core: expose v2 model listing API (#25821)
This commit is contained in:
Dax
@@ -0,0 +1,170 @@
|
||||
import {
|
||||
type LanguageModelV3Prompt,
|
||||
type SharedV3ProviderOptions,
|
||||
UnsupportedFunctionalityError,
|
||||
} from "@ai-sdk/provider"
|
||||
import type { OpenAICompatibleChatPrompt } from "./openai-compatible-api-types"
|
||||
import { convertToBase64 } from "@ai-sdk/provider-utils"
|
||||
|
||||
function getOpenAIMetadata(message: { providerOptions?: SharedV3ProviderOptions }) {
|
||||
return message?.providerOptions?.copilot ?? {}
|
||||
}
|
||||
|
||||
export function convertToOpenAICompatibleChatMessages(prompt: LanguageModelV3Prompt): OpenAICompatibleChatPrompt {
|
||||
const messages: OpenAICompatibleChatPrompt = []
|
||||
for (const { role, content, ...message } of prompt) {
|
||||
const metadata = getOpenAIMetadata({ ...message })
|
||||
switch (role) {
|
||||
case "system": {
|
||||
messages.push({
|
||||
role: "system",
|
||||
content: content,
|
||||
...metadata,
|
||||
})
|
||||
break
|
||||
}
|
||||
|
||||
case "user": {
|
||||
if (content.length === 1 && content[0].type === "text") {
|
||||
messages.push({
|
||||
role: "user",
|
||||
content: content[0].text,
|
||||
...getOpenAIMetadata(content[0]),
|
||||
})
|
||||
break
|
||||
}
|
||||
|
||||
messages.push({
|
||||
role: "user",
|
||||
content: content.map((part) => {
|
||||
const partMetadata = getOpenAIMetadata(part)
|
||||
switch (part.type) {
|
||||
case "text": {
|
||||
return { type: "text", text: part.text, ...partMetadata }
|
||||
}
|
||||
case "file": {
|
||||
if (part.mediaType.startsWith("image/")) {
|
||||
const mediaType = part.mediaType === "image/*" ? "image/jpeg" : part.mediaType
|
||||
|
||||
return {
|
||||
type: "image_url",
|
||||
image_url: {
|
||||
url:
|
||||
part.data instanceof URL
|
||||
? part.data.toString()
|
||||
: `data:${mediaType};base64,${convertToBase64(part.data)}`,
|
||||
},
|
||||
...partMetadata,
|
||||
}
|
||||
} else {
|
||||
throw new UnsupportedFunctionalityError({
|
||||
functionality: `file part media type ${part.mediaType}`,
|
||||
})
|
||||
}
|
||||
}
|
||||
}
|
||||
}),
|
||||
...metadata,
|
||||
})
|
||||
|
||||
break
|
||||
}
|
||||
|
||||
case "assistant": {
|
||||
let text = ""
|
||||
let reasoningText: string | undefined
|
||||
let reasoningOpaque: string | undefined
|
||||
const toolCalls: Array<{
|
||||
id: string
|
||||
type: "function"
|
||||
function: { name: string; arguments: string }
|
||||
}> = []
|
||||
|
||||
for (const part of content) {
|
||||
const partMetadata = getOpenAIMetadata(part)
|
||||
// Check for reasoningOpaque on any part (may be attached to text/tool-call)
|
||||
const partOpaque = (part.providerOptions as { copilot?: { reasoningOpaque?: string } })?.copilot
|
||||
?.reasoningOpaque
|
||||
if (partOpaque && !reasoningOpaque) {
|
||||
reasoningOpaque = partOpaque
|
||||
}
|
||||
|
||||
switch (part.type) {
|
||||
case "text": {
|
||||
text += part.text
|
||||
break
|
||||
}
|
||||
case "reasoning": {
|
||||
if (part.text) reasoningText = part.text
|
||||
break
|
||||
}
|
||||
case "tool-call": {
|
||||
toolCalls.push({
|
||||
id: part.toolCallId,
|
||||
type: "function",
|
||||
function: {
|
||||
name: part.toolName,
|
||||
arguments: JSON.stringify(part.input),
|
||||
},
|
||||
...partMetadata,
|
||||
})
|
||||
break
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
messages.push({
|
||||
role: "assistant",
|
||||
content: text || null,
|
||||
tool_calls: toolCalls.length > 0 ? toolCalls : undefined,
|
||||
reasoning_text: reasoningOpaque ? reasoningText : undefined,
|
||||
reasoning_opaque: reasoningOpaque,
|
||||
...metadata,
|
||||
})
|
||||
|
||||
break
|
||||
}
|
||||
|
||||
case "tool": {
|
||||
for (const toolResponse of content) {
|
||||
if (toolResponse.type === "tool-approval-response") {
|
||||
continue
|
||||
}
|
||||
const output = toolResponse.output
|
||||
|
||||
let contentValue: string
|
||||
switch (output.type) {
|
||||
case "text":
|
||||
case "error-text":
|
||||
contentValue = output.value
|
||||
break
|
||||
case "execution-denied":
|
||||
contentValue = output.reason ?? "Tool execution denied."
|
||||
break
|
||||
case "content":
|
||||
case "json":
|
||||
case "error-json":
|
||||
contentValue = JSON.stringify(output.value)
|
||||
break
|
||||
}
|
||||
|
||||
const toolResponseMetadata = getOpenAIMetadata(toolResponse)
|
||||
messages.push({
|
||||
role: "tool",
|
||||
tool_call_id: toolResponse.toolCallId,
|
||||
content: contentValue,
|
||||
...toolResponseMetadata,
|
||||
})
|
||||
}
|
||||
break
|
||||
}
|
||||
|
||||
default: {
|
||||
const _exhaustiveCheck: never = role
|
||||
throw new Error(`Unsupported role: ${_exhaustiveCheck}`)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return messages
|
||||
}
|
||||
@@ -0,0 +1,15 @@
|
||||
export function getResponseMetadata({
|
||||
id,
|
||||
model,
|
||||
created,
|
||||
}: {
|
||||
id?: string | undefined | null
|
||||
created?: number | undefined | null
|
||||
model?: string | undefined | null
|
||||
}) {
|
||||
return {
|
||||
id: id ?? undefined,
|
||||
modelId: model ?? undefined,
|
||||
timestamp: created != null ? new Date(created * 1000) : undefined,
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,19 @@
|
||||
import type { LanguageModelV3FinishReason } from "@ai-sdk/provider"
|
||||
|
||||
export function mapOpenAICompatibleFinishReason(
|
||||
finishReason: string | null | undefined,
|
||||
): LanguageModelV3FinishReason["unified"] {
|
||||
switch (finishReason) {
|
||||
case "stop":
|
||||
return "stop"
|
||||
case "length":
|
||||
return "length"
|
||||
case "content_filter":
|
||||
return "content-filter"
|
||||
case "function_call":
|
||||
case "tool_calls":
|
||||
return "tool-calls"
|
||||
default:
|
||||
return "other"
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,64 @@
|
||||
import type { JSONValue } from "@ai-sdk/provider"
|
||||
|
||||
export type OpenAICompatibleChatPrompt = Array<OpenAICompatibleMessage>
|
||||
|
||||
export type OpenAICompatibleMessage =
|
||||
| OpenAICompatibleSystemMessage
|
||||
| OpenAICompatibleUserMessage
|
||||
| OpenAICompatibleAssistantMessage
|
||||
| OpenAICompatibleToolMessage
|
||||
|
||||
// Allow for arbitrary additional properties for general purpose
|
||||
// provider-metadata-specific extensibility.
|
||||
type JsonRecord<T = never> = Record<string, JSONValue | JSONValue[] | T | T[] | undefined>
|
||||
|
||||
export interface OpenAICompatibleSystemMessage extends JsonRecord<OpenAICompatibleSystemContentPart> {
|
||||
role: "system"
|
||||
content: string | Array<OpenAICompatibleSystemContentPart>
|
||||
}
|
||||
|
||||
export interface OpenAICompatibleSystemContentPart extends JsonRecord {
|
||||
type: "text"
|
||||
text: string
|
||||
}
|
||||
|
||||
export interface OpenAICompatibleUserMessage extends JsonRecord<OpenAICompatibleContentPart> {
|
||||
role: "user"
|
||||
content: string | Array<OpenAICompatibleContentPart>
|
||||
}
|
||||
|
||||
export type OpenAICompatibleContentPart = OpenAICompatibleContentPartText | OpenAICompatibleContentPartImage
|
||||
|
||||
export interface OpenAICompatibleContentPartImage extends JsonRecord {
|
||||
type: "image_url"
|
||||
image_url: { url: string }
|
||||
}
|
||||
|
||||
export interface OpenAICompatibleContentPartText extends JsonRecord {
|
||||
type: "text"
|
||||
text: string
|
||||
}
|
||||
|
||||
export interface OpenAICompatibleAssistantMessage extends JsonRecord<OpenAICompatibleMessageToolCall> {
|
||||
role: "assistant"
|
||||
content?: string | null
|
||||
tool_calls?: Array<OpenAICompatibleMessageToolCall>
|
||||
// Copilot-specific reasoning fields
|
||||
reasoning_text?: string
|
||||
reasoning_opaque?: string
|
||||
}
|
||||
|
||||
export interface OpenAICompatibleMessageToolCall extends JsonRecord {
|
||||
type: "function"
|
||||
id: string
|
||||
function: {
|
||||
arguments: string
|
||||
name: string
|
||||
}
|
||||
}
|
||||
|
||||
export interface OpenAICompatibleToolMessage extends JsonRecord {
|
||||
role: "tool"
|
||||
content: string
|
||||
tool_call_id: string
|
||||
}
|
||||
@@ -0,0 +1,815 @@
|
||||
import {
|
||||
APICallError,
|
||||
InvalidResponseDataError,
|
||||
type LanguageModelV3,
|
||||
type LanguageModelV3CallOptions,
|
||||
type LanguageModelV3Content,
|
||||
type LanguageModelV3StreamPart,
|
||||
type SharedV3ProviderMetadata,
|
||||
type SharedV3Warning,
|
||||
} from "@ai-sdk/provider"
|
||||
import {
|
||||
combineHeaders,
|
||||
createEventSourceResponseHandler,
|
||||
createJsonErrorResponseHandler,
|
||||
createJsonResponseHandler,
|
||||
type FetchFunction,
|
||||
generateId,
|
||||
isParsableJson,
|
||||
parseProviderOptions,
|
||||
type ParseResult,
|
||||
postJsonToApi,
|
||||
type ResponseHandler,
|
||||
} from "@ai-sdk/provider-utils"
|
||||
import { z } from "zod/v4"
|
||||
import { convertToOpenAICompatibleChatMessages } from "./convert-to-openai-compatible-chat-messages"
|
||||
import { getResponseMetadata } from "./get-response-metadata"
|
||||
import { mapOpenAICompatibleFinishReason } from "./map-openai-compatible-finish-reason"
|
||||
import { type OpenAICompatibleChatModelId, openaiCompatibleProviderOptions } from "./openai-compatible-chat-options"
|
||||
import { defaultOpenAICompatibleErrorStructure, type ProviderErrorStructure } from "../openai-compatible-error"
|
||||
import type { MetadataExtractor } from "./openai-compatible-metadata-extractor"
|
||||
import { prepareTools } from "./openai-compatible-prepare-tools"
|
||||
|
||||
export type OpenAICompatibleChatConfig = {
|
||||
provider: string
|
||||
headers: () => Record<string, string | undefined>
|
||||
url: (options: { modelId: string; path: string }) => string
|
||||
fetch?: FetchFunction
|
||||
includeUsage?: boolean
|
||||
errorStructure?: ProviderErrorStructure<any>
|
||||
metadataExtractor?: MetadataExtractor
|
||||
|
||||
/**
|
||||
* Whether the model supports structured outputs.
|
||||
*/
|
||||
supportsStructuredOutputs?: boolean
|
||||
|
||||
/**
|
||||
* The supported URLs for the model.
|
||||
*/
|
||||
supportedUrls?: () => LanguageModelV3["supportedUrls"]
|
||||
}
|
||||
|
||||
export class OpenAICompatibleChatLanguageModel implements LanguageModelV3 {
|
||||
readonly specificationVersion = "v3"
|
||||
|
||||
readonly supportsStructuredOutputs: boolean
|
||||
|
||||
readonly modelId: OpenAICompatibleChatModelId
|
||||
private readonly config: OpenAICompatibleChatConfig
|
||||
private readonly failedResponseHandler: ResponseHandler<APICallError>
|
||||
private readonly chunkSchema // type inferred via constructor
|
||||
|
||||
constructor(modelId: OpenAICompatibleChatModelId, config: OpenAICompatibleChatConfig) {
|
||||
this.modelId = modelId
|
||||
this.config = config
|
||||
|
||||
// initialize error handling:
|
||||
const errorStructure = config.errorStructure ?? defaultOpenAICompatibleErrorStructure
|
||||
this.chunkSchema = createOpenAICompatibleChatChunkSchema(errorStructure.errorSchema)
|
||||
this.failedResponseHandler = createJsonErrorResponseHandler(errorStructure)
|
||||
|
||||
this.supportsStructuredOutputs = config.supportsStructuredOutputs ?? false
|
||||
}
|
||||
|
||||
get provider(): string {
|
||||
return this.config.provider
|
||||
}
|
||||
|
||||
private get providerOptionsName(): string {
|
||||
return this.config.provider.split(".")[0].trim()
|
||||
}
|
||||
|
||||
get supportedUrls() {
|
||||
return this.config.supportedUrls?.() ?? {}
|
||||
}
|
||||
|
||||
private async getArgs({
|
||||
prompt,
|
||||
maxOutputTokens,
|
||||
temperature,
|
||||
topP,
|
||||
topK,
|
||||
frequencyPenalty,
|
||||
presencePenalty,
|
||||
providerOptions,
|
||||
stopSequences,
|
||||
responseFormat,
|
||||
seed,
|
||||
toolChoice,
|
||||
tools,
|
||||
}: LanguageModelV3CallOptions) {
|
||||
const warnings: SharedV3Warning[] = []
|
||||
|
||||
// Parse provider options
|
||||
const compatibleOptions = Object.assign(
|
||||
(await parseProviderOptions({
|
||||
provider: "copilot",
|
||||
providerOptions,
|
||||
schema: openaiCompatibleProviderOptions,
|
||||
})) ?? {},
|
||||
(await parseProviderOptions({
|
||||
provider: this.providerOptionsName,
|
||||
providerOptions,
|
||||
schema: openaiCompatibleProviderOptions,
|
||||
})) ?? {},
|
||||
)
|
||||
|
||||
if (topK != null) {
|
||||
warnings.push({ type: "unsupported", feature: "topK" })
|
||||
}
|
||||
|
||||
if (responseFormat?.type === "json" && responseFormat.schema != null && !this.supportsStructuredOutputs) {
|
||||
warnings.push({
|
||||
type: "unsupported",
|
||||
feature: "responseFormat",
|
||||
details: "JSON response format schema is only supported with structuredOutputs",
|
||||
})
|
||||
}
|
||||
|
||||
const {
|
||||
tools: openaiTools,
|
||||
toolChoice: openaiToolChoice,
|
||||
toolWarnings,
|
||||
} = prepareTools({
|
||||
tools,
|
||||
toolChoice,
|
||||
})
|
||||
|
||||
return {
|
||||
args: {
|
||||
// model id:
|
||||
model: this.modelId,
|
||||
|
||||
// model specific settings:
|
||||
user: compatibleOptions.user,
|
||||
|
||||
// standardized settings:
|
||||
max_tokens: maxOutputTokens,
|
||||
temperature,
|
||||
top_p: topP,
|
||||
frequency_penalty: frequencyPenalty,
|
||||
presence_penalty: presencePenalty,
|
||||
response_format:
|
||||
responseFormat?.type === "json"
|
||||
? this.supportsStructuredOutputs === true && responseFormat.schema != null
|
||||
? {
|
||||
type: "json_schema",
|
||||
json_schema: {
|
||||
schema: responseFormat.schema,
|
||||
name: responseFormat.name ?? "response",
|
||||
description: responseFormat.description,
|
||||
},
|
||||
}
|
||||
: { type: "json_object" }
|
||||
: undefined,
|
||||
|
||||
stop: stopSequences,
|
||||
seed,
|
||||
...Object.fromEntries(
|
||||
Object.entries(providerOptions?.[this.providerOptionsName] ?? {}).filter(
|
||||
([key]) => !Object.keys(openaiCompatibleProviderOptions.shape).includes(key),
|
||||
),
|
||||
),
|
||||
|
||||
reasoning_effort: compatibleOptions.reasoningEffort,
|
||||
verbosity: compatibleOptions.textVerbosity,
|
||||
|
||||
// messages:
|
||||
messages: convertToOpenAICompatibleChatMessages(prompt),
|
||||
|
||||
// tools:
|
||||
tools: openaiTools,
|
||||
tool_choice: openaiToolChoice,
|
||||
|
||||
// thinking_budget
|
||||
thinking_budget: compatibleOptions.thinking_budget,
|
||||
},
|
||||
warnings: [...warnings, ...toolWarnings],
|
||||
}
|
||||
}
|
||||
|
||||
async doGenerate(options: LanguageModelV3CallOptions) {
|
||||
const { args, warnings } = await this.getArgs({ ...options })
|
||||
|
||||
const body = JSON.stringify(args)
|
||||
|
||||
const {
|
||||
responseHeaders,
|
||||
value: responseBody,
|
||||
rawValue: rawResponse,
|
||||
} = await postJsonToApi({
|
||||
url: this.config.url({
|
||||
path: "/chat/completions",
|
||||
modelId: this.modelId,
|
||||
}),
|
||||
headers: combineHeaders(this.config.headers(), options.headers),
|
||||
body: args,
|
||||
failedResponseHandler: this.failedResponseHandler,
|
||||
successfulResponseHandler: createJsonResponseHandler(OpenAICompatibleChatResponseSchema),
|
||||
abortSignal: options.abortSignal,
|
||||
fetch: this.config.fetch,
|
||||
})
|
||||
|
||||
const choice = responseBody.choices[0]
|
||||
const content: Array<LanguageModelV3Content> = []
|
||||
|
||||
// text content:
|
||||
const text = choice.message.content
|
||||
if (text != null && text.length > 0) {
|
||||
content.push({
|
||||
type: "text",
|
||||
text,
|
||||
providerMetadata: choice.message.reasoning_opaque
|
||||
? { copilot: { reasoningOpaque: choice.message.reasoning_opaque } }
|
||||
: undefined,
|
||||
})
|
||||
}
|
||||
|
||||
// reasoning content (Copilot uses reasoning_text):
|
||||
const reasoning = choice.message.reasoning_text
|
||||
if (reasoning != null && reasoning.length > 0) {
|
||||
content.push({
|
||||
type: "reasoning",
|
||||
text: reasoning,
|
||||
// Include reasoning_opaque for Copilot multi-turn reasoning
|
||||
providerMetadata: choice.message.reasoning_opaque
|
||||
? { copilot: { reasoningOpaque: choice.message.reasoning_opaque } }
|
||||
: undefined,
|
||||
})
|
||||
}
|
||||
|
||||
// tool calls:
|
||||
if (choice.message.tool_calls != null) {
|
||||
for (const toolCall of choice.message.tool_calls) {
|
||||
content.push({
|
||||
type: "tool-call",
|
||||
toolCallId: toolCall.id ?? generateId(),
|
||||
toolName: toolCall.function.name,
|
||||
input: toolCall.function.arguments!,
|
||||
providerMetadata: choice.message.reasoning_opaque
|
||||
? { copilot: { reasoningOpaque: choice.message.reasoning_opaque } }
|
||||
: undefined,
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
// provider metadata:
|
||||
const providerMetadata: SharedV3ProviderMetadata = {
|
||||
[this.providerOptionsName]: {},
|
||||
...(await this.config.metadataExtractor?.extractMetadata?.({
|
||||
parsedBody: rawResponse,
|
||||
})),
|
||||
}
|
||||
const completionTokenDetails = responseBody.usage?.completion_tokens_details
|
||||
if (completionTokenDetails?.accepted_prediction_tokens != null) {
|
||||
providerMetadata[this.providerOptionsName].acceptedPredictionTokens =
|
||||
completionTokenDetails?.accepted_prediction_tokens
|
||||
}
|
||||
if (completionTokenDetails?.rejected_prediction_tokens != null) {
|
||||
providerMetadata[this.providerOptionsName].rejectedPredictionTokens =
|
||||
completionTokenDetails?.rejected_prediction_tokens
|
||||
}
|
||||
|
||||
return {
|
||||
content,
|
||||
finishReason: {
|
||||
unified: mapOpenAICompatibleFinishReason(choice.finish_reason),
|
||||
raw: choice.finish_reason ?? undefined,
|
||||
},
|
||||
usage: {
|
||||
inputTokens: {
|
||||
total: responseBody.usage?.prompt_tokens ?? undefined,
|
||||
noCache: undefined,
|
||||
cacheRead: responseBody.usage?.prompt_tokens_details?.cached_tokens ?? undefined,
|
||||
cacheWrite: undefined,
|
||||
},
|
||||
outputTokens: {
|
||||
total: responseBody.usage?.completion_tokens ?? undefined,
|
||||
text: undefined,
|
||||
reasoning: responseBody.usage?.completion_tokens_details?.reasoning_tokens ?? undefined,
|
||||
},
|
||||
raw: responseBody.usage ?? undefined,
|
||||
},
|
||||
providerMetadata,
|
||||
request: { body },
|
||||
response: {
|
||||
...getResponseMetadata(responseBody),
|
||||
headers: responseHeaders,
|
||||
body: rawResponse,
|
||||
},
|
||||
warnings,
|
||||
}
|
||||
}
|
||||
|
||||
async doStream(options: LanguageModelV3CallOptions) {
|
||||
const { args, warnings } = await this.getArgs({ ...options })
|
||||
|
||||
const body = {
|
||||
...args,
|
||||
stream: true,
|
||||
|
||||
// only include stream_options when in strict compatibility mode:
|
||||
stream_options: this.config.includeUsage ? { include_usage: true } : undefined,
|
||||
}
|
||||
|
||||
const metadataExtractor = this.config.metadataExtractor?.createStreamExtractor()
|
||||
|
||||
const { responseHeaders, value: response } = await postJsonToApi({
|
||||
url: this.config.url({
|
||||
path: "/chat/completions",
|
||||
modelId: this.modelId,
|
||||
}),
|
||||
headers: combineHeaders(this.config.headers(), options.headers),
|
||||
body,
|
||||
failedResponseHandler: this.failedResponseHandler,
|
||||
successfulResponseHandler: createEventSourceResponseHandler(this.chunkSchema),
|
||||
abortSignal: options.abortSignal,
|
||||
fetch: this.config.fetch,
|
||||
})
|
||||
|
||||
const toolCalls: Array<{
|
||||
id: string
|
||||
type: "function"
|
||||
function: {
|
||||
name: string
|
||||
arguments: string
|
||||
}
|
||||
hasFinished: boolean
|
||||
}> = []
|
||||
|
||||
let finishReason: {
|
||||
unified: ReturnType<typeof mapOpenAICompatibleFinishReason>
|
||||
raw: string | undefined
|
||||
} = {
|
||||
unified: "other",
|
||||
raw: undefined,
|
||||
}
|
||||
const usage: {
|
||||
completionTokens: number | undefined
|
||||
completionTokensDetails: {
|
||||
reasoningTokens: number | undefined
|
||||
acceptedPredictionTokens: number | undefined
|
||||
rejectedPredictionTokens: number | undefined
|
||||
}
|
||||
promptTokens: number | undefined
|
||||
promptTokensDetails: {
|
||||
cachedTokens: number | undefined
|
||||
}
|
||||
totalTokens: number | undefined
|
||||
} = {
|
||||
completionTokens: undefined,
|
||||
completionTokensDetails: {
|
||||
reasoningTokens: undefined,
|
||||
acceptedPredictionTokens: undefined,
|
||||
rejectedPredictionTokens: undefined,
|
||||
},
|
||||
promptTokens: undefined,
|
||||
promptTokensDetails: {
|
||||
cachedTokens: undefined,
|
||||
},
|
||||
totalTokens: undefined,
|
||||
}
|
||||
let isFirstChunk = true
|
||||
const providerOptionsName = this.providerOptionsName
|
||||
let isActiveReasoning = false
|
||||
let isActiveText = false
|
||||
let reasoningOpaque: string | undefined
|
||||
|
||||
return {
|
||||
stream: response.pipeThrough(
|
||||
new TransformStream<ParseResult<z.infer<typeof this.chunkSchema>>, LanguageModelV3StreamPart>({
|
||||
start(controller) {
|
||||
controller.enqueue({ type: "stream-start", warnings })
|
||||
},
|
||||
|
||||
// TODO we lost type safety on Chunk, most likely due to the error schema. MUST FIX
|
||||
transform(chunk, controller) {
|
||||
// Emit raw chunk if requested (before anything else)
|
||||
if (options.includeRawChunks) {
|
||||
controller.enqueue({ type: "raw", rawValue: chunk.rawValue })
|
||||
}
|
||||
|
||||
// handle failed chunk parsing / validation:
|
||||
if (!chunk.success) {
|
||||
finishReason = {
|
||||
unified: "error",
|
||||
raw: undefined,
|
||||
}
|
||||
controller.enqueue({ type: "error", error: chunk.error })
|
||||
return
|
||||
}
|
||||
const value = chunk.value
|
||||
|
||||
metadataExtractor?.processChunk(chunk.rawValue)
|
||||
|
||||
// handle error chunks:
|
||||
if ("error" in value) {
|
||||
finishReason = {
|
||||
unified: "error",
|
||||
raw: undefined,
|
||||
}
|
||||
controller.enqueue({ type: "error", error: value.error.message })
|
||||
return
|
||||
}
|
||||
|
||||
if (isFirstChunk) {
|
||||
isFirstChunk = false
|
||||
|
||||
controller.enqueue({
|
||||
type: "response-metadata",
|
||||
...getResponseMetadata(value),
|
||||
})
|
||||
}
|
||||
|
||||
if (value.usage != null) {
|
||||
const {
|
||||
prompt_tokens,
|
||||
completion_tokens,
|
||||
total_tokens,
|
||||
prompt_tokens_details,
|
||||
completion_tokens_details,
|
||||
} = value.usage
|
||||
|
||||
usage.promptTokens = prompt_tokens ?? undefined
|
||||
usage.completionTokens = completion_tokens ?? undefined
|
||||
usage.totalTokens = total_tokens ?? undefined
|
||||
if (completion_tokens_details?.reasoning_tokens != null) {
|
||||
usage.completionTokensDetails.reasoningTokens = completion_tokens_details?.reasoning_tokens
|
||||
}
|
||||
if (completion_tokens_details?.accepted_prediction_tokens != null) {
|
||||
usage.completionTokensDetails.acceptedPredictionTokens =
|
||||
completion_tokens_details?.accepted_prediction_tokens
|
||||
}
|
||||
if (completion_tokens_details?.rejected_prediction_tokens != null) {
|
||||
usage.completionTokensDetails.rejectedPredictionTokens =
|
||||
completion_tokens_details?.rejected_prediction_tokens
|
||||
}
|
||||
if (prompt_tokens_details?.cached_tokens != null) {
|
||||
usage.promptTokensDetails.cachedTokens = prompt_tokens_details?.cached_tokens
|
||||
}
|
||||
}
|
||||
|
||||
const choice = value.choices[0]
|
||||
|
||||
if (choice?.finish_reason != null) {
|
||||
finishReason = {
|
||||
unified: mapOpenAICompatibleFinishReason(choice.finish_reason),
|
||||
raw: choice.finish_reason ?? undefined,
|
||||
}
|
||||
}
|
||||
|
||||
if (choice?.delta == null) {
|
||||
return
|
||||
}
|
||||
|
||||
const delta = choice.delta
|
||||
|
||||
// Capture reasoning_opaque for Copilot multi-turn reasoning
|
||||
if (delta.reasoning_opaque) {
|
||||
if (reasoningOpaque != null) {
|
||||
throw new InvalidResponseDataError({
|
||||
data: delta,
|
||||
message:
|
||||
"Multiple reasoning_opaque values received in a single response. Only one thinking part per response is supported.",
|
||||
})
|
||||
}
|
||||
reasoningOpaque = delta.reasoning_opaque
|
||||
}
|
||||
|
||||
// enqueue reasoning before text deltas (Copilot uses reasoning_text):
|
||||
const reasoningContent = delta.reasoning_text
|
||||
if (reasoningContent) {
|
||||
if (!isActiveReasoning) {
|
||||
controller.enqueue({
|
||||
type: "reasoning-start",
|
||||
id: "reasoning-0",
|
||||
})
|
||||
isActiveReasoning = true
|
||||
}
|
||||
|
||||
controller.enqueue({
|
||||
type: "reasoning-delta",
|
||||
id: "reasoning-0",
|
||||
delta: reasoningContent,
|
||||
})
|
||||
}
|
||||
|
||||
if (delta.content) {
|
||||
// If reasoning was active and we're starting text, end reasoning first
|
||||
// This handles the case where reasoning_opaque and content come in the same chunk
|
||||
if (isActiveReasoning && !isActiveText) {
|
||||
controller.enqueue({
|
||||
type: "reasoning-end",
|
||||
id: "reasoning-0",
|
||||
providerMetadata: reasoningOpaque ? { copilot: { reasoningOpaque } } : undefined,
|
||||
})
|
||||
isActiveReasoning = false
|
||||
}
|
||||
|
||||
if (!isActiveText) {
|
||||
controller.enqueue({
|
||||
type: "text-start",
|
||||
id: "txt-0",
|
||||
providerMetadata: reasoningOpaque ? { copilot: { reasoningOpaque } } : undefined,
|
||||
})
|
||||
isActiveText = true
|
||||
}
|
||||
|
||||
controller.enqueue({
|
||||
type: "text-delta",
|
||||
id: "txt-0",
|
||||
delta: delta.content,
|
||||
})
|
||||
}
|
||||
|
||||
if (delta.tool_calls != null) {
|
||||
// If reasoning was active and we're starting tool calls, end reasoning first
|
||||
// This handles the case where reasoning goes directly to tool calls with no content
|
||||
if (isActiveReasoning) {
|
||||
controller.enqueue({
|
||||
type: "reasoning-end",
|
||||
id: "reasoning-0",
|
||||
providerMetadata: reasoningOpaque ? { copilot: { reasoningOpaque } } : undefined,
|
||||
})
|
||||
isActiveReasoning = false
|
||||
}
|
||||
for (const toolCallDelta of delta.tool_calls) {
|
||||
const index = toolCallDelta.index
|
||||
|
||||
if (toolCalls[index] == null) {
|
||||
if (toolCallDelta.id == null) {
|
||||
throw new InvalidResponseDataError({
|
||||
data: toolCallDelta,
|
||||
message: `Expected 'id' to be a string.`,
|
||||
})
|
||||
}
|
||||
|
||||
if (toolCallDelta.function?.name == null) {
|
||||
throw new InvalidResponseDataError({
|
||||
data: toolCallDelta,
|
||||
message: `Expected 'function.name' to be a string.`,
|
||||
})
|
||||
}
|
||||
|
||||
controller.enqueue({
|
||||
type: "tool-input-start",
|
||||
id: toolCallDelta.id,
|
||||
toolName: toolCallDelta.function.name,
|
||||
})
|
||||
|
||||
toolCalls[index] = {
|
||||
id: toolCallDelta.id,
|
||||
type: "function",
|
||||
function: {
|
||||
name: toolCallDelta.function.name,
|
||||
arguments: toolCallDelta.function.arguments ?? "",
|
||||
},
|
||||
hasFinished: false,
|
||||
}
|
||||
|
||||
const toolCall = toolCalls[index]
|
||||
|
||||
if (toolCall.function?.name != null && toolCall.function?.arguments != null) {
|
||||
// send delta if the argument text has already started:
|
||||
if (toolCall.function.arguments.length > 0) {
|
||||
controller.enqueue({
|
||||
type: "tool-input-delta",
|
||||
id: toolCall.id,
|
||||
delta: toolCall.function.arguments,
|
||||
})
|
||||
}
|
||||
|
||||
// check if tool call is complete
|
||||
// (some providers send the full tool call in one chunk):
|
||||
if (isParsableJson(toolCall.function.arguments)) {
|
||||
controller.enqueue({
|
||||
type: "tool-input-end",
|
||||
id: toolCall.id,
|
||||
})
|
||||
|
||||
controller.enqueue({
|
||||
type: "tool-call",
|
||||
toolCallId: toolCall.id ?? generateId(),
|
||||
toolName: toolCall.function.name,
|
||||
input: toolCall.function.arguments,
|
||||
providerMetadata: reasoningOpaque ? { copilot: { reasoningOpaque } } : undefined,
|
||||
})
|
||||
toolCall.hasFinished = true
|
||||
}
|
||||
}
|
||||
|
||||
continue
|
||||
}
|
||||
|
||||
// existing tool call, merge if not finished
|
||||
const toolCall = toolCalls[index]
|
||||
|
||||
if (toolCall.hasFinished) {
|
||||
continue
|
||||
}
|
||||
|
||||
if (toolCallDelta.function?.arguments != null) {
|
||||
toolCall.function!.arguments += toolCallDelta.function?.arguments ?? ""
|
||||
}
|
||||
|
||||
// send delta
|
||||
controller.enqueue({
|
||||
type: "tool-input-delta",
|
||||
id: toolCall.id,
|
||||
delta: toolCallDelta.function.arguments ?? "",
|
||||
})
|
||||
|
||||
// check if tool call is complete
|
||||
if (
|
||||
toolCall.function?.name != null &&
|
||||
toolCall.function?.arguments != null &&
|
||||
isParsableJson(toolCall.function.arguments)
|
||||
) {
|
||||
controller.enqueue({
|
||||
type: "tool-input-end",
|
||||
id: toolCall.id,
|
||||
})
|
||||
|
||||
controller.enqueue({
|
||||
type: "tool-call",
|
||||
toolCallId: toolCall.id ?? generateId(),
|
||||
toolName: toolCall.function.name,
|
||||
input: toolCall.function.arguments,
|
||||
providerMetadata: reasoningOpaque ? { copilot: { reasoningOpaque } } : undefined,
|
||||
})
|
||||
toolCall.hasFinished = true
|
||||
}
|
||||
}
|
||||
}
|
||||
},
|
||||
|
||||
flush(controller) {
|
||||
if (isActiveReasoning) {
|
||||
controller.enqueue({
|
||||
type: "reasoning-end",
|
||||
id: "reasoning-0",
|
||||
// Include reasoning_opaque for Copilot multi-turn reasoning
|
||||
providerMetadata: reasoningOpaque ? { copilot: { reasoningOpaque } } : undefined,
|
||||
})
|
||||
}
|
||||
|
||||
if (isActiveText) {
|
||||
controller.enqueue({ type: "text-end", id: "txt-0" })
|
||||
}
|
||||
|
||||
// go through all tool calls and send the ones that are not finished
|
||||
for (const toolCall of toolCalls.filter((toolCall) => !toolCall.hasFinished)) {
|
||||
controller.enqueue({
|
||||
type: "tool-input-end",
|
||||
id: toolCall.id,
|
||||
})
|
||||
|
||||
controller.enqueue({
|
||||
type: "tool-call",
|
||||
toolCallId: toolCall.id ?? generateId(),
|
||||
toolName: toolCall.function.name,
|
||||
input: toolCall.function.arguments,
|
||||
})
|
||||
}
|
||||
|
||||
const providerMetadata: SharedV3ProviderMetadata = {
|
||||
[providerOptionsName]: {},
|
||||
// Include reasoning_opaque for Copilot multi-turn reasoning
|
||||
...(reasoningOpaque ? { copilot: { reasoningOpaque } } : {}),
|
||||
...metadataExtractor?.buildMetadata(),
|
||||
}
|
||||
if (usage.completionTokensDetails.acceptedPredictionTokens != null) {
|
||||
providerMetadata[providerOptionsName].acceptedPredictionTokens =
|
||||
usage.completionTokensDetails.acceptedPredictionTokens
|
||||
}
|
||||
if (usage.completionTokensDetails.rejectedPredictionTokens != null) {
|
||||
providerMetadata[providerOptionsName].rejectedPredictionTokens =
|
||||
usage.completionTokensDetails.rejectedPredictionTokens
|
||||
}
|
||||
|
||||
controller.enqueue({
|
||||
type: "finish",
|
||||
finishReason,
|
||||
usage: {
|
||||
inputTokens: {
|
||||
total: usage.promptTokens,
|
||||
noCache:
|
||||
usage.promptTokens != undefined && usage.promptTokensDetails.cachedTokens != undefined
|
||||
? usage.promptTokens - usage.promptTokensDetails.cachedTokens
|
||||
: undefined,
|
||||
cacheRead: usage.promptTokensDetails.cachedTokens,
|
||||
cacheWrite: undefined,
|
||||
},
|
||||
outputTokens: {
|
||||
total: usage.completionTokens,
|
||||
text: undefined,
|
||||
reasoning: usage.completionTokensDetails.reasoningTokens,
|
||||
},
|
||||
raw: {
|
||||
prompt_tokens: usage.promptTokens ?? null,
|
||||
completion_tokens: usage.completionTokens ?? null,
|
||||
total_tokens: usage.totalTokens ?? null,
|
||||
},
|
||||
},
|
||||
providerMetadata,
|
||||
})
|
||||
},
|
||||
}),
|
||||
),
|
||||
request: { body },
|
||||
response: { headers: responseHeaders },
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
const openaiCompatibleTokenUsageSchema = z
|
||||
.object({
|
||||
prompt_tokens: z.number().nullish(),
|
||||
completion_tokens: z.number().nullish(),
|
||||
total_tokens: z.number().nullish(),
|
||||
prompt_tokens_details: z
|
||||
.object({
|
||||
cached_tokens: z.number().nullish(),
|
||||
})
|
||||
.nullish(),
|
||||
completion_tokens_details: z
|
||||
.object({
|
||||
reasoning_tokens: z.number().nullish(),
|
||||
accepted_prediction_tokens: z.number().nullish(),
|
||||
rejected_prediction_tokens: z.number().nullish(),
|
||||
})
|
||||
.nullish(),
|
||||
})
|
||||
.nullish()
|
||||
|
||||
// limited version of the schema, focussed on what is needed for the implementation
|
||||
// this approach limits breakages when the API changes and increases efficiency
|
||||
const OpenAICompatibleChatResponseSchema = z.object({
|
||||
id: z.string().nullish(),
|
||||
created: z.number().nullish(),
|
||||
model: z.string().nullish(),
|
||||
choices: z.array(
|
||||
z.object({
|
||||
message: z.object({
|
||||
role: z.literal("assistant").nullish(),
|
||||
content: z.string().nullish(),
|
||||
// Copilot-specific reasoning fields
|
||||
reasoning_text: z.string().nullish(),
|
||||
reasoning_opaque: z.string().nullish(),
|
||||
tool_calls: z
|
||||
.array(
|
||||
z.object({
|
||||
id: z.string().nullish(),
|
||||
function: z.object({
|
||||
name: z.string(),
|
||||
arguments: z.string(),
|
||||
}),
|
||||
}),
|
||||
)
|
||||
.nullish(),
|
||||
}),
|
||||
finish_reason: z.string().nullish(),
|
||||
}),
|
||||
),
|
||||
usage: openaiCompatibleTokenUsageSchema,
|
||||
})
|
||||
|
||||
// limited version of the schema, focussed on what is needed for the implementation
|
||||
// this approach limits breakages when the API changes and increases efficiency
|
||||
const createOpenAICompatibleChatChunkSchema = <ERROR_SCHEMA extends z.core.$ZodType>(errorSchema: ERROR_SCHEMA) =>
|
||||
z.union([
|
||||
z.object({
|
||||
id: z.string().nullish(),
|
||||
created: z.number().nullish(),
|
||||
model: z.string().nullish(),
|
||||
choices: z.array(
|
||||
z.object({
|
||||
delta: z
|
||||
.object({
|
||||
role: z.enum(["assistant"]).nullish(),
|
||||
content: z.string().nullish(),
|
||||
// Copilot-specific reasoning fields
|
||||
reasoning_text: z.string().nullish(),
|
||||
reasoning_opaque: z.string().nullish(),
|
||||
tool_calls: z
|
||||
.array(
|
||||
z.object({
|
||||
index: z.number(),
|
||||
id: z.string().nullish(),
|
||||
function: z.object({
|
||||
name: z.string().nullish(),
|
||||
arguments: z.string().nullish(),
|
||||
}),
|
||||
}),
|
||||
)
|
||||
.nullish(),
|
||||
})
|
||||
.nullish(),
|
||||
finish_reason: z.string().nullish(),
|
||||
}),
|
||||
),
|
||||
usage: openaiCompatibleTokenUsageSchema,
|
||||
}),
|
||||
errorSchema,
|
||||
])
|
||||
@@ -0,0 +1,28 @@
|
||||
import { z } from "zod/v4"
|
||||
|
||||
export type OpenAICompatibleChatModelId = string
|
||||
|
||||
export const openaiCompatibleProviderOptions = z.object({
|
||||
/**
|
||||
* A unique identifier representing your end-user, which can help the provider to
|
||||
* monitor and detect abuse.
|
||||
*/
|
||||
user: z.string().optional(),
|
||||
|
||||
/**
|
||||
* Reasoning effort for reasoning models. Defaults to `medium`.
|
||||
*/
|
||||
reasoningEffort: z.string().optional(),
|
||||
|
||||
/**
|
||||
* Controls the verbosity of the generated text. Defaults to `medium`.
|
||||
*/
|
||||
textVerbosity: z.string().optional(),
|
||||
|
||||
/**
|
||||
* Copilot thinking_budget used for Anthropic models.
|
||||
*/
|
||||
thinking_budget: z.number().optional(),
|
||||
})
|
||||
|
||||
export type OpenAICompatibleProviderOptions = z.infer<typeof openaiCompatibleProviderOptions>
|
||||
@@ -0,0 +1,44 @@
|
||||
import type { SharedV3ProviderMetadata } from "@ai-sdk/provider"
|
||||
|
||||
/**
|
||||
Extracts provider-specific metadata from API responses.
|
||||
Used to standardize metadata handling across different LLM providers while allowing
|
||||
provider-specific metadata to be captured.
|
||||
*/
|
||||
export type MetadataExtractor = {
|
||||
/**
|
||||
* Extracts provider metadata from a complete, non-streaming response.
|
||||
*
|
||||
* @param parsedBody - The parsed response JSON body from the provider's API.
|
||||
*
|
||||
* @returns Provider-specific metadata or undefined if no metadata is available.
|
||||
* The metadata should be under a key indicating the provider id.
|
||||
*/
|
||||
extractMetadata: ({ parsedBody }: { parsedBody: unknown }) => Promise<SharedV3ProviderMetadata | undefined>
|
||||
|
||||
/**
|
||||
* Creates an extractor for handling streaming responses. The returned object provides
|
||||
* methods to process individual chunks and build the final metadata from the accumulated
|
||||
* stream data.
|
||||
*
|
||||
* @returns An object with methods to process chunks and build metadata from a stream
|
||||
*/
|
||||
createStreamExtractor: () => {
|
||||
/**
|
||||
* Process an individual chunk from the stream. Called for each chunk in the response stream
|
||||
* to accumulate metadata throughout the streaming process.
|
||||
*
|
||||
* @param parsedChunk - The parsed JSON response chunk from the provider's API
|
||||
*/
|
||||
processChunk(parsedChunk: unknown): void
|
||||
|
||||
/**
|
||||
* Builds the metadata object after all chunks have been processed.
|
||||
* Called at the end of the stream to generate the complete provider metadata.
|
||||
*
|
||||
* @returns Provider-specific metadata or undefined if no metadata is available.
|
||||
* The metadata should be under a key indicating the provider id.
|
||||
*/
|
||||
buildMetadata(): SharedV3ProviderMetadata | undefined
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,83 @@
|
||||
import { type LanguageModelV3CallOptions, type SharedV3Warning, UnsupportedFunctionalityError } from "@ai-sdk/provider"
|
||||
|
||||
export function prepareTools({
|
||||
tools,
|
||||
toolChoice,
|
||||
}: {
|
||||
tools: LanguageModelV3CallOptions["tools"]
|
||||
toolChoice?: LanguageModelV3CallOptions["toolChoice"]
|
||||
}): {
|
||||
tools:
|
||||
| undefined
|
||||
| Array<{
|
||||
type: "function"
|
||||
function: {
|
||||
name: string
|
||||
description: string | undefined
|
||||
parameters: unknown
|
||||
}
|
||||
}>
|
||||
toolChoice: { type: "function"; function: { name: string } } | "auto" | "none" | "required" | undefined
|
||||
toolWarnings: SharedV3Warning[]
|
||||
} {
|
||||
// when the tools array is empty, change it to undefined to prevent errors:
|
||||
tools = tools?.length ? tools : undefined
|
||||
|
||||
const toolWarnings: SharedV3Warning[] = []
|
||||
|
||||
if (tools == null) {
|
||||
return { tools: undefined, toolChoice: undefined, toolWarnings }
|
||||
}
|
||||
|
||||
const openaiCompatTools: Array<{
|
||||
type: "function"
|
||||
function: {
|
||||
name: string
|
||||
description: string | undefined
|
||||
parameters: unknown
|
||||
}
|
||||
}> = []
|
||||
|
||||
for (const tool of tools) {
|
||||
if (tool.type === "provider") {
|
||||
toolWarnings.push({ type: "unsupported", feature: `tool type: ${tool.type}` })
|
||||
} else {
|
||||
openaiCompatTools.push({
|
||||
type: "function",
|
||||
function: {
|
||||
name: tool.name,
|
||||
description: tool.description,
|
||||
parameters: tool.inputSchema,
|
||||
},
|
||||
})
|
||||
}
|
||||
}
|
||||
|
||||
if (toolChoice == null) {
|
||||
return { tools: openaiCompatTools, toolChoice: undefined, toolWarnings }
|
||||
}
|
||||
|
||||
const type = toolChoice.type
|
||||
|
||||
switch (type) {
|
||||
case "auto":
|
||||
case "none":
|
||||
case "required":
|
||||
return { tools: openaiCompatTools, toolChoice: type, toolWarnings }
|
||||
case "tool":
|
||||
return {
|
||||
tools: openaiCompatTools,
|
||||
toolChoice: {
|
||||
type: "function",
|
||||
function: { name: toolChoice.toolName },
|
||||
},
|
||||
toolWarnings,
|
||||
}
|
||||
default: {
|
||||
const _exhaustiveCheck: never = type
|
||||
throw new UnsupportedFunctionalityError({
|
||||
functionality: `tool choice type: ${_exhaustiveCheck}`,
|
||||
})
|
||||
}
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user