OpenAI "text-embedding-3-large" model: `vector's dimensions must be <= [2048]; got 3072`

[amotl]

Problem

When trying to store text embeddings using OpenAI's text-embedding-3-large into CrateDB, LangChain trips.

OpenAIEmbeddings(model="text-embedding-3-large")

sqlalchemy.exc.ProgrammingError: (crate.client.exceptions.ProgrammingError)
SQLParseException[Field [6] vector's dimensions must be <= [2048]; got 3072]
[SQL: INSERT INTO langchain_embedding (id, collection_id, embedding, document, cmetadata) VALUES (?, ?, ?, ?, ?) ON CONFLICT (id) DO UPDATE SET embedding = excluded.embedding, document = excluded.document, cmetadata = excluded.cmetadata]

Details

Processing the SQL DDL statement using FLOAT_VECTOR(3072) succeeds.

CREATE TABLE langchain_embedding (
  id STRING NOT NULL,
  collection_id STRING,
  embedding FLOAT_VECTOR(3072),
  document STRING,
  cmetadata OBJECT,
  PRIMARY KEY (id)
)

The error happens when processing the INSERT statement.

INSERT INTO langchain_embedding (id, collection_id, embedding, document, cmetadata)
VALUES (?, ?, ?, ?, ?) ON CONFLICT (id) DO
UPDATE
SET embedding = excluded.embedding,
    document = excluded.document,
    cmetadata = excluded.cmetadata

---

This is the engine's FLOAT_VECTOR dimension cap surfacing at insert time — the DDL accepts 3072, the write path enforces 2048 — and it hits the defaults of current models: OpenAI's text-embedding-3-large and Gemini embeddings both produce 3072 dimensions out of the box, so this is many users' first-run experience. On the integration side: fail at construction with a clear message instead of deep inside an INSERT, and pass through the dimensions parameter both providers support, so users can fit under the cap deliberately. The cap itself needs a core-side conversation.