anything-llm/server/utils/AiProviders/openAi/index.js

243 lines
6.8 KiB
JavaScript
Raw Normal View History

const { OpenAiEmbedder } = require("../../EmbeddingEngines/openAi");
const { chatPrompt } = require("../../chats");
class OpenAiLLM {
constructor(embedder = null, modelPreference = null) {
const { Configuration, OpenAIApi } = require("openai");
if (!process.env.OPEN_AI_KEY) throw new Error("No OpenAI API key was set.");
2023-06-08 06:31:35 +02:00
const config = new Configuration({
apiKey: process.env.OPEN_AI_KEY,
});
this.openai = new OpenAIApi(config);
this.model =
modelPreference || process.env.OPEN_MODEL_PREF || "gpt-3.5-turbo";
this.limits = {
history: this.promptWindowLimit() * 0.15,
system: this.promptWindowLimit() * 0.15,
user: this.promptWindowLimit() * 0.7,
};
if (!embedder)
console.warn(
"No embedding provider defined for OpenAiLLM - falling back to OpenAiEmbedder for embedding!"
);
this.embedder = !embedder ? new OpenAiEmbedder() : embedder;
this.defaultTemp = 0.7;
}
#appendContext(contextTexts = []) {
if (!contextTexts || !contextTexts.length) return "";
return (
"\nContext:\n" +
contextTexts
.map((text, i) => {
return `[CONTEXT ${i}]:\n${text}\n[END CONTEXT ${i}]\n\n`;
})
.join("")
);
}
streamingEnabled() {
return "streamChat" in this && "streamGetChatCompletion" in this;
}
promptWindowLimit() {
switch (this.model) {
case "gpt-3.5-turbo":
return 4096;
case "gpt-3.5-turbo-1106":
return 16385;
case "gpt-4":
return 8192;
case "gpt-4-1106-preview":
return 128000;
case "gpt-4-turbo-preview":
return 128000;
case "gpt-4-32k":
return 32000;
default:
return 4096; // assume a fine-tune 3.5
}
2023-06-04 04:28:07 +02:00
}
async isValidChatCompletionModel(modelName = "") {
const validModels = [
"gpt-4",
"gpt-3.5-turbo",
"gpt-3.5-turbo-1106",
"gpt-4-1106-preview",
"gpt-4-turbo-preview",
"gpt-4-32k",
];
const isPreset = validModels.some((model) => modelName === model);
if (isPreset) return true;
const model = await this.openai
.retrieveModel(modelName)
.then((res) => res.data)
.catch(() => null);
return !!model;
2023-06-04 04:28:07 +02:00
}
constructPrompt({
systemPrompt = "",
contextTexts = [],
chatHistory = [],
userPrompt = "",
}) {
const prompt = {
role: "system",
content: `${systemPrompt}${this.#appendContext(contextTexts)}`,
};
return [prompt, ...chatHistory, { role: "user", content: userPrompt }];
}
2023-06-08 06:31:35 +02:00
async isSafe(input = "") {
const { flagged = false, categories = {} } = await this.openai
.createModeration({ input })
2023-06-04 04:28:07 +02:00
.then((json) => {
const res = json.data;
2023-06-08 06:31:35 +02:00
if (!res.hasOwnProperty("results"))
throw new Error("OpenAI moderation: No results!");
if (res.results.length === 0)
throw new Error("OpenAI moderation: No results length!");
return res.results[0];
})
.catch((error) => {
throw new Error(
`OpenAI::CreateModeration failed with: ${error.message}`
);
2023-06-08 06:31:35 +02:00
});
2023-06-04 04:28:07 +02:00
if (!flagged) return { safe: true, reasons: [] };
2023-06-08 06:31:35 +02:00
const reasons = Object.keys(categories)
.map((category) => {
const value = categories[category];
if (value === true) {
return category.replace("/", " or ");
} else {
return null;
}
})
.filter((reason) => !!reason);
2023-06-04 04:28:07 +02:00
2023-06-08 06:31:35 +02:00
return { safe: false, reasons };
2023-06-04 04:28:07 +02:00
}
async sendChat(chatHistory = [], prompt, workspace = {}, rawHistory = []) {
if (!(await this.isValidChatCompletionModel(this.model)))
2023-06-08 06:31:35 +02:00
throw new Error(
`OpenAI chat: ${this.model} is not valid for chat completion!`
2023-06-08 06:31:35 +02:00
);
2023-06-04 04:28:07 +02:00
2023-06-08 06:31:35 +02:00
const textResponse = await this.openai
.createChatCompletion({
model: this.model,
temperature: Number(workspace?.openAiTemp ?? this.defaultTemp),
2023-06-08 06:31:35 +02:00
n: 1,
messages: await this.compressMessages(
{
systemPrompt: chatPrompt(workspace),
userPrompt: prompt,
chatHistory,
},
rawHistory
),
2023-06-04 04:28:07 +02:00
})
2023-06-08 06:31:35 +02:00
.then((json) => {
const res = json.data;
if (!res.hasOwnProperty("choices"))
throw new Error("OpenAI chat: No results!");
if (res.choices.length === 0)
throw new Error("OpenAI chat: No results length!");
return res.choices[0].message.content;
})
.catch((error) => {
throw new Error(
`OpenAI::createChatCompletion failed with: ${error.message}`
);
2023-06-08 06:31:35 +02:00
});
2023-06-04 04:28:07 +02:00
2023-06-08 06:31:35 +02:00
return textResponse;
2023-06-04 04:28:07 +02:00
}
async streamChat(chatHistory = [], prompt, workspace = {}, rawHistory = []) {
if (!(await this.isValidChatCompletionModel(this.model)))
throw new Error(
`OpenAI chat: ${this.model} is not valid for chat completion!`
);
const streamRequest = await this.openai.createChatCompletion(
{
model: this.model,
stream: true,
temperature: Number(workspace?.openAiTemp ?? this.defaultTemp),
n: 1,
messages: await this.compressMessages(
{
systemPrompt: chatPrompt(workspace),
userPrompt: prompt,
chatHistory,
},
rawHistory
),
},
{ responseType: "stream" }
);
return streamRequest;
}
async getChatCompletion(messages = null, { temperature = 0.7 }) {
if (!(await this.isValidChatCompletionModel(this.model)))
throw new Error(
`OpenAI chat: ${this.model} is not valid for chat completion!`
);
const { data } = await this.openai.createChatCompletion({
model: this.model,
messages,
temperature,
});
if (!data.hasOwnProperty("choices")) return null;
return data.choices[0].message.content;
}
async streamGetChatCompletion(messages = null, { temperature = 0.7 }) {
if (!(await this.isValidChatCompletionModel(this.model)))
throw new Error(
`OpenAI chat: ${this.model} is not valid for chat completion!`
);
const streamRequest = await this.openai.createChatCompletion(
{
model: this.model,
stream: true,
messages,
temperature,
},
{ responseType: "stream" }
);
return streamRequest;
}
// Simple wrapper for dynamic embedder & normalize interface for all LLM implementations
async embedTextInput(textInput) {
return await this.embedder.embedTextInput(textInput);
}
async embedChunks(textChunks = []) {
return await this.embedder.embedChunks(textChunks);
}
async compressMessages(promptArgs = {}, rawHistory = []) {
const { messageArrayCompressor } = require("../../helpers/chat");
const messageArray = this.constructPrompt(promptArgs);
return await messageArrayCompressor(this, messageArray, rawHistory);
}
2023-06-04 04:28:07 +02:00
}
module.exports = {
OpenAiLLM,
2023-06-04 04:28:07 +02:00
};