Embed

Agent embeds corpus passages and user queries with the same input_type, then wonders why retrieval recall is poor.

Set input_type='search_document' when embedding corpus passages for indexing and input_type='search_query' when embedding the query at search time. The asymmetric input_type is required for the retrieval embedding space to align.

A downstream classifier is trained on embeddings produced with input_type='search_document', then serves with input_type='classification'.

Use input_type='classification' for classification features and input_type='clustering' for clustering, and keep the same input_type at train and inference time. Mixing input_types between training and serving shifts the embedding distribution and degrades the model.

Documents longer than the embed model's max token length are sent whole; the agent assumes the full document is represented in the vector.

Chunk documents to the embed model's documented max token length and embed chunks (storing chunk→parent mapping) rather than relying on silent truncation that drops the tail. Exact max length is [REQUIRES-VERIFICATION] per model version.