The community member is using llamaindex with pgvector and text-embedding-large-3 with an embedding size of 1536. They are encountering an error when using the vector store, as the expected dimensions are 3072 but the embeddings are 1536. Another community member suggests swapping to ada-2, which works but produces incorrect results. The community members discuss that there may be a hardcoded 3072 dimension size somewhere, and confirm that the embeddings are stored with 1536 dimensions. A solution is provided to set the embed_dim to 3072 in the PGVectorStore, which resolves the initial error. However, the community member then encounters a new issue where the process stops mid-generation and gives an error, with no information in the backend logs. They ask for suggestions on where to look to resolve this new issue.
Hey folks! I'm using llamaindex with pgvector. I'm usiong text-embedding-large-3 with embedding size of 1536. When using the vector store and asking a question, I get this error: sqlalchemy.exc.StatementError: (builtins.ValueError) expected 1536 dimensions, not 3072 This is odd because embedding with 3072 doesn't work with pgvector due to postgres max dimensions. I'm assuming that there must be a dimension size that is hardcoded somewhere? Does anyone have a similar issue?
Swapping to ada-2 works, even though a different embedding model was used to embed it. The results are wrong, but at least it doesn't give an error. I'm assuming there is a hardcoded 3072 somewhere for what it is expecting. But we can actually use 1536 dimensions with embedding-3-large. Any thoughts, friends?
