Fix metadata deserialization in async mode for PGVector

[shamspias]

Problem

When using asynchronous methods with PGVector (async_mode=True), the metadata field retrieved from the database may be of type Fragment (from asyncpg) or other non-dict types. This causes a ValidationError when the Document class expects metadata to be a dictionary.

Solution

This pull request modifies the _results_to_docs_and_scores method to ensure that metadata is correctly converted into a dictionary before creating Document instances. The method now handles different possible types of metadata and attempts to deserialize it into a dict.

Changes

• Updated _results_to_docs_and_scores method in PGVector class to handle metadata deserialization for different types (e.g., dict, str, Fragment).

Testing

• Tested with async_mode=True and confirmed that the metadata field is correctly deserialized and no longer causes validation errors.
• Ensured that the change does not affect the behavior when async_mode=False.

Related Issues

• Metadata field not properly deserialized when using async_mode=True with PGVector #124

[jaimeescano]

Facing this very same issue. Wondering when this commit could be merged.
Big credits to @shamspias for providing the fix.

Regards

[shamspias]

@eyurtsev, please let me know if there is anything else I need to do. If not, please merge it.

[simadimonyan]

Did you fix the issue? I have related: langchain-ai/langchain#28029

[shamspias]

Did you fix the issue? I have related: langchain-ai/langchain#28029

yes I did fixed the issue you can use the branch if you like to

[samsiuatpurple]

Good work! It has fixed an issue in my langgraph app. Can a maintainer merge this as it's a blocker for us. Thanks!

[lucelsbyleverageAI]

Hi @shamspias @samsiuatpurple - I just ran into the same issue and used your PR code. It prevented the error but now it's just returning empty metadata. Assume you aren't having the same?

[shamspias]

Hi @shamspias @samsiuatpurple - I just ran into the same issue and used your PR code. It prevented the error but now it's just returning empty metadata. Assume you aren't having the same?

It works fine, I run several project with this PR.

[fritzebner]

Can someone tell us when this change will be in a release? When is the next release for langchain-postgres?

[shamspias]

Can someone tell us when this change will be in a release? When is the next release for langchain-postgres?

same questions

[ccurme]

Hi @shamspias, are you interested in iterating on this? It still needs some work:

• The code breaks out the gate (json is not imported)
• Most importantly, there are no tests.

The most useful thing would be a unit test or reproducible example demonstrating how to generate non-dict metadata.