Generates vector embeddings for one or more inputs using a catalog
embedding model (model_category = embedding).

Request/response shapes match OpenAI's embeddings API (without the
OpenAI-only user abuse-monitoring field).

OpenAI params: model, input (string | string[] | token ids),
optional dimensions, optional encoding_format (float|base64).

Gemini text knobs (gemini-embedding-001 only): optional taskType,
optional title. These are rejected for gemini-embedding-2 (use task
prefixes in the input text instead).

Gemini multimodal (gemini-embedding-2 only): input may also be an
array of chat-style content blocks (text / image / audio / video)
for one aggregated embedding, or an array of strings and/or content-block
arrays for one vector per item. Media must be inline source.type=base64.
OpenAI models and gemini-embedding-001 reject content blocks (400).

Billing: charged per prompt/input token from llm_model_pricing_tiers
(Agent700 markup × 1.40). Completion token rates are zero for embedding models.
Usage is logged with request_channel = embeddings.

List valid models with GET /api/models?category=embedding. Embedding models
cannot be assigned to agents or used with POST /api/chat.