fix: use jsonb instead of json for pgvector

[devinburnette]

Description

When trying to integrate pgvector as a vector store with langchain, I got the following error:

2023-10-16 18:27:54.326 EDT [87858] ERROR:  operator does not exist: json @> unknown at character 106
2023-10-16 18:27:54.326 EDT [87858] HINT:  No operator matches the given name and argument types. You might need to add explicit type casts.
2023-10-16 18:27:54.326 EDT [87858] STATEMENT:
              SELECT *, embedding <=> $1 as "_distance"
              FROM langchain_pg_embedding
              WHERE cmetadata @> $2
              ORDER BY "_distance" ASC
              LIMIT $3;

The problem

From my brief research, it appears that the @> operator only works with jsonb. So I'm trying to figure out how this ever worked in the first place, but I'm not a postgres expert.

Testing / Reproduced

To test this, I ran the following against my database after langchain created it and this seemed to fix my issue:

ALTER TABLE langchain_pg_embedding ALTER COLUMN cmetadata SET DATA TYPE jsonb;

I'm using this python version of langchain for running my document loaders and storing my embeddings in PG, but then I have a chat client that's using the langchain-js lib to query it, so I happened to then go see if pgvector had the same problem over there, and it doesn't.
https://github.com/langchain-ai/langchainjs/blob/aa9b618786757c301e93caeb3253c56b56d05602/langchain/src/vectorstores/pgvector.ts#L244C38-L244C43

Let me know if the fix isn't as simple as this or if there's a better solution that should be considered like type casting or something else?

cc:
@baskaryan, @eyurtsev, @hwchase17

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

1 Ignored Deployment

Name	Status	Preview	Comments	Updated (UTC)
langchain	⬜️ Ignored (Inspect)	Visit Preview		Oct 17, 2023 2:30pm

[devinburnette]

also not sure why this was commented out, but seems like it should be part of the setup as well /cc @hwchase17

[devinburnette]

without this, i get the following error:

sqlalchemy.exc.ProgrammingError: (psycopg2.errors.UndefinedObject) type "vector" does not exist
LINE 4:  embedding VECTOR,
                   ^

[SQL:
CREATE TABLE langchain_pg_embedding (
	collection_id UUID,
	embedding VECTOR,
	document VARCHAR,
	cmetadata JSON,
	custom_id VARCHAR,
	uuid UUID NOT NULL,
	PRIMARY KEY (uuid),
	FOREIGN KEY(collection_id) REFERENCES langchain_pg_collection (uuid) ON DELETE CASCADE
)

]

[Sbargaoui]

The creation query doesn't raise an exception if it fails but rather a warning, proposed fix in #11930

[eyurtsev]

Closing in favor of: #18992