feat(ai): support image-guided generation (#37998)
Co-authored-by: Aiden Cline <rekram1-node@users.noreply.github.com>
This commit is contained in:
opencode-agent[bot]
GitHub
parent
fa4ec98e1d
commit
e93fa06d05
19 changed files with 756 additions and 80 deletions
|
|
@ -1,6 +1,13 @@
|
|||
import { Effect, Encoding, Schema } from "effect"
|
||||
import { Headers, HttpClientRequest } from "effect/unstable/http"
|
||||
import { GeneratedImage, ImageModel, ImageResponse, type ImageRequestFor, type ImageRoute } from "../image"
|
||||
import {
|
||||
GeneratedImage,
|
||||
ImageModel,
|
||||
ImageResponse,
|
||||
type ImageInput,
|
||||
type ImageRequestFor,
|
||||
type ImageRoute,
|
||||
} from "../image"
|
||||
import { Auth, type Definition as AuthDefinition } from "../route/auth"
|
||||
import {
|
||||
InvalidProviderOutputReason,
|
||||
|
|
@ -12,6 +19,7 @@ import {
|
|||
type ProviderMetadata,
|
||||
} from "../schema"
|
||||
import { ProviderShared } from "./shared"
|
||||
import { ImageInputs } from "./utils/image-input"
|
||||
|
||||
const ADAPTER = "google-images"
|
||||
export const DEFAULT_BASE_URL = "https://generativelanguage.googleapis.com/v1beta"
|
||||
|
|
@ -31,7 +39,7 @@ export type GoogleImageOptions = {
|
|||
export type GoogleImageBody = Record<string, unknown> & {
|
||||
readonly contents: ReadonlyArray<{
|
||||
readonly role: "user"
|
||||
readonly parts: ReadonlyArray<{ readonly text: string }>
|
||||
readonly parts: ReadonlyArray<Record<string, unknown>>
|
||||
}>
|
||||
readonly generationConfig: Record<string, unknown>
|
||||
}
|
||||
|
|
@ -116,15 +124,6 @@ const nativeOptions = (options: GoogleImageOptions | undefined) => {
|
|||
)
|
||||
}
|
||||
|
||||
const body = (request: ImageRequestFor<GoogleImageOptions>, overlay: Record<string, unknown> | undefined) =>
|
||||
mergeJsonRecords(
|
||||
{
|
||||
contents: [{ role: "user", parts: [{ text: request.prompt }] }],
|
||||
generationConfig: nativeOptions(request.options),
|
||||
},
|
||||
overlay,
|
||||
) as GoogleImageBody
|
||||
|
||||
const invalidOutput = (message: string, providerMetadata?: ProviderMetadata) =>
|
||||
new LLMError({
|
||||
module: ADAPTER,
|
||||
|
|
@ -143,8 +142,16 @@ export const model = (input: ModelInput) => {
|
|||
const route: ImageRoute<GoogleImageOptions> = {
|
||||
id: ADAPTER,
|
||||
generate: Effect.fn("GoogleImages.generate")(function* (request: ImageRequestFor<GoogleImageOptions>, execute) {
|
||||
const imageParts = yield* Effect.forEach(request.images ?? [], googleImagePart)
|
||||
const http = mergeHttpOptions(request.model.http, request.http)
|
||||
const text = ProviderShared.encodeJson(body(request, http?.body))
|
||||
const requestBody = mergeJsonRecords(
|
||||
{
|
||||
contents: [{ role: "user", parts: [{ text: request.prompt }, ...imageParts] }],
|
||||
generationConfig: nativeOptions(request.options),
|
||||
},
|
||||
http?.body,
|
||||
) as GoogleImageBody
|
||||
const text = ProviderShared.encodeJson(requestBody)
|
||||
const url = applyQuery(
|
||||
`${(input.baseURL ?? DEFAULT_BASE_URL).replace(/\/$/, "")}/models/${request.model.id}:generateContent`,
|
||||
http?.query,
|
||||
|
|
@ -278,6 +285,30 @@ export const model = (input: ModelInput) => {
|
|||
return ImageModel.make<GoogleImageOptions>({ id: input.id, provider: "google", route, http: input.http })
|
||||
}
|
||||
|
||||
const googleImagePart = (image: ImageInput): Effect.Effect<Record<string, unknown>, LLMError> => {
|
||||
if (image.type === "bytes")
|
||||
return Effect.succeed({ inlineData: { mimeType: image.mediaType, data: Encoding.encodeBase64(image.data) } })
|
||||
if (image.type === "file-uri") return Effect.succeed({ fileData: { mimeType: image.mediaType, fileUri: image.uri } })
|
||||
if (image.type === "url")
|
||||
return ImageInputs.decodeDataUrl(image.url, ADAPTER).pipe(
|
||||
Effect.flatMap((decoded) => {
|
||||
if (decoded === undefined)
|
||||
return Effect.fail(
|
||||
ImageInputs.invalid(
|
||||
ADAPTER,
|
||||
"Google generateContent does not fetch public image URLs; use bytes, a data URL, or a Gemini file URI",
|
||||
),
|
||||
)
|
||||
return Effect.succeed({
|
||||
inlineData: { mimeType: decoded.mediaType, data: Encoding.encodeBase64(decoded.data) },
|
||||
})
|
||||
}),
|
||||
)
|
||||
return Effect.fail(
|
||||
ImageInputs.invalid(ADAPTER, "Google generateContent requires Gemini file URIs rather than provider file IDs"),
|
||||
)
|
||||
}
|
||||
|
||||
export const GoogleImages = {
|
||||
model,
|
||||
} as const
|
||||
|
|
|
|||
|
|
@ -1,6 +1,13 @@
|
|||
import { Effect, Encoding, Schema } from "effect"
|
||||
import { Headers, HttpClientRequest } from "effect/unstable/http"
|
||||
import { ImageModel, GeneratedImage, ImageResponse, type ImageRequestFor, type ImageRoute } from "../image"
|
||||
import { Headers, HttpClientRequest, HttpClientResponse } from "effect/unstable/http"
|
||||
import {
|
||||
ImageModel,
|
||||
GeneratedImage,
|
||||
ImageResponse,
|
||||
type ImageInput,
|
||||
type ImageRequestFor,
|
||||
type ImageRoute,
|
||||
} from "../image"
|
||||
import { Auth, type Definition as AuthDefinition } from "../route/auth"
|
||||
import {
|
||||
InvalidProviderOutputReason,
|
||||
|
|
@ -11,15 +18,18 @@ import {
|
|||
type HttpOptions,
|
||||
} from "../schema"
|
||||
import { ProviderShared } from "./shared"
|
||||
import { ImageInputs } from "./utils/image-input"
|
||||
import { OpenAIImage } from "./utils/openai-image"
|
||||
|
||||
const ADAPTER = "openai-images"
|
||||
export const DEFAULT_BASE_URL = "https://api.openai.com/v1"
|
||||
export const PATH = "/images/generations"
|
||||
export const EDIT_PATH = "/images/edits"
|
||||
|
||||
export type OpenAIImageString<Known extends string> = Known | (string & {})
|
||||
|
||||
export type OpenAIImageOptions = {
|
||||
readonly mask?: ImageInput
|
||||
readonly n?: number
|
||||
readonly size?: OpenAIImageString<
|
||||
"auto" | "256x256" | "512x512" | "1024x1024" | "1536x1024" | "1024x1536" | "1792x1024" | "1024x1792"
|
||||
|
|
@ -64,9 +74,9 @@ export interface ModelInput {
|
|||
readonly http?: HttpOptions
|
||||
}
|
||||
|
||||
const nativeOptions = (options: Record<string, unknown> | undefined) => {
|
||||
const nativeOptions = (options: OpenAIImageOptions | undefined) => {
|
||||
if (!options) return undefined
|
||||
const { outputFormat, outputCompression, ...native } = options
|
||||
const { mask: _, outputFormat, outputCompression, ...native } = options
|
||||
return {
|
||||
output_format: outputFormat,
|
||||
output_compression: outputCompression,
|
||||
|
|
@ -92,14 +102,89 @@ export const model = (input: ModelInput) => {
|
|||
const route: ImageRoute<OpenAIImageOptions> = {
|
||||
id: ADAPTER,
|
||||
generate: Effect.fn("OpenAIImages.generate")(function* (request: ImageRequestFor<OpenAIImageOptions>, execute) {
|
||||
const mask = request.options?.mask
|
||||
if (mask !== undefined && (request.images?.length ?? 0) === 0)
|
||||
return yield* ImageInputs.invalid(ADAPTER, "An OpenAI image mask requires at least one input image")
|
||||
const http = mergeHttpOptions(request.model.http, request.http)
|
||||
const sourceImages = request.images ?? []
|
||||
const multipartImages = yield* Effect.forEach(sourceImages, (image) => {
|
||||
if (image.type === "bytes") return Effect.succeed({ data: image.data, mediaType: image.mediaType })
|
||||
if (image.type === "url") return ImageInputs.decodeDataUrl(image.url, ADAPTER)
|
||||
return Effect.succeed(undefined)
|
||||
})
|
||||
const multipartMask =
|
||||
mask === undefined
|
||||
? undefined
|
||||
: mask.type === "bytes"
|
||||
? { data: mask.data, mediaType: mask.mediaType }
|
||||
: mask.type === "url"
|
||||
? yield* ImageInputs.decodeDataUrl(mask.url, ADAPTER)
|
||||
: undefined
|
||||
const useMultipart =
|
||||
sourceImages.length > 0 &&
|
||||
multipartImages.every((image) => image !== undefined) &&
|
||||
(mask === undefined || multipartMask !== undefined)
|
||||
const path = sourceImages.length === 0 ? PATH : EDIT_PATH
|
||||
const url = applyQuery(`${(input.baseURL ?? DEFAULT_BASE_URL).replace(/\/$/, "")}${path}`, http?.query)
|
||||
|
||||
if (useMultipart) {
|
||||
const form = new FormData()
|
||||
form.append("model", request.model.id)
|
||||
form.append("prompt", request.prompt)
|
||||
Object.entries(mergeJsonRecords(nativeOptions(request.options), http?.body) ?? {}).forEach(([key, value]) => {
|
||||
if (["model", "prompt", "image", "image[]", "images", "mask"].includes(key)) return
|
||||
form.append(key, typeof value === "string" ? value : ProviderShared.encodeJson(value))
|
||||
})
|
||||
multipartImages.forEach((image, index) => {
|
||||
if (image === undefined) return
|
||||
form.append("image[]", imageBlob(image.data, image.mediaType), `image-${index}`)
|
||||
})
|
||||
if (multipartMask !== undefined)
|
||||
form.append("mask", imageBlob(multipartMask.data, multipartMask.mediaType), "mask")
|
||||
const headers = yield* Auth.toEffect(input.auth)({
|
||||
request,
|
||||
method: "POST",
|
||||
url,
|
||||
body: "[multipart/form-data]",
|
||||
headers: Headers.remove(Headers.fromInput({ ...input.headers, ...http?.headers }), "content-type"),
|
||||
})
|
||||
const response = yield* execute(
|
||||
HttpClientRequest.post(url).pipe(HttpClientRequest.setHeaders(headers), HttpClientRequest.bodyFormData(form)),
|
||||
)
|
||||
return yield* parseResponse(response, request.options, http?.body)
|
||||
}
|
||||
|
||||
const references = sourceImages.map((image) => {
|
||||
if (image.type === "bytes") return { image_url: ImageInputs.dataUrl(image) }
|
||||
if (image.type === "url") return { image_url: image.url }
|
||||
if (image.type === "file-id") return { file_id: image.id }
|
||||
return undefined
|
||||
})
|
||||
if (references.some((image) => image === undefined))
|
||||
return yield* ImageInputs.invalid(ADAPTER, "OpenAI Images accepts image URLs, data URLs, bytes, and file IDs")
|
||||
const maskReference =
|
||||
mask === undefined
|
||||
? undefined
|
||||
: mask.type === "bytes"
|
||||
? { image_url: ImageInputs.dataUrl(mask) }
|
||||
: mask.type === "url"
|
||||
? { image_url: mask.url }
|
||||
: mask.type === "file-id"
|
||||
? { file_id: mask.id }
|
||||
: undefined
|
||||
if (mask !== undefined && maskReference === undefined)
|
||||
return yield* ImageInputs.invalid(ADAPTER, "OpenAI Images accepts masks as URLs, data URLs, bytes, or file IDs")
|
||||
const requestBody = mergeJsonRecords(
|
||||
{ model: request.model.id, prompt: request.prompt },
|
||||
{
|
||||
model: request.model.id,
|
||||
prompt: request.prompt,
|
||||
images: references.length === 0 ? undefined : references,
|
||||
mask: maskReference,
|
||||
},
|
||||
nativeOptions(request.options),
|
||||
http?.body,
|
||||
) as OpenAIImageBody
|
||||
const text = ProviderShared.encodeJson(requestBody)
|
||||
const url = applyQuery(`${(input.baseURL ?? DEFAULT_BASE_URL).replace(/\/$/, "")}${PATH}`, http?.query)
|
||||
const headers = yield* Auth.toEffect(input.auth)({
|
||||
request,
|
||||
method: "POST",
|
||||
|
|
@ -113,58 +198,73 @@ export const model = (input: ModelInput) => {
|
|||
HttpClientRequest.bodyText(text, "application/json"),
|
||||
),
|
||||
)
|
||||
const payload = yield* response.json.pipe(
|
||||
Effect.mapError(() => invalidOutput("Failed to read the OpenAI Images response")),
|
||||
)
|
||||
const decoded = yield* Schema.decodeUnknownEffect(OpenAIImageResponse)(payload).pipe(
|
||||
Effect.mapError(() => invalidOutput("OpenAI Images returned an invalid response")),
|
||||
)
|
||||
const format =
|
||||
decoded.output_format ?? (typeof requestBody.output_format === "string" ? requestBody.output_format : "png")
|
||||
const images = yield* Effect.forEach(decoded.data, (item, index) => {
|
||||
if (item.b64_json)
|
||||
return Effect.fromResult(Encoding.decodeBase64(item.b64_json)).pipe(
|
||||
Effect.mapError(() => invalidOutput(`OpenAI Images result ${index} contains invalid base64 data`)),
|
||||
Effect.map(
|
||||
(data) =>
|
||||
new GeneratedImage({
|
||||
mediaType: `image/${format}`,
|
||||
data,
|
||||
providerMetadata:
|
||||
item.revised_prompt === undefined ? undefined : { openai: { revisedPrompt: item.revised_prompt } },
|
||||
}),
|
||||
),
|
||||
)
|
||||
if (item.url)
|
||||
return Effect.succeed(
|
||||
new GeneratedImage({
|
||||
mediaType: `image/${format}`,
|
||||
data: item.url,
|
||||
providerMetadata:
|
||||
item.revised_prompt === undefined ? undefined : { openai: { revisedPrompt: item.revised_prompt } },
|
||||
}),
|
||||
)
|
||||
return Effect.fail(invalidOutput(`OpenAI Images result ${index} has neither image data nor a URL`))
|
||||
})
|
||||
if (images.length === 0) return yield* invalidOutput("OpenAI Images returned no images")
|
||||
return new ImageResponse({
|
||||
images,
|
||||
usage:
|
||||
decoded.usage === undefined
|
||||
? undefined
|
||||
: new Usage({
|
||||
inputTokens: decoded.usage.input_tokens,
|
||||
outputTokens: decoded.usage.output_tokens,
|
||||
totalTokens: decoded.usage.total_tokens,
|
||||
providerMetadata: { openai: decoded.usage },
|
||||
}),
|
||||
providerMetadata: { openai: { outputFormat: format } },
|
||||
})
|
||||
return yield* parseResponse(response, request.options, http?.body)
|
||||
}),
|
||||
}
|
||||
return ImageModel.make<OpenAIImageOptions>({ id: input.id, provider: "openai", route, http: input.http })
|
||||
}
|
||||
|
||||
const parseResponse = Effect.fn("OpenAIImages.parseResponse")(function* (
|
||||
response: HttpClientResponse.HttpClientResponse,
|
||||
options: OpenAIImageOptions | undefined,
|
||||
overlay: Record<string, unknown> | undefined,
|
||||
) {
|
||||
const payload = yield* response.json.pipe(
|
||||
Effect.mapError(() => invalidOutput("Failed to read the OpenAI Images response")),
|
||||
)
|
||||
const decoded = yield* Schema.decodeUnknownEffect(OpenAIImageResponse)(payload).pipe(
|
||||
Effect.mapError(() => invalidOutput("OpenAI Images returned an invalid response")),
|
||||
)
|
||||
const requestBody = mergeJsonRecords(nativeOptions(options), overlay)
|
||||
const format =
|
||||
decoded.output_format ?? (typeof requestBody?.output_format === "string" ? requestBody.output_format : "png")
|
||||
const images = yield* Effect.forEach(decoded.data, (item, index) => {
|
||||
if (item.b64_json)
|
||||
return Effect.fromResult(Encoding.decodeBase64(item.b64_json)).pipe(
|
||||
Effect.mapError(() => invalidOutput(`OpenAI Images result ${index} contains invalid base64 data`)),
|
||||
Effect.map(
|
||||
(data) =>
|
||||
new GeneratedImage({
|
||||
mediaType: `image/${format}`,
|
||||
data,
|
||||
providerMetadata:
|
||||
item.revised_prompt === undefined ? undefined : { openai: { revisedPrompt: item.revised_prompt } },
|
||||
}),
|
||||
),
|
||||
)
|
||||
if (item.url)
|
||||
return Effect.succeed(
|
||||
new GeneratedImage({
|
||||
mediaType: `image/${format}`,
|
||||
data: item.url,
|
||||
providerMetadata:
|
||||
item.revised_prompt === undefined ? undefined : { openai: { revisedPrompt: item.revised_prompt } },
|
||||
}),
|
||||
)
|
||||
return Effect.fail(invalidOutput(`OpenAI Images result ${index} has neither image data nor a URL`))
|
||||
})
|
||||
if (images.length === 0) return yield* invalidOutput("OpenAI Images returned no images")
|
||||
return new ImageResponse({
|
||||
images,
|
||||
usage:
|
||||
decoded.usage === undefined
|
||||
? undefined
|
||||
: new Usage({
|
||||
inputTokens: decoded.usage.input_tokens,
|
||||
outputTokens: decoded.usage.output_tokens,
|
||||
totalTokens: decoded.usage.total_tokens,
|
||||
providerMetadata: { openai: decoded.usage },
|
||||
}),
|
||||
providerMetadata: { openai: { outputFormat: format } },
|
||||
})
|
||||
})
|
||||
|
||||
const imageBlob = (data: Uint8Array, mediaType: string) => {
|
||||
const buffer = new ArrayBuffer(data.byteLength)
|
||||
new Uint8Array(buffer).set(data)
|
||||
return new Blob([buffer], { type: mediaType })
|
||||
}
|
||||
|
||||
export const OpenAIImages = {
|
||||
model,
|
||||
} as const
|
||||
|
|
|
|||
34
packages/ai/src/protocols/utils/image-input.ts
Normal file
34
packages/ai/src/protocols/utils/image-input.ts
Normal file
|
|
@ -0,0 +1,34 @@
|
|||
import { Effect, Encoding } from "effect"
|
||||
import type { ImageInput } from "../../image"
|
||||
import { InvalidRequestReason, LLMError } from "../../schema"
|
||||
|
||||
const invalid = (module: string, message: string) =>
|
||||
new LLMError({
|
||||
module,
|
||||
method: "generate",
|
||||
reason: new InvalidRequestReason({ message }),
|
||||
})
|
||||
|
||||
export const dataUrl = (input: Extract<ImageInput, { readonly type: "bytes" }>) =>
|
||||
`data:${input.mediaType};base64,${Encoding.encodeBase64(input.data)}`
|
||||
|
||||
export const decodeDataUrl = (
|
||||
url: string,
|
||||
module: string,
|
||||
): Effect.Effect<{ readonly mediaType: string; readonly data: Uint8Array } | undefined, LLMError> => {
|
||||
if (!url.startsWith("data:")) return Effect.succeed(undefined)
|
||||
const match = /^data:([^;,]+);base64,(.*)$/s.exec(url)
|
||||
if (!match) return Effect.fail(invalid(module, "Image data URLs must contain a MIME type and base64 data"))
|
||||
return Effect.fromResult(Encoding.decodeBase64(match[2])).pipe(
|
||||
Effect.mapError(() => invalid(module, "Image data URL contains invalid base64 data")),
|
||||
Effect.map((data) => ({ mediaType: match[1], data })),
|
||||
)
|
||||
}
|
||||
|
||||
export const invalidImageInput = invalid
|
||||
|
||||
export const ImageInputs = {
|
||||
dataUrl,
|
||||
decodeDataUrl,
|
||||
invalid: invalidImageInput,
|
||||
} as const
|
||||
|
|
@ -11,10 +11,12 @@ import {
|
|||
type HttpOptions,
|
||||
} from "../schema"
|
||||
import { ProviderShared, optionalNull } from "./shared"
|
||||
import { ImageInputs } from "./utils/image-input"
|
||||
|
||||
const ADAPTER = "xai-images"
|
||||
export const DEFAULT_BASE_URL = "https://api.x.ai/v1"
|
||||
export const PATH = "/images/generations"
|
||||
export const EDIT_PATH = "/images/edits"
|
||||
|
||||
export type XAIImageString<Known extends string> = Known | (string & {})
|
||||
|
||||
|
|
@ -111,13 +113,29 @@ export const model = (input: ModelInput) => {
|
|||
id: ADAPTER,
|
||||
generate: Effect.fn("XAIImages.generate")(function* (request: ImageRequestFor<XAIImageOptions>, execute) {
|
||||
const http = mergeHttpOptions(request.model.http, request.http)
|
||||
const imageReferences = (request.images ?? []).map((image) => {
|
||||
if (image.type === "bytes") return { url: ImageInputs.dataUrl(image), type: "image_url" as const }
|
||||
if (image.type === "url") return { url: image.url, type: "image_url" as const }
|
||||
if (image.type === "file-id") return { file_id: image.id }
|
||||
return undefined
|
||||
})
|
||||
if (imageReferences.some((image) => image === undefined))
|
||||
return yield* ImageInputs.invalid(ADAPTER, "xAI Images accepts image URLs, data URLs, bytes, and file IDs")
|
||||
const requestBody = mergeJsonRecords(
|
||||
{ model: request.model.id, prompt: request.prompt },
|
||||
{
|
||||
model: request.model.id,
|
||||
prompt: request.prompt,
|
||||
image: imageReferences.length === 1 ? imageReferences[0] : undefined,
|
||||
images: imageReferences.length > 1 ? imageReferences : undefined,
|
||||
},
|
||||
nativeOptions(request.options),
|
||||
http?.body,
|
||||
) as XAIImageBody
|
||||
const text = ProviderShared.encodeJson(requestBody)
|
||||
const url = applyQuery(`${(input.baseURL ?? DEFAULT_BASE_URL).replace(/\/$/, "")}${PATH}`, http?.query)
|
||||
const url = applyQuery(
|
||||
`${(input.baseURL ?? DEFAULT_BASE_URL).replace(/\/$/, "")}${imageReferences.length === 0 ? PATH : EDIT_PATH}`,
|
||||
http?.query,
|
||||
)
|
||||
const headers = yield* Auth.toEffect(input.auth)({
|
||||
request,
|
||||
method: "POST",
|
||||
|
|
|
|||
|
|
@ -4,6 +4,7 @@ import { GeneratedImage, ImageModel, ImageResponse, type ImageRequestFor, type I
|
|||
import { Auth, type Definition as AuthDefinition } from "../route/auth"
|
||||
import { InvalidProviderOutputReason, LLMError, mergeHttpOptions, mergeJsonRecords, type HttpOptions } from "../schema"
|
||||
import { ProviderShared } from "./shared"
|
||||
import { ImageInputs } from "./utils/image-input"
|
||||
|
||||
const ADAPTER = "zai-images"
|
||||
export const DEFAULT_BASE_URL = "https://api.z.ai/api/paas/v4"
|
||||
|
|
@ -74,6 +75,8 @@ export const model = (input: ModelInput) => {
|
|||
const route: ImageRoute<ZAIImageOptions> = {
|
||||
id: ADAPTER,
|
||||
generate: Effect.fn("ZAIImages.generate")(function* (request: ImageRequestFor<ZAIImageOptions>, execute) {
|
||||
if ((request.images?.length ?? 0) > 0)
|
||||
return yield* ImageInputs.invalid(ADAPTER, "Z.ai hosted image generation does not support image inputs")
|
||||
const http = mergeHttpOptions(request.model.http, request.http)
|
||||
const requestBody = mergeJsonRecords(
|
||||
{ model: request.model.id, prompt: request.prompt },
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue