Google Gemini (embeddings)

Gemini covers more modalities than any other embedding provider on the list — text, image, audio, video, PDF — all in one vector space, on gemini-embedding-2. The v1 line stays useful too: text-only but with the full Google task enum (SEMANTIC_SIMILARITY, CLASSIFICATION, CLUSTERING, QUESTION_ANSWERING, …).

Install

pnpm add @effect-uai/core @effect-uai/google effect

Wire it up

import { Config, Effect, Layer } from "effect"
import { FetchHttpClient } from "effect/unstable/http"
import { layer as geminiEmbeddingLayer } from "@effect-uai/google/GeminiEmbedding"

const provider = Layer.unwrap(
  Effect.gen(function* () {
    const apiKey = yield* Config.redacted("GOOGLE_API_KEY")
    return geminiEmbeddingLayer({ apiKey })
  }),
)

const mainLayer = provider.pipe(Layer.provide(FetchHttpClient.layer))

geminiEmbeddingLayer registers two service tags from one underlying implementation:

GeminiEmbedding — the typed tag. Yield this when you want the full task enum, document title, or the typed model union.
EmbeddingModel — the generic tag for provider-portable code.

Request shape

interface GeminiEmbedRequest extends Omit<CommonEmbedRequest, "model" | "task" | "encoding"> {
  readonly model: GoogleEmbeddingModel
  readonly task?: GoogleEmbeddingTask
  readonly title?: string
  readonly dimensions?: number
}

type GoogleEmbeddingTask =
  | "query"
  | "document" // cross-provider denominator
  | "similarity"
  | "classification"
  | "clustering"
  | "qa"
  | "fact_verification"
  | "code_query"

task — Gemini’s full task enum. Mapped internally to wire values (RETRIEVAL_QUERY, SEMANTIC_SIMILARITY, …). Honoured by gemini-embedding-001. Ignored by gemini-embedding-2 — that model expects task hints as prefix text in the prompt itself.
title — optional document title for RETRIEVAL_DOCUMENT tasks on gemini-embedding-001. Ignored elsewhere.
dimensions — Matryoshka truncation. gemini-embedding-2 supports 128–3072; gemini-embedding-001 supports 128/256/512/1408.

Calling it

Multimodal call on v2:

import { GeminiEmbedding } from "@effect-uai/google/GeminiEmbedding"
import * as Image from "@effect-uai/core/Image"

const program = Effect.gen(function* () {
  const gemini = yield* GeminiEmbedding
  return gemini.embedMany({
    model: "gemini-embedding-2",
    inputs: [
      { text: "A photo of artisan sourdough bread" },
      { image: Image.imageBytes(jpegBytes, "image/jpeg") },
    ],
  })
})

Task-typed call on v1:

const program = Effect.gen(function* () {
  const gemini = yield* GeminiEmbedding
  return gemini.embed({
    model: "gemini-embedding-001",
    input: "What is the boiling point of water?",
    task: "qa",
    title: "Physics FAQ",
  })
})

Models

GoogleEmbeddingModel is a literal union with a (string & {}) tail:

Model	Modalities	Task field	Native dims
`gemini-embedding-2`	text, image, audio, video, PDF	ignored (use prefix)	up to 3072
`gemini-embedding-001`	text only	full enum	128/256/512/1408

Reference: Gemini embeddings docs.

Encoding support

`encoding`	Behaviour
`float32` (default)	Native float32 from `embedding.values`.
`int8` / `binary`	Rejected — Gemini doesn’t ship quantized output.
`sparse` / `multivector`	Rejected — same.

If you need quantized vectors against a Gemini index, Jina ships binary natively.

Image input shapes

Gemini’s embed endpoint takes inline base64. The layer accepts:

base64 — passed through.
bytes — auto-encoded to base64.
url — rejected with AiError.InvalidRequest. Pre-uploading via Google’s Files API isn’t free, so URL-form image inputs fail fast rather than silently doing extra work. Fetch the bytes and pass imageBytes(...) instead.

Errors

HTTP failures map to typed AiError variants:

Status	Error
`429`	`AiError.RateLimited`
`408`/`504`	`AiError.Timeout`
`401`	`AiError.AuthFailed` (`auth`)
`403`	`AiError.AuthFailed` (`permission`)
`402`	`AiError.AuthFailed` (`billing`)
`413`	`AiError.ContextLengthExceeded`
`>= 500`	`AiError.Unavailable`
other 4xx	`AiError.InvalidRequest`

Gemini’s embed endpoints don’t return token-count metadata, so usage.inputTokens comes back undefined rather than estimated.