Embeddings And Vector Stores

A Pinecone index is created with dimension=1536 (for an older embedder) but the configured embed model now outputs 3072-dim vectors; upserts fail or are silently rejected.

Ensure the vector store's configured dimension exactly matches the embedding model's output dimension. On an embedder change that alters dimension, create a new collection/index at the right dimension and re-embed — you cannot mix dimensions in one space. Verify dimension before bulk upsert.

An integrator uses an instruction-tuned/asymmetric embedding model but embeds queries with get_text_embedding instead of get_query_embedding, hurting retrieval quality.

Use the query-side embedding path (get_query_embedding / the retriever's built-in query embedding) so any query prefix/instruction the model expects is applied. For asymmetric models, query and document embeddings are produced differently; LlamaIndex's retriever handles this when used correctly.

Bulk-embedding 500k nodes with a hosted embedding API hits provider rate limits; the integrator retries the whole batch on any 429, repeatedly re-embedding already-done nodes.

Tune embed_batch_size and concurrency to stay within provider limits, retry with backoff at the batch level, and checkpoint progress (via the ingestion docstore/cache) so a 429 does not re-embed completed nodes. Track embedding spend against a budget.