next-chat/app/client/platforms/google.ts

import { ApiPath, Google, REQUEST_TIMEOUT_MS } from "@/app/constant";
import {
  ChatOptions,
  getHeaders,
  LLMApi,
  LLMModel,
  LLMUsage,
  SpeechOptions,
} from "../api";
import { useAccessStore, useAppConfig, useChatStore } from "@/app/store";
import { getClientConfig } from "@/app/config/client";
import { GEMINI_BASE_URL } from "@/app/constant";
import Locale from "../../locales";
import {
  EventStreamContentType,
  fetchEventSource,
} from "@fortaine/fetch-event-source";
import { prettyObject } from "@/app/utils/format";
import {
  getMessageTextContent,
  getMessageImages,
  isVisionModel,
} from "@/app/utils";
import { preProcessImageContent } from "@/app/utils/chat";
import { fetch } from "@/app/utils/stream";

export class GeminiProApi implements LLMApi {
  path(path: string): string {
    const accessStore = useAccessStore.getState();

    let baseUrl = "";
    if (accessStore.useCustomConfig) {
      baseUrl = accessStore.googleUrl;
    }

    const isApp = !!getClientConfig()?.isApp;
    if (baseUrl.length === 0) {
      baseUrl = isApp ? GEMINI_BASE_URL : ApiPath.Google;
    }
    if (baseUrl.endsWith("/")) {
      baseUrl = baseUrl.slice(0, baseUrl.length - 1);
    }
    if (!baseUrl.startsWith("http") && !baseUrl.startsWith(ApiPath.Google)) {
      baseUrl = "https://" + baseUrl;
    }

    console.log("[Proxy Endpoint] ", baseUrl, path);

    let chatPath = [baseUrl, path].join("/");

    chatPath += chatPath.includes("?") ? "&alt=sse" : "?alt=sse";
    return chatPath;
  }
  extractMessage(res: any) {
    console.log("[Response] gemini-pro response: ", res);

    return (
      res?.candidates?.at(0)?.content?.parts.at(0)?.text ||
      res?.error?.message ||
      ""
    );
  }
  speech(options: SpeechOptions): Promise<ArrayBuffer> {
    throw new Error("Method not implemented.");
  }

  async chat(options: ChatOptions): Promise<void> {
    const apiClient = this;
    let multimodal = false;

    // try get base64image from local cache image_url
    const _messages: ChatOptions["messages"] = [];
    for (const v of options.messages) {
      const content = await preProcessImageContent(v.content);
      _messages.push({ role: v.role, content });
    }
    const messages = _messages.map((v) => {
      let parts: any[] = [{ text: getMessageTextContent(v) }];
      if (isVisionModel(options.config.model)) {
        const images = getMessageImages(v);
        if (images.length > 0) {
          multimodal = true;
          parts = parts.concat(
            images.map((image) => {
              const imageType = image.split(";")[0].split(":")[1];
              const imageData = image.split(",")[1];
              return {
                inline_data: {
                  mime_type: imageType,
                  data: imageData,
                },
              };
            }),
          );
        }
      }
      return {
        role: v.role.replace("assistant", "model").replace("system", "user"),
        parts: parts,
      };
    });

    // google requires that role in neighboring messages must not be the same
    for (let i = 0; i < messages.length - 1; ) {
      // Check if current and next item both have the role "model"
      if (messages[i].role === messages[i + 1].role) {
        // Concatenate the 'parts' of the current and next item
        messages[i].parts = messages[i].parts.concat(messages[i + 1].parts);
        // Remove the next item
        messages.splice(i + 1, 1);
      } else {
        // Move to the next item
        i++;
      }
    }
    // if (visionModel && messages.length > 1) {
    //   options.onError?.(new Error("Multiturn chat is not enabled for models/gemini-pro-vision"));
    // }

    const accessStore = useAccessStore.getState();

    const modelConfig = {
      ...useAppConfig.getState().modelConfig,
      ...useChatStore.getState().currentSession().mask.modelConfig,
      ...{
        model: options.config.model,
      },
    };
    const requestPayload = {
      contents: messages,
      generationConfig: {
        // stopSequences: [
        //   "Title"
        // ],
        temperature: modelConfig.temperature,
        maxOutputTokens: modelConfig.max_tokens,
        topP: modelConfig.top_p,
        // "topK": modelConfig.top_k,
      },
      safetySettings: [
        {
          category: "HARM_CATEGORY_HARASSMENT",
          threshold: accessStore.googleSafetySettings,
        },
        {
          category: "HARM_CATEGORY_HATE_SPEECH",
          threshold: accessStore.googleSafetySettings,
        },
        {
          category: "HARM_CATEGORY_SEXUALLY_EXPLICIT",
          threshold: accessStore.googleSafetySettings,
        },
        {
          category: "HARM_CATEGORY_DANGEROUS_CONTENT",
          threshold: accessStore.googleSafetySettings,
        },
      ],
    };

    let shouldStream = !!options.config.stream;
    const controller = new AbortController();
    options.onController?.(controller);
    try {
      // https://github.com/google-gemini/cookbook/blob/main/quickstarts/rest/Streaming_REST.ipynb
      const chatPath = this.path(Google.ChatPath(modelConfig.model));

      const chatPayload = {
        method: "POST",
        body: JSON.stringify(requestPayload),
        signal: controller.signal,
        headers: getHeaders(),
      };

      // make a fetch request
      const requestTimeoutId = setTimeout(
        () => controller.abort(),
        REQUEST_TIMEOUT_MS,
      );

      if (shouldStream) {
        let responseText = "";
        let remainText = "";
        let finished = false;

        const finish = () => {
          if (!finished) {
            finished = true;
            options.onFinish(responseText + remainText);
          }
        };

        // animate response to make it looks smooth
        function animateResponseText() {
          if (finished || controller.signal.aborted) {
            responseText += remainText;
            finish();
            return;
          }

          if (remainText.length > 0) {
            const fetchCount = Math.max(1, Math.round(remainText.length / 60));
            const fetchText = remainText.slice(0, fetchCount);
            responseText += fetchText;
            remainText = remainText.slice(fetchCount);
            options.onUpdate?.(responseText, fetchText);
          }

          requestAnimationFrame(animateResponseText);
        }

        // start animaion
        animateResponseText();

        controller.signal.onabort = finish;

        fetchEventSource(chatPath, {
          fetch: fetch as any,
          ...chatPayload,
          async onopen(res) {
            clearTimeout(requestTimeoutId);
            const contentType = res.headers.get("content-type");
            console.log(
              "[Gemini] request response content type: ",
              contentType,
            );

            if (contentType?.startsWith("text/plain")) {
              responseText = await res.clone().text();
              return finish();
            }

            if (
              !res.ok ||
              !res.headers
                .get("content-type")
                ?.startsWith(EventStreamContentType) ||
              res.status !== 200
            ) {
              const responseTexts = [responseText];
              let extraInfo = await res.clone().text();
              try {
                const resJson = await res.clone().json();
                extraInfo = prettyObject(resJson);
              } catch {}

              if (res.status === 401) {
                responseTexts.push(Locale.Error.Unauthorized);
              }

              if (extraInfo) {
                responseTexts.push(extraInfo);
              }

              responseText = responseTexts.join("\n\n");

              return finish();
            }
          },
          onmessage(msg) {
            if (msg.data === "[DONE]" || finished) {
              return finish();
            }
            const text = msg.data;
            try {
              const json = JSON.parse(text);
              const delta = apiClient.extractMessage(json);

              if (delta) {
                remainText += delta;
              }

              const blockReason = json?.promptFeedback?.blockReason;
              if (blockReason) {
                // being blocked
                console.log(`[Google] [Safety Ratings] result:`, blockReason);
              }
            } catch (e) {
              console.error("[Request] parse error", text, msg);
            }
          },
          onclose() {
            finish();
          },
          onerror(e) {
            options.onError?.(e);
            throw e;
          },
          openWhenHidden: true,
        });
      } else {
        const res = await fetch(chatPath, chatPayload);
        clearTimeout(requestTimeoutId);
        const resJson = await res.json();
        if (resJson?.promptFeedback?.blockReason) {
          // being blocked
          options.onError?.(
            new Error(
              "Message is being blocked for reason: " +
                resJson.promptFeedback.blockReason,
            ),
          );
        }
        const message = apiClient.extractMessage(resJson);
        options.onFinish(message);
      }
    } catch (e) {
      console.log("[Request] failed to make a chat request", e);
      options.onError?.(e as Error);
    }
  }
  usage(): Promise<LLMUsage> {
    throw new Error("Method not implemented.");
  }
  async models(): Promise<LLMModel[]> {
    return [];
  }
}