieve relevant documents based on the specified filter.

Screenshots

Additional context
It seems that the Extract Metadata Retriever currently only extracts unnested key-value pairs as metadata. However, Qdrant requires nested keys for filtering (see the documentation), which is causing compatibility issues.

[HenryHengZJ]

right, because the extracted metadata is simply added as plain JSON object to filter. For vector store like Qdrant which require specifix syntax, it will not work.

i think we can have a field parameter with type=code, that allow users to return the filter in their own way

[kirisame-wang]

@HenryHengZJ Thanks for replying!
Is it possible to set nested schemas (see OpenAI's documentation) using a method similar to #3669 and #3695 to get values with a nested metadata filter?

[jmelati]

Hello! Is this fix, to use Extract Metadata Retriever in conjunction with QDrant, in production? Or is there no solution for this? For now, I'm using PineCone, but I would like to use it in conjunction with QDrant.

[BadrLaajali]

Any news about this topic ?