Bug: PGVector.add_documents/add_embeddings silently truncates documents via zip and falsely reports all IDs as inserted

[pongpatapee]

Related to issue:
langchain-ai/langchain-google#1704

Hi, I ran into an issue with langchain_google_genai's embedding results. It silently returns 1 embedding when using gemini-embedding-2. The silent failure is further hidden by langchain-postgres PGVector.add_embeddings.

PGVector.add_embeddings builds its insert payload using zip(texts, metadatas, embeddings, ids_), which silently truncates to the shortest list. It then returns ids_ (the full input list) regardless of how many rows were actually written, so the caller sees N IDs returned with no exception, but only 1 row in the database.

Reproduce

Here's a script to reproduce the error:

from langchain_core.documents import Document
from langchain_google_genai import GoogleGenerativeAIEmbeddings
from langchain_postgres import PGVector
import psycopg2
import os
from dotenv import load_dotenv

load_dotenv()

GOOGLE_API_KEY = os.getenv("GOOGLE_API_KEY", "")
DATABASE_URL = os.getenv("DATABASE_URL", "")

embeddings = GoogleGenerativeAIEmbeddings(
    model="gemini-embedding-2",
    api_key=GOOGLE_API_KEY,
)

docs = [
    Document(
        page_content="there are cats in the pond",
        metadata={"id": 1, "location": "pond", "topic": "animals"},
    ),
    Document(
        page_content="ducks are also found in the pond",
        metadata={"id": 2, "location": "pond", "topic": "animals"},
    ),
    Document(
        page_content="fresh apples are available at the market",
        metadata={"id": 3, "location": "market", "topic": "food"},
    ),
]

result = embeddings.embed_documents([d.page_content for d in docs])
print(
    f"[BUG 1 - langchain-google-genai] embed_documents: passed {len(docs)} texts, got {len(result)} vectors"
)
# Expected: 3, Actual: 1


vector_store = PGVector(
    embeddings=embeddings,
    collection_name="rag_docs",
    connection=DATABASE_URL,
)

inserted_ids = vector_store.add_documents(docs)
print(
    f"[BUG 2 - langchain-postgres] add_documents: passed {len(docs)} docs, returned {len(inserted_ids)} IDs (false report)"
)
# Expected: 3 docs, 1 inserted (since embeddings only return 1 result)
# Actual: 3 docs, 3 inserted (hiding that only one embedding was processed)


conn = psycopg2.connect(DATABASE_URL)
cur = conn.cursor()
cur.execute("SELECT COUNT(*) FROM langchain_pg_embedding")
count = cur.fetchone()[0]
print(f"[BUG 2 - langchain-postgres] actual rows in DB: {count}")
# Actual: 1 row in DB

Root Cause

# langchain_postgres/vectorstores.py

def add_embeddings(
    self,
    texts: Sequence[str],
    embeddings: List[List[float]],
    metadatas: Optional[List[dict]] = None,
    ids: Optional[List[str]] = None,
    **kwargs: Any,
) -> List[str]:
    ...
    data = [
        {
            "id": id,
            "collection_id": collection.uuid,
            "embedding": embedding,
            "document": text,
            "cmetadata": metadata or {},
        }
        for text, metadata, embedding, id in zip(
            texts, metadatas, embeddings, ids_
        ) # embeddings truncates the loop
    ]
    ...

    return ids_ # returns full list

Suggested Fix

Before inserting the data, we can validate that the number of embeddings match the number of text inputs

 if len(embeddings) != len(texts):
      raise ValueError(
          f"Embedding count mismatch: expected {len(texts)}, got {len(embeddings)}. No documents inserted."
      )

Library Versions

langchain==1.2.15
langchain-classic==1.0.4
langchain-community==0.4.1
langchain-core==1.3.2
langchain-google-genai==4.2.2
langchain-postgres==0.0.17
langchain-protocol==0.0.12
langchain-text-splitters==1.1.2