Integrated Inference Embed And Rerank

Operator embeds with multilingual-e5-large (1024 dim) via /embed and upserts into an index created with dimension=1536.

Match index dimension to embedder output. multilingual-e5-large=1024, llama-text-embed-v2=1024 (configurable), pinecone-sparse-english-v0=sparse. Recreate the index at the correct dimension; do not pad or truncate. Document the embedder choice with the index.

Operator calls /embed with input_type=passage for both corpus and queries on an asymmetric model (e.g. multilingual-e5-large).

Asymmetric embedders require input_type=passage for documents being indexed and input_type=query for search queries — the prefix changes the embedding. Mixing degrades retrieval. Verify per-model defaults [REQUIRES-VERIFICATION for the exact list of asymmetric models].

Operator queries top_k=100 from ANN, then reranks all 100 with /rerank, then takes top 5.

Set rerank top_n=5 explicitly. /rerank scores all input pairs and returns the top_n sorted descending. Trimming at the rerank layer avoids extra client-side sort and keeps the contract obvious in logs.