AsyncPGVectorStore Metadata Filtering Not Working

[bharatht19]

I am seeing the below error while using meta data filtering as below .Please suggest.

Code

async def _build_chain():
    store = await get_vector_store()
    FILTER = {'source': {"$eq":"data/handbooks/employee_handbook_2025.docx"}}
    
    retriever = store.as_retriever(
    search_kwargs={
        "filter": FILTER
    }
)

DB Table

CREATE TABLE IF NOT EXISTS langchain_pg_embedding (
langchain_id uuid PRIMARY KEY DEFAULT gen_random_uuid(), -- Auto-generate UUIDs
content text, -- Raw text chunk
embedding vector(1536), -- Vector dimension matches embedding model
langchain_metadata jsonb -- Metadata
);

Error:

cka-app | sqlalchemy.exc.ProgrammingError: (sqlalchemy.dialects.postgresql.asyncpg.ProgrammingError) <class 'asyncpg.exceptions.UndefinedColumnError'>: column "source" does not exist
cka-app | [SQL: SELECT "langchain_id", "content", "embedding", "langchain_metadata", cosine_distance("embedding", $1) as distance
cka-app | FROM "public"."langchain_pg_embedding" WHERE source = $2 ORDER BY "embedding" <=> $1 LIMIT $3;
cka-app | ]
cka-app | [parameters: ('[0.019000347703695297, 0.008494899608194828, -0.0014854423934593797, -0.008188657462596893, -0.03637627884745598, -0.04138267785310745, -0.0131018515 ... (33892 characters truncated) ... 6108424216508865, -0.023460835218429565, 0.022315755486488342, 0.001256592688150704, 0.0203052069991827, 0.017788691446185112, -0.020238632336258888]', 'data/handbooks/employee_handbook_2025.docx', 4)]
cka-app | (Background on this error at: https://sqlalche.me/e/20/f405)

[averikitsch]

Hi @bharatht19 can you share the code behind function get_vector_store()?

[bharatht19]

@averikitsch - below is the code

import os
from langchain_openai import OpenAIEmbeddings
from langchain_postgres.v2.engine import PGEngine
from langchain_postgres.v2.async_vectorstore import AsyncPGVectorStore

PG_CONN_STR = os.getenv("DATABASE_URL")

PG_ENGINE = PGEngine.from_connection_string(PG_CONN_STR)

embeddings = OpenAIEmbeddings(model="text-embedding-3-small")


async def get_vector_store()->AsyncPGVectorStore:
    return await AsyncPGVectorStore.create(
        engine=PG_ENGINE,
        embedding_service=embeddings,
        table_name="langchain_pg_embedding",
        metadata_json_column="langchain_metadata"
    )

[bharatht19]

@averikitsch Did you get a chance to look in to this.This is stopping me from using metadata in my RAG pipeline. Is this a bug?

[dishaprakash]

Hi @bharatht19
PGVectorStore does not support filtering metadata stored in a JSONB column. For filtering to work, each metadata field must have its own separate column. This approach is designed to make filtering quicker and efficient in Postgres.

As a workaround, you can

• Mention the metadata column during the VectorStore initialization before adding the data and make sure the JSON key's value is being added into the metadata column OR

• Use SQL to ALTER your already existing table by adding a new column for the JSON key you want to filter and copying the data into it.
  Please see heading Search for documents with JSON Filter in this how-to guide.

Thank you for opening this issue, we will update the PGVectorStore documentation to reflect the same!

[mrodriguez2]

What's the reasoning behind removing the JSON filtering in the new PGVectorStore class? It's a pretty breaking change with any current Langchain implementation utilizing PGVector.

Langchain4j's PGVector implementation, for example, supports metadataStorageConfig with the following parameters:

• COLUMN_PER_KEY
• COMBINED_JSON
• COMBINED_JSONB

Maybe this would be a better approach?

[dishaprakash]

Our current design only permits filtering on the dedicated metadata columns, not on the JSONB field itself. This is a deliberate choice for performance, as filtering on standard columns are much more efficient in a PostgreSQL database. We recommend converting any filterable metadata from a JSONB field to a standard column.

This issue will be marked as a feature request for JSONB filtering in PGVectorStore