Files
OAI-Proxy/src/shared/api-schemas/mistral-ai.ts
T

165 lines
5.4 KiB
TypeScript

import { z } from "zod";
import { OPENAI_OUTPUT_MAX } from "./openai";
import { Template } from "@huggingface/jinja";
import { APIFormatTransformer } from "./index";
import { logger } from "../../logger";
const MistralChatMessageSchema = z.object({
role: z.enum(["system", "user", "assistant", "tool"]), // TODO: implement tools
content: z.string(),
prefix: z.boolean().optional(),
});
const MistralMessagesSchema = z.array(MistralChatMessageSchema).refine(
(input) => {
const prefixIdx = input.findIndex((msg) => Boolean(msg.prefix));
if (prefixIdx === -1) return true; // no prefix messages
const lastIdx = input.length - 1;
const lastMsg = input[lastIdx];
return prefixIdx === lastIdx && lastMsg.role === "assistant";
},
{
message:
"`prefix` can only be set to `true` on the last message, and only for an assistant message.",
}
);
// https://docs.mistral.ai/api#operation/createChatCompletion
const BaseMistralAIV1CompletionsSchema = z.object({
model: z.string(),
messages: MistralMessagesSchema.optional(),
prompt: z.string().optional(),
temperature: z.number().optional().default(0.7),
top_p: z.number().optional().default(1),
max_tokens: z.coerce
.number()
.int()
.nullish()
.transform((v) => Math.min(v ?? OPENAI_OUTPUT_MAX, OPENAI_OUTPUT_MAX)),
stream: z.boolean().optional().default(false),
// Mistral docs say that `stop` can be a string or array but AWS Mistral
// blows up if a string is passed. We must convert it to an array.
stop: z
.union([z.string(), z.array(z.string())])
.optional()
.default([])
.transform((v) => (Array.isArray(v) ? v : [v])),
random_seed: z.number().int().min(0).optional(),
response_format: z.enum(["text", "json_object"]).optional().default("text"),
safe_prompt: z.boolean().optional().default(false),
});
export const MistralAIV1ChatCompletionsSchema =
BaseMistralAIV1CompletionsSchema.and(
z.object({ messages: MistralMessagesSchema })
);
export const MistralAIV1TextCompletionsSchema =
BaseMistralAIV1CompletionsSchema.and(z.object({ prompt: z.string() }));
/*
Slightly more strict version that only allows a subset of the parameters. AWS
Mistral helpfully returns no details if unsupported parameters are passed so
this list comes from trial and error as of 2024-08-12.
*/
const BaseAWSMistralAIV1CompletionsSchema =
BaseMistralAIV1CompletionsSchema.pick({
temperature: true,
top_p: true,
max_tokens: true,
stop: true,
random_seed: true,
// response_format: true,
// safe_prompt: true,
}).strip();
export const AWSMistralV1ChatCompletionsSchema =
BaseAWSMistralAIV1CompletionsSchema.and(
z.object({ messages: MistralMessagesSchema })
);
export const AWSMistralV1TextCompletionsSchema =
BaseAWSMistralAIV1CompletionsSchema.and(z.object({ prompt: z.string() }));
export type MistralAIChatMessage = z.infer<typeof MistralChatMessageSchema>;
export function fixMistralPrompt(
messages: MistralAIChatMessage[]
): MistralAIChatMessage[] {
// Mistral uses OpenAI format but has some additional requirements:
// - Only one system message per request, and it must be the first message if
// present.
// - Final message must be a user message, unless it has `prefix: true`.
// - Cannot have multiple messages from the same role in a row.
// While frontends should be able to handle this, we can fix it here in the
// meantime.
const fixed = messages.reduce<MistralAIChatMessage[]>((acc, msg) => {
if (acc.length === 0) {
acc.push(msg);
return acc;
}
const copy = { ...msg };
// Reattribute subsequent system messages to the user
if (msg.role === "system") {
copy.role = "user";
}
// Consolidate multiple messages from the same role
const last = acc[acc.length - 1];
if (last.role === copy.role) {
last.content += "\n\n" + copy.content;
} else {
acc.push(copy);
}
return acc;
}, []);
// If the last message is an assistant message, mark it as a prefix. An
// assistant message at the end of the conversation without `prefix: true`
// results in an error.
if (fixed[fixed.length - 1].role === "assistant") {
fixed[fixed.length - 1].prefix = true;
}
return fixed;
}
let jinjaTemplate: Template;
let renderTemplate: (messages: MistralAIChatMessage[]) => string;
function renderMistralPrompt(messages: MistralAIChatMessage[]) {
if (!jinjaTemplate) {
logger.warn("Lazy loading mistral chat template...");
const { chatTemplate, bosToken, eosToken } =
require("./templates/mistral-template").MISTRAL_TEMPLATE;
jinjaTemplate = new Template(chatTemplate);
renderTemplate = (messages) =>
jinjaTemplate.render({
messages,
bos_token: bosToken,
eos_token: eosToken,
});
}
return renderTemplate(messages);
}
/**
* Attempts to convert a Mistral chat completions request to a text completions,
* using the official prompt template published by Mistral.
*/
export const transformMistralChatToText: APIFormatTransformer<
typeof MistralAIV1TextCompletionsSchema
> = async (req) => {
const { body } = req;
const result = MistralAIV1ChatCompletionsSchema.safeParse(body);
if (!result.success) {
req.log.warn(
{ issues: result.error.issues, body },
"Invalid Mistral chat completions request"
);
throw result.error;
}
const { messages, ...rest } = result.data;
const prompt = renderMistralPrompt(messages);
return { ...rest, prompt, messages: undefined };
};