feat: add support for using a reverse proxy to use the official ChatGPT models

2023-02-12 23:36:42 -06:00 · 2023-02-12 23:36:42 -06:00 · 1777c4551d
commit 1777c4551d
--- a/demos/demo-reverse-proxy.ts
+++ b/demos/demo-reverse-proxy.ts
@ -0,0 +1,82 @@
+import dotenv from 'dotenv-safe'
+import { oraPromise } from 'ora'
+
+import { ChatGPTAPI } from '../src'
+
+dotenv.config()
+
+/**
+ * Demo CLI for testing conversation support using a reverse proxy that mimic's.
+ * OpenAI's completions API ChatGPT's unofficial API.
+ *
+ * ```
+ * npx tsx demos/demo-reverse-proxy.ts
+ * ```
+ */
+async function main() {
+  const api = new ChatGPTAPI({
+    apiReverseProxyUrl: 'https://chatgpt.pawan.krd/api/completions',
+    // change this to an `accessToken` extracted from the ChatGPT site's `https://chat.openai.com/api/auth/session` response
+    apiKey: process.env.OPENAI_ACCESS_TOKEN,
+    completionParams: {
+      // override this depending on the ChatGPT model you want to use
+      // NOTE: if you are on a paid plan, you can't use the free model and vice-versa
+      // model: 'text-davinci-002-render' // free, default model
+      model: 'text-davinci-002-render-paid' // paid, default model
+      // model: 'text-davinci-002-render-sha' // paid, turbo model
+    },
+    debug: false
+  })
+
+  const prompt = 'Write a poem about cats.'
+
+  let res = await oraPromise(api.sendMessage(prompt), {
+    text: prompt
+  })
+
+  console.log('\n' + res.text + '\n')
+
+  const prompt2 = 'Can you make it cuter and shorter?'
+
+  res = await oraPromise(
+    api.sendMessage(prompt2, {
+      conversationId: res.conversationId,
+      parentMessageId: res.id
+    }),
+    {
+      text: prompt2
+    }
+  )
+  console.log('\n' + res.text + '\n')
+
+  const prompt3 = 'Now write it in French.'
+
+  res = await oraPromise(
+    api.sendMessage(prompt3, {
+      conversationId: res.conversationId,
+      parentMessageId: res.id
+    }),
+    {
+      text: prompt3
+    }
+  )
+  console.log('\n' + res.text + '\n')
+
+  const prompt4 = 'What were we talking about again?'
+
+  res = await oraPromise(
+    api.sendMessage(prompt4, {
+      conversationId: res.conversationId,
+      parentMessageId: res.id
+    }),
+    {
+      text: prompt4
+    }
+  )
+  console.log('\n' + res.text + '\n')
+}
+
+main().catch((err) => {
+  console.error(err)
+  process.exit(1)
+})
--- a/src/chatgpt-api.ts
+++ b/src/chatgpt-api.ts
@ -23,6 +23,7 @@ const ASSISTANT_LABEL_DEFAULT = 'ChatGPT'
 export class ChatGPTAPI {
  protected _apiKey: string
  protected _apiBaseUrl: string
+  protected _apiReverseProxyUrl: string
  protected _debug: boolean

  protected _completionParams: Omit<types.openai.CompletionParams, 'prompt'>
@ -30,6 +31,8 @@ export class ChatGPTAPI {
  protected _maxResponseTokens: number
  protected _userLabel: string
  protected _assistantLabel: string
+  protected _endToken: string
+  protected _sepToken: string

  protected _getMessageById: types.GetMessageByIdFunction
  protected _upsertMessage: types.UpsertMessageFunction
@ -42,6 +45,7 @@ export class ChatGPTAPI {
   *
   * @param apiKey - OpenAI API key (required).
   * @param apiBaseUrl - Optional override for the OpenAI API base URL.
+   * @param apiReverseProxyUrl - Optional override for a reverse proxy URL to use instead of the OpenAI API completions API.
   * @param debug - Optional enables logging debugging info to stdout.
   * @param completionParams - Param overrides to send to the [OpenAI completion API](https://platform.openai.com/docs/api-reference/completions/create). Options like `temperature` and `presence_penalty` can be tweaked to change the personality of the assistant.
   * @param maxModelTokens - Optional override for the maximum number of tokens allowed by the model's context. Defaults to 4096 for the `text-chat-davinci-002-20230126` model.
@ -56,6 +60,9 @@ export class ChatGPTAPI {
    /** @defaultValue `'https://api.openai.com'` **/
    apiBaseUrl?: string

+    /** @defaultValue `undefined` **/
+    apiReverseProxyUrl?: string
+
    /** @defaultValue `false` **/
    debug?: boolean

@ -80,6 +87,7 @@ export class ChatGPTAPI {
    const {
      apiKey,
      apiBaseUrl = 'https://api.openai.com',
+      apiReverseProxyUrl,
      debug = false,
      messageStore,
      completionParams,
@ -93,15 +101,33 @@ export class ChatGPTAPI {

    this._apiKey = apiKey
    this._apiBaseUrl = apiBaseUrl
+    this._apiReverseProxyUrl = apiReverseProxyUrl
    this._debug = !!debug

    this._completionParams = {
      model: CHATGPT_MODEL,
-      temperature: 0.7,
-      presence_penalty: 0.6,
-      stop: ['<|im_end|>'],
+      temperature: 0.8,
+      top_p: 1.0,
+      presence_penalty: 1.0,
      ...completionParams
    }
+
+    if (this._isChatGPTModel) {
+      this._endToken = '<|im_end|>'
+      this._sepToken = '<|im_sep|>'
+
+      if (!this._completionParams.stop) {
+        this._completionParams.stop = [this._endToken, this._sepToken]
+      }
+    } else {
+      this._endToken = '<|endoftext|>'
+      this._sepToken = this._endToken
+
+      if (!this._completionParams.stop) {
+        this._completionParams.stop = [this._endToken]
+      }
+    }
+
    this._maxModelTokens = maxModelTokens
    this._maxResponseTokens = maxResponseTokens
    this._userLabel = userLabel
@ -190,7 +216,8 @@ export class ChatGPTAPI {

    const responseP = new Promise<types.ChatMessage>(
      async (resolve, reject) => {
-        const url = `${this._apiBaseUrl}/v1/completions`
+        const url =
+          this._apiReverseProxyUrl || `${this._apiBaseUrl}/v1/completions`
        const headers = {
          'Content-Type': 'application/json',
          Authorization: `Bearer ${this._apiKey}`
@ -223,9 +250,13 @@ export class ChatGPTAPI {
                const response: types.openai.CompletionResponse =
                  JSON.parse(data)

-                if (response?.id && response?.choices?.length) {
+                if (response.id) {
                  result.id = response.id
+                }
+
+                if (response?.choices?.length) {
                  result.text += response.choices[0].text
+                  result.detail = response

                  onProgress?.(result)
                }
@ -260,8 +291,22 @@ export class ChatGPTAPI {
              console.log(response)
            }

-            result.id = response.id
-            result.text = response.choices[0].text.trim()
+            if (response.id) {
+              result.id = response.id
+            }
+            if (response?.choices?.length) {
+              result.text = response.choices[0].text.trim()
+            } else {
+              return reject(
+                new Error(
+                  `ChatGPT error: ${
+                    (response as any).detail || response || 'unknown'
+                  }`
+                )
+              )
+            }
+
+            result.detail = response

            return resolve(result)
          } catch (err) {
@ -306,13 +351,13 @@ export class ChatGPTAPI {

    const promptPrefix =
      opts.promptPrefix ||
-      `You are ${this._assistantLabel}, a large language model trained by OpenAI. You answer as concisely as possible for each response (e.g. don’t be verbose). It is very important that you answer as concisely as possible, so please remember this. If you are generating a list, do not have too many items. Keep the number of items short.
-Current date: ${currentDate}\n\n`
+      `Instructions:\nYou are ${this._assistantLabel}, a large language model trained by OpenAI.
+Current date: ${currentDate}${this._sepToken}\n\n`
    const promptSuffix = opts.promptSuffix || `\n\n${this._assistantLabel}:\n`

    const maxNumTokens = this._maxModelTokens - this._maxResponseTokens
    let { parentMessageId } = opts
-    let nextPromptBody = `${this._userLabel}:\n\n${message}${this._completionParams.stop[0]}`
+    let nextPromptBody = `${this._userLabel}:\n\n${message}${this._endToken}`
    let promptBody = ''
    let prompt: string
    let numTokens: number
@ -348,7 +393,7 @@ Current date: ${currentDate}\n\n`
        parentMessageRole === 'user' ? this._userLabel : this._assistantLabel

      // TODO: differentiate between assistant and user messages
-      const parentMessageString = `${parentMessageRoleDesc}:\n\n${parentMessage.text}${this._completionParams.stop[0]}\n\n`
+      const parentMessageString = `${parentMessageRoleDesc}:\n\n${parentMessage.text}${this._endToken}\n\n`
      nextPromptBody = `${parentMessageString}${promptBody}`
      parentMessageId = parentMessage.parentMessageId
    } while (true)
@ -364,24 +409,35 @@ Current date: ${currentDate}\n\n`
  }

  protected async _getTokenCount(text: string) {
-    if (this._completionParams.model === CHATGPT_MODEL) {
+    if (this._isChatGPTModel) {
      // With this model, "<|im_end|>" is 1 token, but tokenizers aren't aware of it yet.
      // Replace it with "<|endoftext|>" (which it does know about) so that the tokenizer can count it as 1 token.
      text = text.replace(/<\|im_end\|>/g, '<|endoftext|>')
+      text = text.replace(/<\|im_sep\|>/g, '<|endoftext|>')
    }

    return gptEncode(text).length
  }

+  protected get _isChatGPTModel() {
+    return (
+      this._completionParams.model.startsWith('text-chat') ||
+      this._completionParams.model.startsWith('text-davinci-002-render')
+    )
+  }
+
  protected async _defaultGetMessageById(
    id: string
  ): Promise<types.ChatMessage> {
-    return this._messageStore.get(id)
+    const res = await this._messageStore.get(id)
+    console.log('getMessageById', id, res)
+    return res
  }

  protected async _defaultUpsertMessage(
    message: types.ChatMessage
  ): Promise<void> {
-    this._messageStore.set(message.id, message)
+    console.log('upsertMessage', message.id, message)
+    await this._messageStore.set(message.id, message)
  }
 }
--- a/src/types.ts
+++ b/src/types.ts
@ -18,6 +18,7 @@ export interface ChatMessage {
  role: Role
  parentMessageId?: string
  conversationId?: string
+  detail?: any
 }

 export class ChatGPTError extends Error {