@genkit-ai/compat-oai
Version:
Genkit AI framework plugin for OpenAI APIs.
348 lines • 9.33 kB
JavaScript
import { GenerationCommonConfigSchema, Message, modelRef, z } from "genkit";
const VisualDetailLevelSchema = z.enum(["auto", "low", "high"]).optional();
const ChatCompletionCommonConfigSchema = GenerationCommonConfigSchema.extend({
temperature: z.number().min(0).max(2).optional(),
frequencyPenalty: z.number().min(-2).max(2).optional(),
logProbs: z.boolean().optional(),
presencePenalty: z.number().min(-2).max(2).optional(),
topLogProbs: z.number().int().min(0).max(20).optional()
});
function toOpenAIRole(role) {
switch (role) {
case "user":
return "user";
case "model":
return "assistant";
case "system":
return "system";
case "tool":
return "tool";
default:
throw new Error(`role ${role} doesn't map to an OpenAI role.`);
}
}
function toOpenAITool(tool) {
return {
type: "function",
function: {
name: tool.name,
parameters: tool.inputSchema !== null ? tool.inputSchema : void 0
}
};
}
function toOpenAITextAndMedia(part, visualDetailLevel) {
if (part.text) {
return {
type: "text",
text: part.text
};
} else if (part.media) {
return {
type: "image_url",
image_url: {
url: part.media.url,
detail: visualDetailLevel
}
};
}
throw Error(
`Unsupported genkit part fields encountered for current message role: ${JSON.stringify(part)}.`
);
}
function toOpenAIMessages(messages, visualDetailLevel = "auto") {
const apiMessages = [];
for (const message of messages) {
const msg = new Message(message);
const role = toOpenAIRole(message.role);
switch (role) {
case "user":
const content = msg.content.map(
(part) => toOpenAITextAndMedia(part, visualDetailLevel)
);
const onlyTextContent = content.some((item) => item.type !== "text");
if (!onlyTextContent) {
content.forEach((item) => {
if (item.type === "text") {
apiMessages.push({
role,
content: item.text
});
}
});
} else {
apiMessages.push({
role,
content
});
}
break;
case "system":
apiMessages.push({
role,
content: msg.text
});
break;
case "assistant": {
const toolCalls = msg.content.filter(
(part) => Boolean(part.toolRequest)
).map((part) => ({
id: part.toolRequest.ref ?? "",
type: "function",
function: {
name: part.toolRequest.name,
arguments: JSON.stringify(part.toolRequest.input)
}
}));
if (toolCalls.length > 0) {
apiMessages.push({
role,
tool_calls: toolCalls
});
} else {
apiMessages.push({
role,
content: msg.text
});
}
break;
}
case "tool": {
const toolResponseParts = msg.toolResponseParts();
toolResponseParts.map((part) => {
apiMessages.push({
role,
tool_call_id: part.toolResponse.ref ?? "",
content: typeof part.toolResponse.output === "string" ? part.toolResponse.output : JSON.stringify(part.toolResponse.output)
});
});
break;
}
}
}
return apiMessages;
}
const finishReasonMap = {
length: "length",
stop: "stop",
tool_calls: "stop",
content_filter: "blocked"
};
function fromOpenAIToolCall(toolCall, choice) {
if (!toolCall.function) {
throw Error(
`Unexpected openAI chunk choice. tool_calls was provided but one or more tool_calls is missing.`
);
}
const f = toolCall.function;
if (choice.finish_reason === "tool_calls") {
return {
toolRequest: {
name: f.name,
ref: toolCall.id,
input: f.arguments ? JSON.parse(f.arguments) : f.arguments
}
};
} else {
return {
toolRequest: {
name: f.name,
ref: toolCall.id,
input: ""
}
};
}
}
function fromOpenAIChoice(choice, jsonMode = false) {
const toolRequestParts = choice.message.tool_calls?.map(
(toolCall) => fromOpenAIToolCall(toolCall, choice)
);
return {
finishReason: finishReasonMap[choice.finish_reason] || "other",
message: {
role: "model",
content: toolRequestParts ? (
// Note: Not sure why I have to cast here exactly.
// Otherwise it thinks toolRequest must be 'undefined' if provided
toolRequestParts
) : [
jsonMode ? { data: JSON.parse(choice.message.content) } : { text: choice.message.content }
]
}
};
}
function fromOpenAIChunkChoice(choice, jsonMode = false) {
const toolRequestParts = choice.delta.tool_calls?.map(
(toolCall) => fromOpenAIToolCall(toolCall, choice)
);
return {
finishReason: choice.finish_reason ? finishReasonMap[choice.finish_reason] || "other" : "unknown",
message: {
role: "model",
content: toolRequestParts ? (
// Note: Not sure why I have to cast here exactly.
// Otherwise it thinks toolRequest must be 'undefined' if provided
toolRequestParts
) : [
jsonMode ? { data: JSON.parse(choice.delta.content) } : { text: choice.delta.content }
]
}
};
}
function toOpenAIRequestBody(modelName, request, requestBuilder) {
const messages = toOpenAIMessages(
request.messages,
request.config?.visualDetailLevel
);
const {
temperature,
maxOutputTokens,
// unused
topK,
// unused
topP: top_p,
frequencyPenalty: frequency_penalty,
logProbs: logprobs,
presencePenalty: presence_penalty,
topLogProbs: top_logprobs,
stopSequences: stop,
version: modelVersion,
tools: toolsFromConfig,
...restOfConfig
} = request.config ?? {};
const tools = request.tools?.map(toOpenAITool) ?? [];
if (toolsFromConfig) {
tools.push(...toolsFromConfig);
}
let body = {
model: modelVersion ?? modelName,
messages,
tools: tools.length > 0 ? tools : void 0,
temperature,
top_p,
stop,
frequency_penalty,
presence_penalty,
top_logprobs,
logprobs
};
if (requestBuilder) {
requestBuilder(request, body);
} else {
body = { ...body, ...restOfConfig };
}
const response_format = request.output?.format;
if (response_format === "json") {
body.response_format = {
type: "json_object"
};
} else if (response_format === "text") {
body.response_format = {
type: "text"
};
}
for (const key in body) {
if (!body[key] || Array.isArray(body[key]) && !body[key].length)
delete body[key];
}
return body;
}
function openAIModelRunner(name, client, requestBuilder) {
return async (request, options) => {
let response;
const body = toOpenAIRequestBody(name, request, requestBuilder);
if (options?.streamingRequested) {
const stream = client.beta.chat.completions.stream(
{
...body,
stream: true,
stream_options: {
include_usage: true
}
},
{ signal: options?.abortSignal }
);
for await (const chunk of stream) {
chunk.choices?.forEach((chunk2) => {
const c = fromOpenAIChunkChoice(chunk2);
options?.sendChunk({
index: chunk2.index,
content: c.message?.content ?? []
});
});
}
response = await stream.finalChatCompletion();
} else {
response = await client.chat.completions.create(body, {
signal: options?.abortSignal
});
}
const standardResponse = {
usage: {
inputTokens: response.usage?.prompt_tokens,
outputTokens: response.usage?.completion_tokens,
totalTokens: response.usage?.total_tokens
},
raw: response
};
if (response.choices.length === 0) {
return standardResponse;
} else {
const choice = response.choices[0];
return {
...fromOpenAIChoice(choice, request.output?.format === "json"),
...standardResponse
};
}
};
}
function defineCompatOpenAIModel(params) {
const { ai, name, client, modelRef: modelRef2, requestBuilder } = params;
const modelName = name.substring(name.indexOf("/") + 1);
return ai.defineModel(
{
name,
apiVersion: "v2",
...modelRef2?.info,
configSchema: modelRef2?.configSchema
},
openAIModelRunner(modelName, client, requestBuilder)
);
}
const GENERIC_MODEL_INFO = {
supports: {
multiturn: true,
media: true,
tools: true,
toolChoice: true,
systemRole: true
}
};
function compatOaiModelRef(params) {
const {
name,
info = GENERIC_MODEL_INFO,
configSchema,
config = void 0
} = params;
return modelRef({
name,
configSchema: configSchema || ChatCompletionCommonConfigSchema,
info,
config
});
}
export {
ChatCompletionCommonConfigSchema,
compatOaiModelRef,
defineCompatOpenAIModel,
fromOpenAIChoice,
fromOpenAIChunkChoice,
fromOpenAIToolCall,
openAIModelRunner,
toOpenAIMessages,
toOpenAIRequestBody,
toOpenAIRole,
toOpenAITextAndMedia,
toOpenAITool
};
//# sourceMappingURL=model.mjs.map