feat: Enable ingestion of precomputed embeddings (#2317)

llama_stack/apis/vector_io/vector_io.py

@@ -19,8 +19,16 @@ from llama_stack.schema_utils import json_schema_type, webmethod
 
 
 class Chunk(BaseModel):
+    """
+    A chunk of content that can be inserted into a vector database.
+    :param content: The content of the chunk, which can be interleaved text, images, or other types.
+    :param embedding: Optional embedding for the chunk. If not provided, it will be computed later.
+    :param metadata: Metadata associated with the chunk, such as document ID, source, or other relevant information.
+    """
+
     content: InterleavedContent
     metadata: dict[str, Any] = Field(default_factory=dict)
+    embedding: list[float] | None = None
 
 
 @json_schema_type
@@ -50,7 +58,10 @@ class VectorIO(Protocol):
         """Insert chunks into a vector database.
 
         :param vector_db_id: The identifier of the vector database to insert the chunks into.
-        :param chunks: The chunks to insert.
+        :param chunks: The chunks to insert. Each `Chunk` should contain content which can be interleaved text, images, or other types.
+            `metadata`: `dict[str, Any]` and `embedding`: `List[float]` are optional.
+            If `metadata` is provided, you configure how Llama Stack formats the chunk during generation.
+            If `embedding` is not provided, it will be computed later.
         :param ttl_seconds: The time to live of the chunks.
         """
         ...