embedding-strategy

Retrieval pipelines return irrelevant results when the same embedding path processes both document ingestion and user queries, causing a semantic mismatch. Using separate embedding paths for documents and queries resolves this mismatch and improves search relevance.

Asymmetric embedding enforces separate task types for document ingestion and query processing. It applies RETRIEVAL_DOCUMENT for ingestion and RETRIEVAL_QUERY for querying, reducing semantic mismatch and improving retrieval quality across document collections.

Configure the embedding pipeline to route documents through the RETRIEVAL_DOCUMENT task type and route user queries through RETRIEVAL_QUERY. This strategy separates ingestion embeddings from query embeddings to maintain task-type separation and improve search relevance.

Yes, the pipeline allows embedding model selection and configuration for both ingestion and querying. It includes safeguards to validate correct task_type usage, ensuring that RETRIEVAL_DOCUMENT and RETRIEVAL_QUERY are applied properly across the pipeline.

You should use separate task types when documents and queries suffer from semantic mismatch in your retrieval pipeline. Enforcing asymmetric task types prevents retrieval errors by ensuring ingestion and query processing use distinct embedding paths.

Using a single embedding path for both ingestion and querying causes semantic mismatch between the document text and query intent. This leads to retrieval errors and reduced relevance, which asymmetric task_type separation is designed to prevent.