mirror of
https://github.com/ChatGPTNextWeb/ChatGPT-Next-Web.git
synced 2026-02-19 21:04:25 +08:00
Compare commits
98 Commits
Leizhenpen
...
cb8144e168
| Author | SHA1 | Date | |
|---|---|---|---|
|
|
cb8144e168 | ||
|
|
00816304df | ||
|
|
645ecbcad8 | ||
|
|
65dc0e26d2 | ||
|
|
81b14b7b8d | ||
|
|
40c00374e7 | ||
|
|
e94566d258 | ||
|
|
6d72a04854 | ||
|
|
b0f78e9d1c | ||
|
|
29b9a20acf | ||
|
|
89b1774996 | ||
|
|
26f79aa6e6 | ||
|
|
92615dae32 | ||
|
|
9643adcf15 | ||
|
|
e839940a26 | ||
|
|
0ec1ae6276 | ||
|
|
cb0422b8f2 | ||
|
|
e455840ab3 | ||
|
|
19437c7aa1 | ||
|
|
12d38aa4b2 | ||
|
|
2a9f7d72fc | ||
|
|
372a327522 | ||
|
|
50a241b715 | ||
|
|
44a1cf6d6e | ||
|
|
a0886875e1 | ||
|
|
93337b2b92 | ||
|
|
4b2f447474 | ||
|
|
7830b37a90 | ||
|
|
fb3437ca01 | ||
|
|
f5ae086d3c | ||
|
|
26b9fa97cd | ||
|
|
603415f9e1 | ||
|
|
5ac651ad8e | ||
|
|
ad49cd0454 | ||
|
|
57dc44a54f | ||
|
|
4254fd34f9 | ||
|
|
0c55850641 | ||
|
|
a75b9f7fbe | ||
|
|
d9d2a27d8f | ||
|
|
471b17831a | ||
|
|
8ce2cf5c3b | ||
|
|
9c648e5b64 | ||
|
|
b39b3f7a5e | ||
|
|
448babd27f | ||
|
|
e6633753a4 | ||
|
|
15d0600642 | ||
|
|
9a473048b1 | ||
|
|
2fe848ecf3 | ||
|
|
0abfd279e4 | ||
|
|
5bd7e28f82 | ||
|
|
6f7a635030 | ||
|
|
2ccdd1706a | ||
|
|
a19ba6933a | ||
|
|
513cf1b206 | ||
|
|
238eb70986 | ||
|
|
a6337e9f23 | ||
|
|
ff88421904 | ||
|
|
a85db21e1f | ||
|
|
b0c1ccd0a0 | ||
|
|
bd68df1d9b | ||
|
|
f60c237b16 | ||
|
|
9d3f1d2529 | ||
|
|
dfeb9e7f27 | ||
|
|
b2d5e0e309 | ||
|
|
225ad30898 | ||
|
|
6bc1612720 | ||
|
|
24261d20b3 | ||
|
|
1b5a81c7ad | ||
|
|
70f066c15f | ||
|
|
bfa433919f | ||
|
|
afb0752d5d | ||
|
|
f12058463b | ||
|
|
09e4f95272 | ||
|
|
0e09697274 | ||
|
|
82a368a3aa | ||
|
|
9e04198bdc | ||
|
|
4204890d90 | ||
|
|
5d5456c1c5 | ||
|
|
f0c23cc6aa | ||
|
|
3bf55d3530 | ||
|
|
952d8835a3 | ||
|
|
c55cea5853 | ||
|
|
ca17e90c52 | ||
|
|
cae20af24d | ||
|
|
1f66d3779c | ||
|
|
045adc3567 | ||
|
|
1998cf5ced | ||
|
|
1164e1bdf6 | ||
|
|
d55c752e1e | ||
|
|
e3c18bb123 | ||
|
|
f532731e2a | ||
|
|
58837f6dec | ||
|
|
afbf5eb541 | ||
|
|
0f276f59bb | ||
|
|
fc391168e9 | ||
|
|
dca4a0e48f | ||
|
|
722c28839f | ||
|
|
ff356f0c8c |
@@ -76,6 +76,12 @@ ANTHROPIC_URL=
|
||||
### (optional)
|
||||
WHITE_WEBDAV_ENDPOINTS=
|
||||
|
||||
|
||||
### bedrock (optional)
|
||||
AWS_REGION=
|
||||
AWS_ACCESS_KEY=AKIA
|
||||
AWS_SECRET_KEY=
|
||||
|
||||
### siliconflow Api key (optional)
|
||||
SILICONFLOW_API_KEY=
|
||||
|
||||
|
||||
@@ -1,6 +1,7 @@
|
||||
import { ApiPath } from "@/app/constant";
|
||||
import { NextRequest } from "next/server";
|
||||
import { handle as openaiHandler } from "../../openai";
|
||||
import { handle as bedrockHandler } from "../../bedrock";
|
||||
import { handle as azureHandler } from "../../azure";
|
||||
import { handle as googleHandler } from "../../google";
|
||||
import { handle as anthropicHandler } from "../../anthropic";
|
||||
@@ -23,12 +24,15 @@ async function handle(
|
||||
const apiPath = `/api/${params.provider}`;
|
||||
console.log(`[${params.provider} Route] params `, params);
|
||||
switch (apiPath) {
|
||||
case ApiPath.Bedrock:
|
||||
return bedrockHandler(req, { params });
|
||||
case ApiPath.Azure:
|
||||
return azureHandler(req, { params });
|
||||
case ApiPath.Google:
|
||||
return googleHandler(req, { params });
|
||||
case ApiPath.Anthropic:
|
||||
return anthropicHandler(req, { params });
|
||||
|
||||
case ApiPath.Baidu:
|
||||
return baiduHandler(req, { params });
|
||||
case ApiPath.ByteDance:
|
||||
|
||||
@@ -52,7 +52,6 @@ export function auth(req: NextRequest, modelProvider: ModelProvider) {
|
||||
msg: "you are not allowed to access with your own api key",
|
||||
};
|
||||
}
|
||||
|
||||
// if user does not provide an api key, inject system api key
|
||||
if (!apiKey) {
|
||||
const serverConfig = getServerSideConfig();
|
||||
@@ -101,6 +100,14 @@ export function auth(req: NextRequest, modelProvider: ModelProvider) {
|
||||
case ModelProvider.ChatGLM:
|
||||
systemApiKey = serverConfig.chatglmApiKey;
|
||||
break;
|
||||
case ModelProvider.Bedrock:
|
||||
systemApiKey =
|
||||
serverConfig.awsRegion +
|
||||
":" +
|
||||
serverConfig.awsAccessKey +
|
||||
":" +
|
||||
serverConfig.awsSecretKey;
|
||||
break;
|
||||
case ModelProvider.SiliconFlow:
|
||||
systemApiKey = serverConfig.siliconFlowApiKey;
|
||||
break;
|
||||
|
||||
177
app/api/bedrock.ts
Normal file
177
app/api/bedrock.ts
Normal file
@@ -0,0 +1,177 @@
|
||||
import { NextRequest, NextResponse } from "next/server";
|
||||
import { auth } from "./auth";
|
||||
import {
|
||||
sign,
|
||||
decrypt,
|
||||
getBedrockEndpoint,
|
||||
BedrockCredentials,
|
||||
} from "../utils/aws";
|
||||
import { getServerSideConfig } from "../config/server";
|
||||
import { ModelProvider } from "../constant";
|
||||
import { prettyObject } from "../utils/format";
|
||||
|
||||
const ALLOWED_PATH = new Set(["chat", "models"]);
|
||||
|
||||
async function getBedrockCredentials(
|
||||
req: NextRequest,
|
||||
): Promise<BedrockCredentials> {
|
||||
// Get AWS credentials from server config first
|
||||
const config = getServerSideConfig();
|
||||
let awsRegion = config.awsRegion;
|
||||
let awsAccessKey = config.awsAccessKey;
|
||||
let awsSecretKey = config.awsSecretKey;
|
||||
|
||||
// If server-side credentials are not available, parse from Authorization header
|
||||
if (!awsRegion || !awsAccessKey || !awsSecretKey) {
|
||||
const authHeader = req.headers.get("Authorization");
|
||||
if (!authHeader || !authHeader.startsWith("Bearer ")) {
|
||||
throw new Error("Missing or invalid Authorization header");
|
||||
}
|
||||
|
||||
const [_, credentials] = authHeader.split("Bearer ");
|
||||
const [encryptedRegion, encryptedAccessKey, encryptedSecretKey] =
|
||||
credentials.split(":");
|
||||
|
||||
if (!encryptedRegion || !encryptedAccessKey || !encryptedSecretKey) {
|
||||
throw new Error("Invalid Authorization header format");
|
||||
}
|
||||
const encryptionKey = req.headers.get("XEncryptionKey") || "";
|
||||
// Decrypt the credentials
|
||||
[awsRegion, awsAccessKey, awsSecretKey] = await Promise.all([
|
||||
decrypt(encryptedRegion, encryptionKey),
|
||||
decrypt(encryptedAccessKey, encryptionKey),
|
||||
decrypt(encryptedSecretKey, encryptionKey),
|
||||
]);
|
||||
|
||||
if (!awsRegion || !awsAccessKey || !awsSecretKey) {
|
||||
throw new Error(
|
||||
"Failed to decrypt AWS credentials. Please ensure ENCRYPTION_KEY is set correctly.",
|
||||
);
|
||||
}
|
||||
}
|
||||
|
||||
return {
|
||||
region: awsRegion,
|
||||
accessKeyId: awsAccessKey,
|
||||
secretAccessKey: awsSecretKey,
|
||||
};
|
||||
}
|
||||
|
||||
async function requestBedrock(req: NextRequest) {
|
||||
const controller = new AbortController();
|
||||
const timeoutId = setTimeout(() => controller.abort(), 10 * 60 * 1000);
|
||||
|
||||
try {
|
||||
// Get credentials and model info
|
||||
const credentials = await getBedrockCredentials(req);
|
||||
const modelId = req.headers.get("XModelID");
|
||||
const shouldStream = req.headers.get("ShouldStream") !== "false";
|
||||
|
||||
if (!modelId) {
|
||||
throw new Error("Missing model ID");
|
||||
}
|
||||
|
||||
// Parse and validate request body
|
||||
const bodyText = await req.clone().text();
|
||||
if (!bodyText) {
|
||||
throw new Error("Request body is empty");
|
||||
}
|
||||
let bodyJson;
|
||||
try {
|
||||
bodyJson = JSON.parse(bodyText);
|
||||
} catch (e) {
|
||||
throw new Error(`Invalid JSON in request body: ${e}`);
|
||||
}
|
||||
console.log("[Bedrock Request] Initiating request");
|
||||
// Get endpoint and prepare request
|
||||
const endpoint = getBedrockEndpoint(
|
||||
credentials.region,
|
||||
modelId,
|
||||
shouldStream,
|
||||
);
|
||||
const requestBody: any = {
|
||||
...bodyJson,
|
||||
};
|
||||
// Sign request
|
||||
const headers = await sign({
|
||||
method: "POST",
|
||||
url: endpoint,
|
||||
region: credentials.region,
|
||||
accessKeyId: credentials.accessKeyId,
|
||||
secretAccessKey: credentials.secretAccessKey,
|
||||
body: JSON.stringify(requestBody),
|
||||
service: "bedrock",
|
||||
isStreaming: shouldStream,
|
||||
});
|
||||
|
||||
// Make request to AWS Bedrock
|
||||
// console.log(
|
||||
// "[Bedrock Request] Final Body:",
|
||||
// JSON.stringify(requestBody, null, 2),
|
||||
// );
|
||||
const res = await fetch(endpoint, {
|
||||
method: "POST",
|
||||
headers,
|
||||
body: JSON.stringify(requestBody),
|
||||
redirect: "manual",
|
||||
// @ts-ignore
|
||||
duplex: "half",
|
||||
signal: controller.signal,
|
||||
});
|
||||
|
||||
if (!res.ok) {
|
||||
const error = await res.text();
|
||||
console.error("[Bedrock Error] Request failed with status:", res.status);
|
||||
try {
|
||||
const errorJson = JSON.parse(error);
|
||||
throw new Error(errorJson.message || error);
|
||||
} catch {
|
||||
throw new Error(
|
||||
`Bedrock request failed with status ${res.status}: ${
|
||||
error || "No error message"
|
||||
}`,
|
||||
);
|
||||
}
|
||||
}
|
||||
|
||||
if (!res.body) {
|
||||
console.error("[Bedrock Error] Empty response body");
|
||||
throw new Error(
|
||||
"Empty response from Bedrock. Please check AWS credentials and permissions.",
|
||||
);
|
||||
}
|
||||
return res;
|
||||
} catch (e) {
|
||||
console.error("[Bedrock Request Error]:", e);
|
||||
throw e;
|
||||
} finally {
|
||||
clearTimeout(timeoutId);
|
||||
}
|
||||
}
|
||||
|
||||
export async function handle(
|
||||
req: NextRequest,
|
||||
{ params }: { params: { path: string[] } },
|
||||
) {
|
||||
const subpath = params.path.join("/");
|
||||
if (!ALLOWED_PATH.has(subpath)) {
|
||||
return NextResponse.json(
|
||||
{ error: true, msg: "you are not allowed to request " + subpath },
|
||||
{ status: 403 },
|
||||
);
|
||||
}
|
||||
|
||||
const authResult = auth(req, ModelProvider.Bedrock);
|
||||
if (authResult.error) {
|
||||
return NextResponse.json(authResult, {
|
||||
status: 401,
|
||||
});
|
||||
}
|
||||
|
||||
try {
|
||||
return await requestBedrock(req);
|
||||
} catch (e) {
|
||||
console.error("Handler error:", e);
|
||||
return NextResponse.json(prettyObject(e));
|
||||
}
|
||||
}
|
||||
@@ -23,8 +23,10 @@ import { SparkApi } from "./platforms/iflytek";
|
||||
import { DeepSeekApi } from "./platforms/deepseek";
|
||||
import { XAIApi } from "./platforms/xai";
|
||||
import { ChatGLMApi } from "./platforms/glm";
|
||||
import { BedrockApi } from "./platforms/bedrock";
|
||||
import { SiliconflowApi } from "./platforms/siliconflow";
|
||||
|
||||
|
||||
export const ROLES = ["system", "user", "assistant"] as const;
|
||||
export type MessageRole = (typeof ROLES)[number];
|
||||
|
||||
@@ -132,6 +134,9 @@ export class ClientApi {
|
||||
|
||||
constructor(provider: ModelProvider = ModelProvider.GPT) {
|
||||
switch (provider) {
|
||||
case ModelProvider.Bedrock:
|
||||
this.llm = new BedrockApi();
|
||||
break;
|
||||
case ModelProvider.GeminiPro:
|
||||
this.llm = new GeminiProApi();
|
||||
break;
|
||||
@@ -247,6 +252,7 @@ export function getHeaders(ignoreHeaders: boolean = false) {
|
||||
|
||||
function getConfig() {
|
||||
const modelConfig = chatStore.currentSession().mask.modelConfig;
|
||||
const isBedrock = modelConfig.providerName === ServiceProvider.Bedrock;
|
||||
const isGoogle = modelConfig.providerName === ServiceProvider.Google;
|
||||
const isAzure = modelConfig.providerName === ServiceProvider.Azure;
|
||||
const isAnthropic = modelConfig.providerName === ServiceProvider.Anthropic;
|
||||
@@ -287,6 +293,7 @@ export function getHeaders(ignoreHeaders: boolean = false) {
|
||||
: ""
|
||||
: accessStore.openaiApiKey;
|
||||
return {
|
||||
isBedrock,
|
||||
isGoogle,
|
||||
isAzure,
|
||||
isAnthropic,
|
||||
@@ -315,6 +322,7 @@ export function getHeaders(ignoreHeaders: boolean = false) {
|
||||
}
|
||||
|
||||
const {
|
||||
isBedrock,
|
||||
isGoogle,
|
||||
isAzure,
|
||||
isAnthropic,
|
||||
@@ -335,17 +343,23 @@ export function getHeaders(ignoreHeaders: boolean = false) {
|
||||
|
||||
const authHeader = getAuthHeader();
|
||||
|
||||
const bearerToken = getBearerToken(
|
||||
apiKey,
|
||||
isAzure || isAnthropic || isGoogle,
|
||||
);
|
||||
|
||||
if (bearerToken) {
|
||||
headers[authHeader] = bearerToken;
|
||||
} else if (isEnabledAccessControl && validString(accessStore.accessCode)) {
|
||||
headers["Authorization"] = getBearerToken(
|
||||
ACCESS_CODE_PREFIX + accessStore.accessCode,
|
||||
if (isBedrock) {
|
||||
if (apiKey) {
|
||||
headers[authHeader] = getBearerToken(apiKey);
|
||||
}
|
||||
} else {
|
||||
const bearerToken = getBearerToken(
|
||||
apiKey,
|
||||
isAzure || isAnthropic || isGoogle,
|
||||
);
|
||||
|
||||
if (bearerToken) {
|
||||
headers[authHeader] = bearerToken;
|
||||
} else if (isEnabledAccessControl && validString(accessStore.accessCode)) {
|
||||
headers["Authorization"] = getBearerToken(
|
||||
ACCESS_CODE_PREFIX + accessStore.accessCode,
|
||||
);
|
||||
}
|
||||
}
|
||||
|
||||
return headers;
|
||||
@@ -353,6 +367,8 @@ export function getHeaders(ignoreHeaders: boolean = false) {
|
||||
|
||||
export function getClientApi(provider: ServiceProvider): ClientApi {
|
||||
switch (provider) {
|
||||
case ServiceProvider.Bedrock:
|
||||
return new ClientApi(ModelProvider.Bedrock);
|
||||
case ServiceProvider.Google:
|
||||
return new ClientApi(ModelProvider.GeminiPro);
|
||||
case ServiceProvider.Anthropic:
|
||||
|
||||
859
app/client/platforms/bedrock.ts
Normal file
859
app/client/platforms/bedrock.ts
Normal file
@@ -0,0 +1,859 @@
|
||||
"use client";
|
||||
import { ChatOptions, getHeaders, LLMApi, SpeechOptions } from "../api";
|
||||
import {
|
||||
useAppConfig,
|
||||
usePluginStore,
|
||||
useChatStore,
|
||||
useAccessStore,
|
||||
ChatMessageTool,
|
||||
} from "@/app/store";
|
||||
import { preProcessImageContent } from "@/app/utils/chat";
|
||||
import { getMessageTextContent, isVisionModel } from "@/app/utils";
|
||||
import { ApiPath, BEDROCK_BASE_URL, REQUEST_TIMEOUT_MS } from "@/app/constant";
|
||||
import { getClientConfig } from "@/app/config/client";
|
||||
import {
|
||||
extractMessage,
|
||||
processMessage,
|
||||
processChunks,
|
||||
parseEventData,
|
||||
sign,
|
||||
} from "@/app/utils/aws";
|
||||
import { prettyObject } from "@/app/utils/format";
|
||||
import Locale from "@/app/locales";
|
||||
import { encrypt } from "@/app/utils/aws";
|
||||
|
||||
const ClaudeMapper = {
|
||||
assistant: "assistant",
|
||||
user: "user",
|
||||
system: "user",
|
||||
} as const;
|
||||
|
||||
const MistralMapper = {
|
||||
system: "system",
|
||||
user: "user",
|
||||
assistant: "assistant",
|
||||
} as const;
|
||||
type MistralRole = keyof typeof MistralMapper;
|
||||
|
||||
interface Tool {
|
||||
function?: {
|
||||
name?: string;
|
||||
description?: string;
|
||||
parameters?: any;
|
||||
};
|
||||
}
|
||||
const isApp = !!getClientConfig()?.isApp;
|
||||
// const isApp = true;
|
||||
async function getBedrockHeaders(
|
||||
modelId: string,
|
||||
chatPath: string,
|
||||
finalRequestBody: any,
|
||||
shouldStream: boolean,
|
||||
): Promise<Record<string, string>> {
|
||||
const accessStore = useAccessStore.getState();
|
||||
const bedrockHeaders = isApp
|
||||
? await sign({
|
||||
method: "POST",
|
||||
url: chatPath,
|
||||
region: accessStore.awsRegion,
|
||||
accessKeyId: accessStore.awsAccessKey,
|
||||
secretAccessKey: accessStore.awsSecretKey,
|
||||
body: finalRequestBody,
|
||||
service: "bedrock",
|
||||
headers: {},
|
||||
isStreaming: shouldStream,
|
||||
})
|
||||
: getHeaders();
|
||||
|
||||
if (!isApp) {
|
||||
const { awsRegion, awsAccessKey, awsSecretKey, encryptionKey } =
|
||||
accessStore;
|
||||
|
||||
const bedrockHeadersConfig = {
|
||||
XModelID: modelId,
|
||||
XEncryptionKey: encryptionKey,
|
||||
ShouldStream: String(shouldStream),
|
||||
Authorization: await createAuthHeader(
|
||||
awsRegion,
|
||||
awsAccessKey,
|
||||
awsSecretKey,
|
||||
encryptionKey,
|
||||
),
|
||||
};
|
||||
|
||||
Object.assign(bedrockHeaders, bedrockHeadersConfig);
|
||||
}
|
||||
|
||||
return bedrockHeaders;
|
||||
}
|
||||
|
||||
// Helper function to create Authorization header
|
||||
async function createAuthHeader(
|
||||
region: string,
|
||||
accessKey: string,
|
||||
secretKey: string,
|
||||
encryptionKey: string,
|
||||
): Promise<string> {
|
||||
const encryptedValues = await Promise.all([
|
||||
encrypt(region, encryptionKey),
|
||||
encrypt(accessKey, encryptionKey),
|
||||
encrypt(secretKey, encryptionKey),
|
||||
]);
|
||||
|
||||
return `Bearer ${encryptedValues.join(":")}`;
|
||||
}
|
||||
|
||||
export class BedrockApi implements LLMApi {
|
||||
speech(options: SpeechOptions): Promise<ArrayBuffer> {
|
||||
throw new Error("Speech not implemented for Bedrock.");
|
||||
}
|
||||
|
||||
formatRequestBody(messages: ChatOptions["messages"], modelConfig: any) {
|
||||
const model = modelConfig.model;
|
||||
const visionModel = isVisionModel(modelConfig.model);
|
||||
|
||||
// Get tools if available
|
||||
const [tools] = usePluginStore
|
||||
.getState()
|
||||
.getAsTools(useChatStore.getState().currentSession().mask?.plugin || []);
|
||||
|
||||
const toolsArray = (tools as Tool[]) || [];
|
||||
|
||||
// Handle Nova models
|
||||
if (model.includes("amazon.nova")) {
|
||||
// Extract system message if present
|
||||
const systemMessage = messages.find((m) => m.role === "system");
|
||||
const conversationMessages = messages.filter((m) => m.role !== "system");
|
||||
|
||||
const requestBody: any = {
|
||||
schemaVersion: "messages-v1",
|
||||
messages: conversationMessages.map((message) => {
|
||||
const content = Array.isArray(message.content)
|
||||
? message.content
|
||||
: [{ text: getMessageTextContent(message) }];
|
||||
|
||||
return {
|
||||
role: message.role,
|
||||
content: content.map((item: any) => {
|
||||
// Handle text content
|
||||
if (item.text || typeof item === "string") {
|
||||
return { text: item.text || item };
|
||||
}
|
||||
// Handle image content
|
||||
if (item.image_url?.url) {
|
||||
const { url = "" } = item.image_url;
|
||||
const colonIndex = url.indexOf(":");
|
||||
const semicolonIndex = url.indexOf(";");
|
||||
const comma = url.indexOf(",");
|
||||
|
||||
// Extract format from mime type
|
||||
const mimeType = url.slice(colonIndex + 1, semicolonIndex);
|
||||
const format = mimeType.split("/")[1];
|
||||
const data = url.slice(comma + 1);
|
||||
|
||||
return {
|
||||
image: {
|
||||
format,
|
||||
source: {
|
||||
bytes: data,
|
||||
},
|
||||
},
|
||||
};
|
||||
}
|
||||
return item;
|
||||
}),
|
||||
};
|
||||
}),
|
||||
inferenceConfig: {
|
||||
temperature: modelConfig.temperature || 0.7,
|
||||
top_p: modelConfig.top_p || 0.9,
|
||||
top_k: modelConfig.top_k || 50,
|
||||
max_new_tokens: modelConfig.max_tokens || 1000,
|
||||
stopSequences: modelConfig.stop || [],
|
||||
},
|
||||
};
|
||||
|
||||
// Add system message if present
|
||||
if (systemMessage) {
|
||||
requestBody.system = [
|
||||
{
|
||||
text: getMessageTextContent(systemMessage),
|
||||
},
|
||||
];
|
||||
}
|
||||
|
||||
// Add tools if available - exact Nova format
|
||||
if (toolsArray.length > 0) {
|
||||
requestBody.toolConfig = {
|
||||
tools: toolsArray.map((tool) => ({
|
||||
toolSpec: {
|
||||
name: tool?.function?.name || "",
|
||||
description: tool?.function?.description || "",
|
||||
inputSchema: {
|
||||
json: {
|
||||
type: "object",
|
||||
properties: tool?.function?.parameters?.properties || {},
|
||||
required: tool?.function?.parameters?.required || [],
|
||||
},
|
||||
},
|
||||
},
|
||||
})),
|
||||
toolChoice: { auto: {} },
|
||||
};
|
||||
}
|
||||
|
||||
return requestBody;
|
||||
}
|
||||
|
||||
// Handle Titan models
|
||||
if (model.startsWith("amazon.titan")) {
|
||||
const inputText = messages
|
||||
.map((message) => {
|
||||
return `${message.role}: ${getMessageTextContent(message)}`;
|
||||
})
|
||||
.join("\n\n");
|
||||
|
||||
return {
|
||||
inputText,
|
||||
textGenerationConfig: {
|
||||
maxTokenCount: modelConfig.max_tokens,
|
||||
temperature: modelConfig.temperature,
|
||||
stopSequences: [],
|
||||
},
|
||||
};
|
||||
}
|
||||
|
||||
// Handle LLaMA models
|
||||
if (model.includes("meta.llama")) {
|
||||
let prompt = "<|begin_of_text|>";
|
||||
|
||||
// Extract system message if present
|
||||
const systemMessage = messages.find((m) => m.role === "system");
|
||||
if (systemMessage) {
|
||||
prompt += `<|start_header_id|>system<|end_header_id|>\n${getMessageTextContent(
|
||||
systemMessage,
|
||||
)}<|eot_id|>`;
|
||||
}
|
||||
|
||||
// Format the conversation
|
||||
const conversationMessages = messages.filter((m) => m.role !== "system");
|
||||
for (const message of conversationMessages) {
|
||||
const role = message.role === "assistant" ? "assistant" : "user";
|
||||
const content = getMessageTextContent(message);
|
||||
prompt += `<|start_header_id|>${role}<|end_header_id|>\n${content}<|eot_id|>`;
|
||||
}
|
||||
|
||||
// Add the final assistant header to prompt completion
|
||||
prompt += "<|start_header_id|>assistant<|end_header_id|>";
|
||||
|
||||
return {
|
||||
prompt,
|
||||
max_gen_len: modelConfig.max_tokens || 512,
|
||||
temperature: modelConfig.temperature || 0.7,
|
||||
top_p: modelConfig.top_p || 0.9,
|
||||
};
|
||||
}
|
||||
|
||||
// Handle Mistral models
|
||||
if (model.includes("mistral.mistral")) {
|
||||
const formattedMessages = messages.map((message) => ({
|
||||
role: MistralMapper[message.role as MistralRole] || "user",
|
||||
content: getMessageTextContent(message),
|
||||
}));
|
||||
|
||||
const requestBody: any = {
|
||||
messages: formattedMessages,
|
||||
max_tokens: modelConfig.max_tokens || 4096,
|
||||
temperature: modelConfig.temperature || 0.7,
|
||||
top_p: modelConfig.top_p || 0.9,
|
||||
};
|
||||
|
||||
// Add tools if available
|
||||
if (toolsArray.length > 0) {
|
||||
requestBody.tool_choice = "auto";
|
||||
requestBody.tools = toolsArray.map((tool) => ({
|
||||
type: "function",
|
||||
function: {
|
||||
name: tool?.function?.name,
|
||||
description: tool?.function?.description,
|
||||
parameters: tool?.function?.parameters,
|
||||
},
|
||||
}));
|
||||
}
|
||||
|
||||
return requestBody;
|
||||
}
|
||||
|
||||
// Handle Claude models
|
||||
const keys = ["system", "user"];
|
||||
// roles must alternate between "user" and "assistant" in claude, so add a fake assistant message between two user messages
|
||||
for (let i = 0; i < messages.length - 1; i++) {
|
||||
const message = messages[i];
|
||||
const nextMessage = messages[i + 1];
|
||||
|
||||
if (keys.includes(message.role) && keys.includes(nextMessage.role)) {
|
||||
messages[i] = [
|
||||
message,
|
||||
{
|
||||
role: "assistant",
|
||||
content: ";",
|
||||
},
|
||||
] as any;
|
||||
}
|
||||
}
|
||||
const prompt = messages
|
||||
.flat()
|
||||
.filter((v) => {
|
||||
if (!v.content) return false;
|
||||
if (typeof v.content === "string" && !v.content.trim()) return false;
|
||||
return true;
|
||||
})
|
||||
.map((v) => {
|
||||
const { role, content } = v;
|
||||
const insideRole = ClaudeMapper[role] ?? "user";
|
||||
|
||||
if (!visionModel || typeof content === "string") {
|
||||
return {
|
||||
role: insideRole,
|
||||
content: getMessageTextContent(v),
|
||||
};
|
||||
}
|
||||
return {
|
||||
role: insideRole,
|
||||
content: content
|
||||
.filter((v) => v.image_url || v.text)
|
||||
.map(({ type, text, image_url }) => {
|
||||
if (type === "text") {
|
||||
return {
|
||||
type,
|
||||
text: text!,
|
||||
};
|
||||
}
|
||||
const { url = "" } = image_url || {};
|
||||
const colonIndex = url.indexOf(":");
|
||||
const semicolonIndex = url.indexOf(";");
|
||||
const comma = url.indexOf(",");
|
||||
|
||||
const mimeType = url.slice(colonIndex + 1, semicolonIndex);
|
||||
const encodeType = url.slice(semicolonIndex + 1, comma);
|
||||
const data = url.slice(comma + 1);
|
||||
|
||||
return {
|
||||
type: "image" as const,
|
||||
source: {
|
||||
type: encodeType,
|
||||
media_type: mimeType,
|
||||
data,
|
||||
},
|
||||
};
|
||||
}),
|
||||
};
|
||||
});
|
||||
|
||||
if (prompt[0]?.role === "assistant") {
|
||||
prompt.unshift({
|
||||
role: "user",
|
||||
content: ";",
|
||||
});
|
||||
}
|
||||
const requestBody: any = {
|
||||
anthropic_version: useAccessStore.getState().bedrockAnthropicVersion,
|
||||
max_tokens: modelConfig.max_tokens,
|
||||
messages: prompt,
|
||||
temperature: modelConfig.temperature,
|
||||
top_p: modelConfig.top_p || 0.9,
|
||||
top_k: modelConfig.top_k || 5,
|
||||
};
|
||||
|
||||
// Add tools if available for Claude models
|
||||
if (toolsArray.length > 0 && model.includes("anthropic.claude")) {
|
||||
requestBody.tools = toolsArray.map((tool) => ({
|
||||
name: tool?.function?.name || "",
|
||||
description: tool?.function?.description || "",
|
||||
input_schema: tool?.function?.parameters || {},
|
||||
}));
|
||||
}
|
||||
|
||||
return requestBody;
|
||||
}
|
||||
|
||||
async chat(options: ChatOptions) {
|
||||
const accessStore = useAccessStore.getState();
|
||||
|
||||
const shouldStream = !!options.config.stream;
|
||||
|
||||
const modelConfig = {
|
||||
...useAppConfig.getState().modelConfig,
|
||||
...useChatStore.getState().currentSession().mask.modelConfig,
|
||||
...{
|
||||
model: options.config.model,
|
||||
},
|
||||
};
|
||||
|
||||
// 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 controller = new AbortController();
|
||||
options.onController?.(controller);
|
||||
|
||||
let finalRequestBody = this.formatRequestBody(messages, modelConfig);
|
||||
|
||||
try {
|
||||
const bedrockAPIPath = `${BEDROCK_BASE_URL}/model/${
|
||||
modelConfig.model
|
||||
}/invoke${shouldStream ? "-with-response-stream" : ""}`;
|
||||
const chatPath = isApp ? bedrockAPIPath : ApiPath.Bedrock + "/chat";
|
||||
|
||||
if (process.env.NODE_ENV !== "production") {
|
||||
console.debug("[Bedrock Client] Request:", {
|
||||
path: chatPath,
|
||||
model: modelConfig.model,
|
||||
messages: messages.length,
|
||||
stream: shouldStream,
|
||||
});
|
||||
}
|
||||
|
||||
if (shouldStream) {
|
||||
const [tools, funcs] = usePluginStore
|
||||
.getState()
|
||||
.getAsTools(
|
||||
useChatStore.getState().currentSession().mask?.plugin || [],
|
||||
);
|
||||
return bedrockStream(
|
||||
modelConfig.model,
|
||||
chatPath,
|
||||
finalRequestBody,
|
||||
funcs,
|
||||
controller,
|
||||
// processToolMessage, include tool_calls message and tool call results
|
||||
(
|
||||
requestPayload: any[],
|
||||
toolCallMessage: any,
|
||||
toolCallResult: any[],
|
||||
) => {
|
||||
const modelId = modelConfig.model;
|
||||
const isMistral = modelId.includes("mistral.mistral");
|
||||
const isClaude = modelId.includes("anthropic.claude");
|
||||
const isNova = modelId.includes("amazon.nova");
|
||||
|
||||
if (isClaude) {
|
||||
// Format for Claude
|
||||
// @ts-ignore
|
||||
requestPayload?.messages?.splice(
|
||||
// @ts-ignore
|
||||
requestPayload?.messages?.length,
|
||||
0,
|
||||
{
|
||||
role: "assistant",
|
||||
content: toolCallMessage.tool_calls.map(
|
||||
(tool: ChatMessageTool) => ({
|
||||
type: "tool_use",
|
||||
id: tool.id,
|
||||
name: tool?.function?.name,
|
||||
input: tool?.function?.arguments
|
||||
? JSON.parse(tool?.function?.arguments)
|
||||
: {},
|
||||
}),
|
||||
),
|
||||
},
|
||||
// @ts-ignore
|
||||
...toolCallResult.map((result) => ({
|
||||
role: "user",
|
||||
content: [
|
||||
{
|
||||
type: "tool_result",
|
||||
tool_use_id: result.tool_call_id,
|
||||
content: result.content,
|
||||
},
|
||||
],
|
||||
})),
|
||||
);
|
||||
} else if (isMistral) {
|
||||
// Format for Mistral
|
||||
// @ts-ignore
|
||||
requestPayload?.messages?.splice(
|
||||
// @ts-ignore
|
||||
requestPayload?.messages?.length,
|
||||
0,
|
||||
{
|
||||
role: "assistant",
|
||||
content: "",
|
||||
// @ts-ignore
|
||||
tool_calls: toolCallMessage.tool_calls.map(
|
||||
(tool: ChatMessageTool) => ({
|
||||
id: tool.id,
|
||||
function: {
|
||||
name: tool?.function?.name,
|
||||
arguments: tool?.function?.arguments || "{}",
|
||||
},
|
||||
}),
|
||||
),
|
||||
},
|
||||
...toolCallResult.map((result) => ({
|
||||
role: "tool",
|
||||
tool_call_id: result.tool_call_id,
|
||||
content: result.content,
|
||||
})),
|
||||
);
|
||||
} else if (isNova) {
|
||||
// Format for Nova - Updated format
|
||||
// @ts-ignore
|
||||
requestPayload?.messages?.splice(
|
||||
// @ts-ignore
|
||||
requestPayload?.messages?.length,
|
||||
0,
|
||||
{
|
||||
role: "assistant",
|
||||
content: [
|
||||
{
|
||||
toolUse: {
|
||||
toolUseId: toolCallMessage.tool_calls[0].id,
|
||||
name: toolCallMessage.tool_calls[0]?.function?.name,
|
||||
input:
|
||||
typeof toolCallMessage.tool_calls[0]?.function
|
||||
?.arguments === "string"
|
||||
? JSON.parse(
|
||||
toolCallMessage.tool_calls[0]?.function
|
||||
?.arguments,
|
||||
)
|
||||
: toolCallMessage.tool_calls[0]?.function
|
||||
?.arguments || {},
|
||||
},
|
||||
},
|
||||
],
|
||||
},
|
||||
{
|
||||
role: "user",
|
||||
content: [
|
||||
{
|
||||
toolResult: {
|
||||
toolUseId: toolCallResult[0].tool_call_id,
|
||||
content: [
|
||||
{
|
||||
json: {
|
||||
content: toolCallResult[0].content,
|
||||
},
|
||||
},
|
||||
],
|
||||
},
|
||||
},
|
||||
],
|
||||
},
|
||||
);
|
||||
} else {
|
||||
console.warn(
|
||||
`[Bedrock Client] Unhandled model type for tool calls: ${modelId}`,
|
||||
);
|
||||
}
|
||||
},
|
||||
options,
|
||||
);
|
||||
} else {
|
||||
try {
|
||||
controller.signal.onabort = () =>
|
||||
options.onFinish("", new Response(null, { status: 400 }));
|
||||
const newHeaders = await getBedrockHeaders(
|
||||
modelConfig.model,
|
||||
chatPath,
|
||||
JSON.stringify(finalRequestBody),
|
||||
shouldStream,
|
||||
);
|
||||
const res = await fetch(chatPath, {
|
||||
method: "POST",
|
||||
headers: newHeaders,
|
||||
body: JSON.stringify(finalRequestBody),
|
||||
});
|
||||
const contentType = res.headers.get("content-type");
|
||||
console.log(
|
||||
"[Bedrock Not Stream Request] response content type: ",
|
||||
contentType,
|
||||
);
|
||||
const resJson = await res.json();
|
||||
const message = extractMessage(resJson);
|
||||
options.onFinish(message, res);
|
||||
} catch (e) {
|
||||
const error =
|
||||
e instanceof Error ? e : new Error("Unknown error occurred");
|
||||
console.error("[Bedrock Client] Chat failed:", error.message);
|
||||
options.onError?.(error);
|
||||
}
|
||||
}
|
||||
} catch (e) {
|
||||
console.error("[Bedrock Client] Chat error:", e);
|
||||
options.onError?.(e as Error);
|
||||
}
|
||||
}
|
||||
|
||||
async usage() {
|
||||
return { used: 0, total: 0 };
|
||||
}
|
||||
|
||||
async models() {
|
||||
return [];
|
||||
}
|
||||
}
|
||||
|
||||
function bedrockStream(
|
||||
modelId: string,
|
||||
chatPath: string,
|
||||
requestPayload: any,
|
||||
funcs: Record<string, Function>,
|
||||
controller: AbortController,
|
||||
processToolMessage: (
|
||||
requestPayload: any,
|
||||
toolCallMessage: any,
|
||||
toolCallResult: any[],
|
||||
) => void,
|
||||
options: any,
|
||||
) {
|
||||
let responseText = "";
|
||||
let remainText = "";
|
||||
let finished = false;
|
||||
let running = false;
|
||||
let runTools: any[] = [];
|
||||
let responseRes: Response;
|
||||
let index = -1;
|
||||
let chunks: Uint8Array[] = [];
|
||||
let pendingChunk: Uint8Array | null = null;
|
||||
|
||||
function animateResponseText() {
|
||||
if (finished || controller.signal.aborted) {
|
||||
responseText += remainText;
|
||||
console.log("[Response Animation] finished");
|
||||
if (responseText?.length === 0) {
|
||||
options.onError?.(new Error("empty response from server"));
|
||||
}
|
||||
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);
|
||||
}
|
||||
|
||||
animateResponseText();
|
||||
|
||||
const finish = () => {
|
||||
if (!finished) {
|
||||
if (!running && runTools.length > 0) {
|
||||
const toolCallMessage = {
|
||||
role: "assistant",
|
||||
tool_calls: [...runTools],
|
||||
};
|
||||
running = true;
|
||||
runTools.splice(0, runTools.length);
|
||||
return Promise.all(
|
||||
toolCallMessage.tool_calls.map((tool) => {
|
||||
options?.onBeforeTool?.(tool);
|
||||
const funcName = tool?.function?.name || tool?.name;
|
||||
if (!funcName || !funcs[funcName]) {
|
||||
console.error(`Function ${funcName} not found in funcs:`, funcs);
|
||||
return Promise.reject(`Function ${funcName} not found`);
|
||||
}
|
||||
return Promise.resolve(
|
||||
funcs[funcName](
|
||||
tool?.function?.arguments
|
||||
? JSON.parse(tool?.function?.arguments)
|
||||
: {},
|
||||
),
|
||||
)
|
||||
.then((res) => {
|
||||
let content = res.data || res?.statusText;
|
||||
content =
|
||||
typeof content === "string"
|
||||
? content
|
||||
: JSON.stringify(content);
|
||||
if (res.status >= 300) {
|
||||
return Promise.reject(content);
|
||||
}
|
||||
return content;
|
||||
})
|
||||
.then((content) => {
|
||||
options?.onAfterTool?.({
|
||||
...tool,
|
||||
content,
|
||||
isError: false,
|
||||
});
|
||||
return content;
|
||||
})
|
||||
.catch((e) => {
|
||||
options?.onAfterTool?.({
|
||||
...tool,
|
||||
isError: true,
|
||||
errorMsg: e.toString(),
|
||||
});
|
||||
return e.toString();
|
||||
})
|
||||
.then((content) => ({
|
||||
name: funcName,
|
||||
role: "tool",
|
||||
content,
|
||||
tool_call_id: tool.id,
|
||||
}));
|
||||
}),
|
||||
).then((toolCallResult) => {
|
||||
processToolMessage(requestPayload, toolCallMessage, toolCallResult);
|
||||
setTimeout(() => {
|
||||
console.debug("[BedrockAPI for toolCallResult] restart");
|
||||
running = false;
|
||||
bedrockChatApi(modelId, chatPath, requestPayload, true);
|
||||
}, 60);
|
||||
});
|
||||
}
|
||||
if (running) {
|
||||
return;
|
||||
}
|
||||
console.debug("[BedrockAPI] end");
|
||||
finished = true;
|
||||
options.onFinish(responseText + remainText, responseRes);
|
||||
}
|
||||
};
|
||||
|
||||
controller.signal.onabort = finish;
|
||||
|
||||
async function bedrockChatApi(
|
||||
modelId: string,
|
||||
chatPath: string,
|
||||
requestPayload: any,
|
||||
shouldStream: boolean,
|
||||
) {
|
||||
const requestTimeoutId = setTimeout(
|
||||
() => controller.abort(),
|
||||
REQUEST_TIMEOUT_MS,
|
||||
);
|
||||
|
||||
const newHeaders = await getBedrockHeaders(
|
||||
modelId,
|
||||
chatPath,
|
||||
JSON.stringify(requestPayload),
|
||||
shouldStream,
|
||||
);
|
||||
try {
|
||||
const res = await fetch(chatPath, {
|
||||
method: "POST",
|
||||
headers: newHeaders,
|
||||
body: JSON.stringify(requestPayload),
|
||||
redirect: "manual",
|
||||
// @ts-ignore
|
||||
duplex: "half",
|
||||
signal: controller.signal,
|
||||
});
|
||||
|
||||
clearTimeout(requestTimeoutId);
|
||||
responseRes = res;
|
||||
|
||||
const contentType = res.headers.get("content-type");
|
||||
// console.log(
|
||||
// "[Bedrock Stream Request] response content type: ",
|
||||
// contentType,
|
||||
// );
|
||||
|
||||
if (contentType?.startsWith("text/plain")) {
|
||||
responseText = await res.text();
|
||||
return finish();
|
||||
}
|
||||
|
||||
if (
|
||||
!res.ok ||
|
||||
res.status !== 200 ||
|
||||
!contentType?.startsWith("application/vnd.amazon.eventstream")
|
||||
) {
|
||||
const responseTexts = [responseText];
|
||||
let extraInfo = await res.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();
|
||||
}
|
||||
|
||||
const reader = res.body?.getReader();
|
||||
if (!reader) {
|
||||
throw new Error("No response body reader available");
|
||||
}
|
||||
|
||||
try {
|
||||
while (true) {
|
||||
const { done, value } = await reader.read();
|
||||
if (done) {
|
||||
if (pendingChunk) {
|
||||
try {
|
||||
const parsed = parseEventData(pendingChunk);
|
||||
if (parsed) {
|
||||
const result = processMessage(
|
||||
parsed,
|
||||
remainText,
|
||||
runTools,
|
||||
index,
|
||||
);
|
||||
remainText = result.remainText;
|
||||
index = result.index;
|
||||
}
|
||||
} catch (e) {
|
||||
console.error("[Final Chunk Process Error]:", e);
|
||||
}
|
||||
}
|
||||
break;
|
||||
}
|
||||
|
||||
chunks.push(value);
|
||||
|
||||
const result = processChunks(
|
||||
chunks,
|
||||
pendingChunk,
|
||||
remainText,
|
||||
runTools,
|
||||
index,
|
||||
);
|
||||
chunks = result.chunks;
|
||||
pendingChunk = result.pendingChunk;
|
||||
remainText = result.remainText;
|
||||
index = result.index;
|
||||
}
|
||||
} catch (err) {
|
||||
console.error(
|
||||
"[Bedrock Stream]:",
|
||||
err instanceof Error ? err.message : "Stream processing failed",
|
||||
);
|
||||
throw new Error("Failed to process stream response");
|
||||
} finally {
|
||||
reader.releaseLock();
|
||||
finish();
|
||||
}
|
||||
} catch (e) {
|
||||
if (e instanceof Error && e.name === "AbortError") {
|
||||
console.log("[Bedrock Client] Aborted by user");
|
||||
return;
|
||||
}
|
||||
console.error(
|
||||
"[Bedrock Request] Failed:",
|
||||
e instanceof Error ? e.message : "Request failed",
|
||||
);
|
||||
options.onError?.(e);
|
||||
throw new Error("Request processing failed");
|
||||
}
|
||||
}
|
||||
|
||||
console.debug("[BedrockAPI] start");
|
||||
bedrockChatApi(modelId, chatPath, requestPayload, true);
|
||||
}
|
||||
@@ -750,4 +750,4 @@
|
||||
transform: translateX(0);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
@@ -1,5 +1,4 @@
|
||||
import { useState, useEffect, useMemo } from "react";
|
||||
|
||||
import styles from "./settings.module.scss";
|
||||
|
||||
import ResetIcon from "../icons/reload.svg";
|
||||
@@ -967,7 +966,89 @@ export function Settings() {
|
||||
</ListItem>
|
||||
</>
|
||||
);
|
||||
|
||||
const bedrockConfigComponent = accessStore.provider ===
|
||||
ServiceProvider.Bedrock && (
|
||||
<>
|
||||
<ListItem
|
||||
title={Locale.Settings.Access.Bedrock.Region.Title}
|
||||
subTitle={Locale.Settings.Access.Bedrock.Region.SubTitle}
|
||||
>
|
||||
<input
|
||||
aria-label={Locale.Settings.Access.Bedrock.Region.Title}
|
||||
type="text"
|
||||
value={accessStore.awsRegion}
|
||||
placeholder="us-west-2"
|
||||
onChange={(e) =>
|
||||
accessStore.update((access) => {
|
||||
const region = e.currentTarget.value;
|
||||
access.awsRegion = region;
|
||||
})
|
||||
}
|
||||
/>
|
||||
</ListItem>
|
||||
<ListItem
|
||||
title={Locale.Settings.Access.Bedrock.AccessKey.Title}
|
||||
subTitle={Locale.Settings.Access.Bedrock.AccessKey.SubTitle}
|
||||
>
|
||||
<PasswordInput
|
||||
aria-label={Locale.Settings.Access.Bedrock.AccessKey.Title}
|
||||
value={accessStore.awsAccessKey}
|
||||
type="text"
|
||||
placeholder={Locale.Settings.Access.Bedrock.AccessKey.Placeholder}
|
||||
onChange={(e) => {
|
||||
accessStore.update((access) => {
|
||||
const accessKey = e.currentTarget.value;
|
||||
access.awsAccessKey = accessKey;
|
||||
});
|
||||
}}
|
||||
maskWhenShow={true}
|
||||
/>
|
||||
</ListItem>
|
||||
<ListItem
|
||||
title={Locale.Settings.Access.Bedrock.SecretKey.Title}
|
||||
subTitle={Locale.Settings.Access.Bedrock.SecretKey.SubTitle}
|
||||
>
|
||||
<PasswordInput
|
||||
aria-label={Locale.Settings.Access.Bedrock.SecretKey.Title}
|
||||
value={accessStore.awsSecretKey}
|
||||
type="text"
|
||||
placeholder={Locale.Settings.Access.Bedrock.SecretKey.Placeholder}
|
||||
onChange={(e) => {
|
||||
accessStore.update((access) => {
|
||||
const secretKey = e.currentTarget.value;
|
||||
access.awsSecretKey = secretKey;
|
||||
});
|
||||
}}
|
||||
maskWhenShow={true}
|
||||
/>
|
||||
</ListItem>
|
||||
<ListItem
|
||||
title={Locale.Settings.Access.Bedrock.EncryptionKey.Title}
|
||||
subTitle={Locale.Settings.Access.Bedrock.EncryptionKey.SubTitle}
|
||||
>
|
||||
<PasswordInput
|
||||
aria-label={Locale.Settings.Access.Bedrock.EncryptionKey.Title}
|
||||
value={accessStore.encryptionKey}
|
||||
type="text"
|
||||
placeholder={Locale.Settings.Access.Bedrock.EncryptionKey.Placeholder}
|
||||
onChange={(e) => {
|
||||
accessStore.update(
|
||||
(access) => (access.encryptionKey = e.currentTarget.value),
|
||||
);
|
||||
}}
|
||||
onBlur={(e) => {
|
||||
const value = e.currentTarget.value;
|
||||
if (!value || value.length < 8) {
|
||||
showToast(Locale.Settings.Access.Bedrock.EncryptionKey.Invalid);
|
||||
accessStore.update((access) => (access.encryptionKey = ""));
|
||||
return;
|
||||
}
|
||||
}}
|
||||
maskWhenShow={true}
|
||||
/>
|
||||
</ListItem>
|
||||
</>
|
||||
);
|
||||
const baiduConfigComponent = accessStore.provider ===
|
||||
ServiceProvider.Baidu && (
|
||||
<>
|
||||
@@ -1808,6 +1889,7 @@ export function Settings() {
|
||||
</ListItem>
|
||||
|
||||
{openAIConfigComponent}
|
||||
{bedrockConfigComponent}
|
||||
{azureConfigComponent}
|
||||
{googleConfigComponent}
|
||||
{anthropicConfigComponent}
|
||||
|
||||
@@ -11,6 +11,7 @@ import MaxIcon from "../icons/max.svg";
|
||||
import MinIcon from "../icons/min.svg";
|
||||
|
||||
import Locale from "../locales";
|
||||
import { maskSensitiveValue } from "../utils/aws";
|
||||
|
||||
import { createRoot } from "react-dom/client";
|
||||
import React, {
|
||||
@@ -270,13 +271,25 @@ export function Input(props: InputProps) {
|
||||
}
|
||||
|
||||
export function PasswordInput(
|
||||
props: HTMLProps<HTMLInputElement> & { aria?: string },
|
||||
props: HTMLProps<HTMLInputElement> & {
|
||||
aria?: string;
|
||||
maskWhenShow?: boolean;
|
||||
},
|
||||
) {
|
||||
const [visible, setVisible] = useState(false);
|
||||
const [isEditing, setIsEditing] = useState(false);
|
||||
const { maskWhenShow, onChange, value, ...inputProps } = props;
|
||||
|
||||
function changeVisibility() {
|
||||
setVisible(!visible);
|
||||
}
|
||||
|
||||
// Get display value - use masked value only when showing and maskWhenShow is true and not editing
|
||||
const displayValue =
|
||||
maskWhenShow && visible && value && !isEditing
|
||||
? maskSensitiveValue(value as string)
|
||||
: value;
|
||||
|
||||
return (
|
||||
<div className={"password-input-container"}>
|
||||
<IconButton
|
||||
@@ -286,7 +299,11 @@ export function PasswordInput(
|
||||
className={"password-eye"}
|
||||
/>
|
||||
<input
|
||||
{...props}
|
||||
{...inputProps}
|
||||
value={displayValue}
|
||||
onChange={onChange}
|
||||
onFocus={() => setIsEditing(true)}
|
||||
onBlur={() => setIsEditing(false)}
|
||||
type={visible ? "text" : "password"}
|
||||
className={"password-input"}
|
||||
/>
|
||||
@@ -552,6 +569,7 @@ export function Selector<T>(props: {
|
||||
</div>
|
||||
);
|
||||
}
|
||||
|
||||
export function FullScreen(props: any) {
|
||||
const { children, right = 10, top = 10, ...rest } = props;
|
||||
const ref = useRef<HTMLDivElement>();
|
||||
|
||||
@@ -13,6 +13,12 @@ declare global {
|
||||
BASE_URL?: string;
|
||||
OPENAI_ORG_ID?: string; // openai only
|
||||
|
||||
// bedrock only
|
||||
AWS_REGION?: string;
|
||||
AWS_ACCESS_KEY?: string;
|
||||
AWS_SECRET_KEY?: string;
|
||||
ENCRYPTION_KEY?: string;
|
||||
|
||||
VERCEL?: string;
|
||||
BUILD_MODE?: "standalone" | "export";
|
||||
BUILD_APP?: string; // is building desktop app
|
||||
@@ -148,7 +154,10 @@ export const getServerSideConfig = () => {
|
||||
}
|
||||
|
||||
const isStability = !!process.env.STABILITY_API_KEY;
|
||||
|
||||
const isBedrock =
|
||||
!!process.env.AWS_REGION &&
|
||||
!!process.env.AWS_ACCESS_KEY &&
|
||||
!!process.env.AWS_SECRET_KEY;
|
||||
const isAzure = !!process.env.AZURE_URL;
|
||||
const isGoogle = !!process.env.GOOGLE_API_KEY;
|
||||
const isAnthropic = !!process.env.ANTHROPIC_API_KEY;
|
||||
@@ -180,6 +189,12 @@ export const getServerSideConfig = () => {
|
||||
apiKey: getApiKey(process.env.OPENAI_API_KEY),
|
||||
openaiOrgId: process.env.OPENAI_ORG_ID,
|
||||
|
||||
isBedrock,
|
||||
awsRegion: process.env.AWS_REGION,
|
||||
awsAccessKey: process.env.AWS_ACCESS_KEY,
|
||||
awsSecretKey: process.env.AWS_SECRET_KEY,
|
||||
encryptionKey: process.env.ENCRYPTION_KEY,
|
||||
|
||||
isStability,
|
||||
stabilityUrl: process.env.STABILITY_URL,
|
||||
stabilityApiKey: getApiKey(process.env.STABILITY_API_KEY),
|
||||
|
||||
@@ -56,6 +56,7 @@ export enum Path {
|
||||
|
||||
export enum ApiPath {
|
||||
Cors = "",
|
||||
Bedrock = "/api/bedrock",
|
||||
Azure = "/api/azure",
|
||||
OpenAI = "/api/openai",
|
||||
Anthropic = "/api/anthropic",
|
||||
@@ -129,6 +130,7 @@ export enum ServiceProvider {
|
||||
XAI = "XAI",
|
||||
ChatGLM = "ChatGLM",
|
||||
DeepSeek = "DeepSeek",
|
||||
Bedrock = "Bedrock",
|
||||
SiliconFlow = "SiliconFlow",
|
||||
}
|
||||
|
||||
@@ -155,6 +157,7 @@ export enum ModelProvider {
|
||||
XAI = "XAI",
|
||||
ChatGLM = "ChatGLM",
|
||||
DeepSeek = "DeepSeek",
|
||||
Bedrock = "Bedrock",
|
||||
SiliconFlow = "SiliconFlow",
|
||||
}
|
||||
|
||||
@@ -255,6 +258,15 @@ export const ChatGLM = {
|
||||
VideoPath: "api/paas/v4/videos/generations",
|
||||
};
|
||||
|
||||
export const Bedrock = {
|
||||
ChatPath: "model", // Simplified path since we'll append the full path in bedrock.ts
|
||||
ApiVersion: "2023-11-01",
|
||||
getEndpoint: (region: string = "us-west-2") =>
|
||||
`https://bedrock-runtime.${region}.amazonaws.com`,
|
||||
};
|
||||
// Get the region from access store for BEDROCK_BASE_URL
|
||||
export const BEDROCK_BASE_URL = Bedrock.getEndpoint();
|
||||
|
||||
export const SiliconFlow = {
|
||||
ExampleEndpoint: SILICONFLOW_BASE_URL,
|
||||
ChatPath: "v1/chat/completions",
|
||||
@@ -463,6 +475,8 @@ export const VISION_MODEL_REGEXES = [
|
||||
/gpt-4-turbo(?!.*preview)/, // Matches "gpt-4-turbo" but not "gpt-4-turbo-preview"
|
||||
/^dall-e-3$/, // Matches exactly "dall-e-3"
|
||||
/glm-4v/,
|
||||
/nova-lite/,
|
||||
/nova-pro/,
|
||||
/vl/i,
|
||||
];
|
||||
|
||||
@@ -496,6 +510,28 @@ const openaiModels = [
|
||||
"o3-mini",
|
||||
];
|
||||
|
||||
const bedrockModels = [
|
||||
// Amazon nova Models
|
||||
"us.amazon.nova-micro-v1:0",
|
||||
"us.amazon.nova-lite-v1:0",
|
||||
"us.amazon.nova-pro-v1:0",
|
||||
// Claude Models
|
||||
"anthropic.claude-3-haiku-20240307-v1:0",
|
||||
"anthropic.claude-3-5-haiku-20241022-v1:0",
|
||||
"anthropic.claude-3-sonnet-20240229-v1:0",
|
||||
"anthropic.claude-3-5-sonnet-20241022-v2:0",
|
||||
"anthropic.claude-3-opus-20240229-v1:0",
|
||||
// Meta Llama Models
|
||||
"us.meta.llama3-1-8b-instruct-v1:0",
|
||||
"us.meta.llama3-1-70b-instruct-v1:0",
|
||||
"us.meta.llama3-2-11b-instruct-v1:0",
|
||||
"us.meta.llama3-2-90b-instruct-v1:0",
|
||||
"us.meta.llama3-3-70b-instruct-v1:0",
|
||||
// Mistral Models
|
||||
"mistral.mistral-large-2402-v1:0",
|
||||
"mistral.mistral-large-2407-v1:0",
|
||||
];
|
||||
|
||||
const googleModels = [
|
||||
"gemini-1.0-pro", // Deprecated on 2/15/2025
|
||||
"gemini-1.5-pro-latest",
|
||||
@@ -763,6 +799,7 @@ export const DEFAULT_MODELS = [
|
||||
sorted: 11,
|
||||
},
|
||||
})),
|
||||
|
||||
...chatglmModels.map((name) => ({
|
||||
name,
|
||||
available: true,
|
||||
@@ -796,6 +833,17 @@ export const DEFAULT_MODELS = [
|
||||
sorted: 14,
|
||||
},
|
||||
})),
|
||||
...bedrockModels.map((name) => ({
|
||||
name,
|
||||
available: true,
|
||||
sorted: seq++,
|
||||
provider: {
|
||||
id: "bedrock",
|
||||
providerName: "Bedrock",
|
||||
providerType: "bedrock",
|
||||
sorted: 15,
|
||||
},
|
||||
})),
|
||||
] as const;
|
||||
|
||||
export const CHAT_PAGE_SIZE = 15;
|
||||
|
||||
@@ -343,6 +343,32 @@ const cn = {
|
||||
SubTitle: "除默认地址外,必须包含 http(s)://",
|
||||
},
|
||||
},
|
||||
Bedrock: {
|
||||
Region: {
|
||||
Title: "AWS 区域",
|
||||
SubTitle: "Bedrock 服务所在的 AWS 区域",
|
||||
Placeholder: "us-west-2",
|
||||
Invalid: "无效的 AWS 区域格式。示例:us-west-2",
|
||||
},
|
||||
AccessKey: {
|
||||
Title: "AWS 访问密钥 ID",
|
||||
SubTitle: "用于 Bedrock 服务的 AWS 访问密钥 ID",
|
||||
Placeholder: "AKIA...",
|
||||
Invalid: "无效的 AWS Access Key 格式。必须为20个字符。",
|
||||
},
|
||||
SecretKey: {
|
||||
Title: "AWS 私有访问密钥",
|
||||
SubTitle: "用于 Bedrock 服务的 AWS 私有访问密钥",
|
||||
Placeholder: "****",
|
||||
Invalid: "无效的 AWS Secret Key 格式。必须为40个字符。",
|
||||
},
|
||||
EncryptionKey: {
|
||||
Title: "加密密钥",
|
||||
SubTitle: "用于配置数据的加密密钥",
|
||||
Placeholder: "输入加密密钥",
|
||||
Invalid: "无效的加密密钥。必须至少包含8个字符!",
|
||||
},
|
||||
},
|
||||
Azure: {
|
||||
ApiKey: {
|
||||
Title: "接口密钥",
|
||||
|
||||
@@ -347,6 +347,33 @@ const en: LocaleType = {
|
||||
SubTitle: "Must start with http(s):// or use /api/openai as default",
|
||||
},
|
||||
},
|
||||
Bedrock: {
|
||||
Region: {
|
||||
Title: "AWS Region",
|
||||
SubTitle: "The AWS region where Bedrock service is located",
|
||||
Placeholder: "us-west-2",
|
||||
Invalid: "Invalid AWS region format. Example: us-west-2",
|
||||
},
|
||||
AccessKey: {
|
||||
Title: "AWS Access Key ID",
|
||||
SubTitle: "Your AWS access key ID for Bedrock service",
|
||||
Placeholder: "AKIA...",
|
||||
Invalid: "Invalid AWS access key format. Must be 20 characters long.",
|
||||
},
|
||||
SecretKey: {
|
||||
Title: "AWS Secret Access Key",
|
||||
SubTitle: "Your AWS secret access key for Bedrock service",
|
||||
Placeholder: "****",
|
||||
Invalid: "Invalid AWS secret key format. Must be 40 characters long.",
|
||||
},
|
||||
EncryptionKey: {
|
||||
Title: "Encryption Key",
|
||||
SubTitle: "Your encryption key for configuration data",
|
||||
Placeholder: "Enter encryption key",
|
||||
Invalid:
|
||||
"Invalid encryption key format. Must no less than 8 characters long!",
|
||||
},
|
||||
},
|
||||
Azure: {
|
||||
ApiKey: {
|
||||
Title: "Azure Api Key",
|
||||
|
||||
@@ -16,6 +16,7 @@ import {
|
||||
DEEPSEEK_BASE_URL,
|
||||
XAI_BASE_URL,
|
||||
CHATGLM_BASE_URL,
|
||||
BEDROCK_BASE_URL,
|
||||
SILICONFLOW_BASE_URL,
|
||||
} from "../constant";
|
||||
import { getHeaders } from "../client/api";
|
||||
@@ -24,36 +25,26 @@ import { createPersistStore } from "../utils/store";
|
||||
import { ensure } from "../utils/clone";
|
||||
import { DEFAULT_CONFIG } from "./config";
|
||||
import { getModelProvider } from "../utils/model";
|
||||
import { encrypt, decrypt } from "../utils/aws";
|
||||
|
||||
let fetchState = 0; // 0 not fetch, 1 fetching, 2 done
|
||||
|
||||
const isApp = getClientConfig()?.buildMode === "export";
|
||||
|
||||
const DEFAULT_OPENAI_URL = isApp ? OPENAI_BASE_URL : ApiPath.OpenAI;
|
||||
|
||||
const DEFAULT_GOOGLE_URL = isApp ? GEMINI_BASE_URL : ApiPath.Google;
|
||||
|
||||
const DEFAULT_ANTHROPIC_URL = isApp ? ANTHROPIC_BASE_URL : ApiPath.Anthropic;
|
||||
|
||||
const DEFAULT_BAIDU_URL = isApp ? BAIDU_BASE_URL : ApiPath.Baidu;
|
||||
|
||||
const DEFAULT_BYTEDANCE_URL = isApp ? BYTEDANCE_BASE_URL : ApiPath.ByteDance;
|
||||
|
||||
const DEFAULT_ALIBABA_URL = isApp ? ALIBABA_BASE_URL : ApiPath.Alibaba;
|
||||
|
||||
const DEFAULT_TENCENT_URL = isApp ? TENCENT_BASE_URL : ApiPath.Tencent;
|
||||
|
||||
const DEFAULT_MOONSHOT_URL = isApp ? MOONSHOT_BASE_URL : ApiPath.Moonshot;
|
||||
|
||||
const DEFAULT_STABILITY_URL = isApp ? STABILITY_BASE_URL : ApiPath.Stability;
|
||||
|
||||
const DEFAULT_IFLYTEK_URL = isApp ? IFLYTEK_BASE_URL : ApiPath.Iflytek;
|
||||
|
||||
const DEFAULT_DEEPSEEK_URL = isApp ? DEEPSEEK_BASE_URL : ApiPath.DeepSeek;
|
||||
|
||||
const DEFAULT_XAI_URL = isApp ? XAI_BASE_URL : ApiPath.XAI;
|
||||
|
||||
const DEFAULT_CHATGLM_URL = isApp ? CHATGLM_BASE_URL : ApiPath.ChatGLM;
|
||||
const DEFAULT_BEDROCK_URL = isApp ? BEDROCK_BASE_URL : ApiPath.Bedrock;
|
||||
|
||||
const DEFAULT_SILICONFLOW_URL = isApp
|
||||
? SILICONFLOW_BASE_URL
|
||||
@@ -128,10 +119,19 @@ const DEFAULT_ACCESS_STATE = {
|
||||
chatglmUrl: DEFAULT_CHATGLM_URL,
|
||||
chatglmApiKey: "",
|
||||
|
||||
// aws bedrock
|
||||
bedrockUrl: DEFAULT_BEDROCK_URL,
|
||||
awsRegion: "",
|
||||
awsAccessKey: "",
|
||||
awsSecretKey: "",
|
||||
encryptionKey: "",
|
||||
bedrockAnthropicVersion: "bedrock-2023-05-31",
|
||||
|
||||
// siliconflow
|
||||
siliconflowUrl: DEFAULT_SILICONFLOW_URL,
|
||||
siliconflowApiKey: "",
|
||||
|
||||
|
||||
// server config
|
||||
needCode: true,
|
||||
hideUserApiKey: false,
|
||||
@@ -148,11 +148,9 @@ const DEFAULT_ACCESS_STATE = {
|
||||
|
||||
export const useAccessStore = createPersistStore(
|
||||
{ ...DEFAULT_ACCESS_STATE },
|
||||
|
||||
(set, get) => ({
|
||||
enabledAccessControl() {
|
||||
this.fetch();
|
||||
|
||||
return get().needCode;
|
||||
},
|
||||
getVisionModels() {
|
||||
@@ -161,7 +159,6 @@ export const useAccessStore = createPersistStore(
|
||||
},
|
||||
edgeVoiceName() {
|
||||
this.fetch();
|
||||
|
||||
return get().edgeTTSVoiceName;
|
||||
},
|
||||
|
||||
@@ -200,6 +197,7 @@ export const useAccessStore = createPersistStore(
|
||||
isValidMoonshot() {
|
||||
return ensure(get(), ["moonshotApiKey"]);
|
||||
},
|
||||
|
||||
isValidIflytek() {
|
||||
return ensure(get(), ["iflytekApiKey"]);
|
||||
},
|
||||
@@ -215,8 +213,19 @@ export const useAccessStore = createPersistStore(
|
||||
return ensure(get(), ["chatglmApiKey"]);
|
||||
},
|
||||
|
||||
|
||||
isValidBedrock() {
|
||||
return ensure(get(), [
|
||||
"awsRegion",
|
||||
"awsAccessKey",
|
||||
"awsSecretKey",
|
||||
"encryptionKey",
|
||||
]);
|
||||
},
|
||||
|
||||
isValidSiliconFlow() {
|
||||
return ensure(get(), ["siliconflowApiKey"]);
|
||||
|
||||
},
|
||||
|
||||
isAuthorized() {
|
||||
@@ -237,11 +246,13 @@ export const useAccessStore = createPersistStore(
|
||||
this.isValidDeepSeek() ||
|
||||
this.isValidXAI() ||
|
||||
this.isValidChatGLM() ||
|
||||
this.isValidBedrock() ||
|
||||
this.isValidSiliconFlow() ||
|
||||
!this.enabledAccessControl() ||
|
||||
(this.enabledAccessControl() && ensure(get(), ["accessCode"]))
|
||||
);
|
||||
},
|
||||
|
||||
fetch() {
|
||||
if (fetchState > 0 || getClientConfig()?.buildMode === "export") return;
|
||||
fetchState = 1;
|
||||
@@ -260,7 +271,6 @@ export const useAccessStore = createPersistStore(
|
||||
DEFAULT_CONFIG.modelConfig.model = model;
|
||||
DEFAULT_CONFIG.modelConfig.providerName = providerName as any;
|
||||
}
|
||||
|
||||
return res;
|
||||
})
|
||||
.then((res: DangerConfig) => {
|
||||
@@ -274,6 +284,43 @@ export const useAccessStore = createPersistStore(
|
||||
fetchState = 2;
|
||||
});
|
||||
},
|
||||
|
||||
// Override the set method to encrypt AWS credentials before storage
|
||||
set: (partial: { [key: string]: any }) => {
|
||||
if (partial.awsAccessKey) {
|
||||
partial.awsAccessKey = encrypt(
|
||||
partial.awsAccessKey,
|
||||
partial.encryptionKey,
|
||||
);
|
||||
}
|
||||
if (partial.awsSecretKey) {
|
||||
partial.awsSecretKey = encrypt(
|
||||
partial.awsSecretKey,
|
||||
partial.encryptionKey,
|
||||
);
|
||||
}
|
||||
if (partial.awsRegion) {
|
||||
partial.awsRegion = encrypt(partial.awsRegion, partial.encryptionKey);
|
||||
}
|
||||
set(partial);
|
||||
},
|
||||
|
||||
// Add getter to decrypt AWS credentials when needed
|
||||
get: () => {
|
||||
const state = get();
|
||||
return {
|
||||
...state,
|
||||
awsRegion: state.awsRegion
|
||||
? decrypt(state.awsRegion, state.encryptionKey)
|
||||
: "",
|
||||
awsAccessKey: state.awsAccessKey
|
||||
? decrypt(state.awsAccessKey, state.encryptionKey)
|
||||
: "",
|
||||
awsSecretKey: state.awsSecretKey
|
||||
? decrypt(state.awsSecretKey, state.encryptionKey)
|
||||
: "",
|
||||
};
|
||||
},
|
||||
}),
|
||||
{
|
||||
name: StoreKey.Access,
|
||||
|
||||
@@ -344,6 +344,13 @@ export function showPlugins(provider: ServiceProvider, model: string) {
|
||||
if (provider == ServiceProvider.Anthropic && !model.includes("claude-2")) {
|
||||
return true;
|
||||
}
|
||||
if (
|
||||
(provider == ServiceProvider.Bedrock && model.includes("claude-3")) ||
|
||||
model.includes("mistral-large") ||
|
||||
model.includes("amazon.nova")
|
||||
) {
|
||||
return true;
|
||||
}
|
||||
if (provider == ServiceProvider.Google && !model.includes("vision")) {
|
||||
return true;
|
||||
}
|
||||
|
||||
696
app/utils/aws.ts
Normal file
696
app/utils/aws.ts
Normal file
@@ -0,0 +1,696 @@
|
||||
// Types and Interfaces
|
||||
export interface BedrockCredentials {
|
||||
region: string;
|
||||
accessKeyId: string;
|
||||
secretAccessKey: string;
|
||||
}
|
||||
|
||||
// Type definitions for better type safety
|
||||
type ParsedEvent = Record<string, any>;
|
||||
type EventResult = ParsedEvent[];
|
||||
|
||||
// Using a dot as separator since it's not used in Base64
|
||||
const SEPARATOR = "~";
|
||||
|
||||
// Unified crypto utilities for both frontend and backend
|
||||
async function generateKey(
|
||||
password: string,
|
||||
salt: Uint8Array,
|
||||
): Promise<CryptoKey> {
|
||||
const enc = new TextEncoder();
|
||||
const keyMaterial = await crypto.subtle.importKey(
|
||||
"raw",
|
||||
enc.encode(password),
|
||||
{ name: "PBKDF2" },
|
||||
false,
|
||||
["deriveBits", "deriveKey"],
|
||||
);
|
||||
|
||||
return crypto.subtle.deriveKey(
|
||||
{
|
||||
name: "PBKDF2",
|
||||
salt,
|
||||
iterations: 100000,
|
||||
hash: "SHA-256",
|
||||
},
|
||||
keyMaterial,
|
||||
{ name: "AES-GCM", length: 256 },
|
||||
false,
|
||||
["encrypt", "decrypt"],
|
||||
);
|
||||
}
|
||||
|
||||
function arrayBufferToBase64(buffer: ArrayBuffer | Uint8Array): string {
|
||||
const bytes = buffer instanceof Uint8Array ? buffer : new Uint8Array(buffer);
|
||||
return btoa(String.fromCharCode(...bytes));
|
||||
}
|
||||
|
||||
function base64ToArrayBuffer(base64: string): Uint8Array {
|
||||
const binaryString = atob(base64);
|
||||
const bytes = new Uint8Array(binaryString.length);
|
||||
for (let i = 0; i < binaryString.length; i++) {
|
||||
bytes[i] = binaryString.charCodeAt(i);
|
||||
}
|
||||
return bytes;
|
||||
}
|
||||
|
||||
export async function encrypt(
|
||||
data: string,
|
||||
encryptionKey: string,
|
||||
): Promise<string> {
|
||||
if (!data) return "";
|
||||
if (!encryptionKey) {
|
||||
throw new Error("Encryption key is required for AWS credential encryption");
|
||||
}
|
||||
|
||||
try {
|
||||
const enc = new TextEncoder();
|
||||
const salt = crypto.getRandomValues(new Uint8Array(16));
|
||||
const iv = crypto.getRandomValues(new Uint8Array(12));
|
||||
const key = await generateKey(encryptionKey, salt);
|
||||
|
||||
const encrypted = await crypto.subtle.encrypt(
|
||||
{
|
||||
name: "AES-GCM",
|
||||
iv,
|
||||
},
|
||||
key,
|
||||
enc.encode(data),
|
||||
);
|
||||
|
||||
// Convert to base64 strings
|
||||
const encryptedBase64 = arrayBufferToBase64(encrypted);
|
||||
const saltBase64 = arrayBufferToBase64(salt);
|
||||
const ivBase64 = arrayBufferToBase64(iv);
|
||||
|
||||
return [saltBase64, ivBase64, encryptedBase64].join(SEPARATOR);
|
||||
} catch (error) {
|
||||
// console.error("[Encryption Error]:", error);
|
||||
throw new Error("Failed to encrypt AWS credentials");
|
||||
}
|
||||
}
|
||||
|
||||
export async function decrypt(
|
||||
encryptedData: string,
|
||||
encryptionKey: string,
|
||||
): Promise<string> {
|
||||
if (!encryptedData) return "";
|
||||
if (!encryptionKey) {
|
||||
throw new Error("Encryption key is required for AWS credential decryption");
|
||||
}
|
||||
|
||||
try {
|
||||
const [saltBase64, ivBase64, cipherBase64] = encryptedData.split(SEPARATOR);
|
||||
|
||||
// Convert base64 strings back to Uint8Arrays
|
||||
const salt = base64ToArrayBuffer(saltBase64);
|
||||
const iv = base64ToArrayBuffer(ivBase64);
|
||||
const cipherData = base64ToArrayBuffer(cipherBase64);
|
||||
|
||||
const key = await generateKey(encryptionKey, salt);
|
||||
|
||||
const decrypted = await crypto.subtle.decrypt(
|
||||
{
|
||||
name: "AES-GCM",
|
||||
iv,
|
||||
},
|
||||
key,
|
||||
cipherData,
|
||||
);
|
||||
|
||||
const dec = new TextDecoder();
|
||||
return dec.decode(decrypted);
|
||||
} catch (error) {
|
||||
throw new Error("Failed to decrypt AWS credentials");
|
||||
}
|
||||
}
|
||||
|
||||
export function maskSensitiveValue(value: string): string {
|
||||
if (!value) return "";
|
||||
if (value.length <= 6) return value;
|
||||
const masked = "*".repeat(value.length - 6);
|
||||
return value.slice(0, 3) + masked + value.slice(-3);
|
||||
}
|
||||
|
||||
// AWS Signing
|
||||
export interface SignParams {
|
||||
method: string;
|
||||
url: string;
|
||||
region: string;
|
||||
accessKeyId: string;
|
||||
secretAccessKey: string;
|
||||
body: string | object;
|
||||
service: string;
|
||||
headers?: Record<string, string>;
|
||||
isStreaming?: boolean;
|
||||
}
|
||||
|
||||
async function createHmac(
|
||||
key: ArrayBuffer | Uint8Array,
|
||||
data: string,
|
||||
): Promise<ArrayBuffer> {
|
||||
const encoder = new TextEncoder();
|
||||
const keyData = key instanceof Uint8Array ? key : new Uint8Array(key);
|
||||
const keyObject = await crypto.subtle.importKey(
|
||||
"raw",
|
||||
keyData,
|
||||
{ name: "HMAC", hash: "SHA-256" },
|
||||
false,
|
||||
["sign"],
|
||||
);
|
||||
return crypto.subtle.sign("HMAC", keyObject, encoder.encode(data));
|
||||
}
|
||||
|
||||
async function getSigningKey(
|
||||
secretKey: string,
|
||||
dateStamp: string,
|
||||
region: string,
|
||||
service: string,
|
||||
): Promise<ArrayBuffer> {
|
||||
const encoder = new TextEncoder();
|
||||
const kDate = await createHmac(encoder.encode("AWS4" + secretKey), dateStamp);
|
||||
const kRegion = await createHmac(kDate, region);
|
||||
const kService = await createHmac(kRegion, service);
|
||||
const kSigning = await createHmac(kService, "aws4_request");
|
||||
return kSigning;
|
||||
}
|
||||
|
||||
function normalizeHeaderValue(value: string): string {
|
||||
return value.replace(/\s+/g, " ").trim();
|
||||
}
|
||||
|
||||
function encodeRFC3986(str: string): string {
|
||||
return encodeURIComponent(str)
|
||||
.replace(
|
||||
/[!'()*]/g,
|
||||
(c) => "%" + c.charCodeAt(0).toString(16).toUpperCase(),
|
||||
)
|
||||
.replace(/[-_.~]/g, (c) => c);
|
||||
}
|
||||
|
||||
function getCanonicalUri(path: string): string {
|
||||
if (!path || path === "/") return "/";
|
||||
|
||||
return (
|
||||
"/" +
|
||||
path
|
||||
.split("/")
|
||||
.map((segment) => {
|
||||
if (!segment) return "";
|
||||
if (segment === "invoke-with-response-stream") return segment;
|
||||
|
||||
if (segment.includes("model/")) {
|
||||
return segment
|
||||
.split(/(model\/)/)
|
||||
.map((part) => {
|
||||
if (part === "model/") return part;
|
||||
return part
|
||||
.split(/([.:])/g)
|
||||
.map((subpart, i) =>
|
||||
i % 2 === 1 ? subpart : encodeRFC3986(subpart),
|
||||
)
|
||||
.join("");
|
||||
})
|
||||
.join("");
|
||||
}
|
||||
|
||||
return encodeRFC3986(segment);
|
||||
})
|
||||
.join("/")
|
||||
);
|
||||
}
|
||||
|
||||
export async function sign({
|
||||
method,
|
||||
url,
|
||||
region,
|
||||
accessKeyId,
|
||||
secretAccessKey,
|
||||
body,
|
||||
service,
|
||||
headers: customHeaders = {},
|
||||
isStreaming = true,
|
||||
}: SignParams): Promise<Record<string, string>> {
|
||||
try {
|
||||
const endpoint = new URL(url);
|
||||
const canonicalUri = getCanonicalUri(endpoint.pathname.slice(1));
|
||||
const canonicalQueryString = endpoint.search.slice(1);
|
||||
|
||||
const now = new Date();
|
||||
const amzDate = now.toISOString().replace(/[:-]|\.\d{3}/g, "");
|
||||
const dateStamp = amzDate.slice(0, 8);
|
||||
|
||||
const bodyString = typeof body === "string" ? body : JSON.stringify(body);
|
||||
const encoder = new TextEncoder();
|
||||
const payloadBuffer = await crypto.subtle.digest(
|
||||
"SHA-256",
|
||||
encoder.encode(bodyString),
|
||||
);
|
||||
const payloadHash = Array.from(new Uint8Array(payloadBuffer))
|
||||
.map((b) => b.toString(16).padStart(2, "0"))
|
||||
.join("");
|
||||
|
||||
const headers: Record<string, string> = {
|
||||
accept: isStreaming
|
||||
? "application/vnd.amazon.eventstream"
|
||||
: "application/json",
|
||||
"content-type": "application/json",
|
||||
host: endpoint.host,
|
||||
"x-amz-content-sha256": payloadHash,
|
||||
"x-amz-date": amzDate,
|
||||
...customHeaders,
|
||||
};
|
||||
|
||||
// Add x-amzn-bedrock-accept header for streaming requests
|
||||
if (isStreaming) {
|
||||
headers["x-amzn-bedrock-accept"] = "*/*";
|
||||
}
|
||||
|
||||
const sortedHeaderKeys = Object.keys(headers).sort((a, b) =>
|
||||
a.toLowerCase().localeCompare(b.toLowerCase()),
|
||||
);
|
||||
|
||||
const canonicalHeaders = sortedHeaderKeys
|
||||
.map(
|
||||
(key) => `${key.toLowerCase()}:${normalizeHeaderValue(headers[key])}\n`,
|
||||
)
|
||||
.join("");
|
||||
|
||||
const signedHeaders = sortedHeaderKeys
|
||||
.map((key) => key.toLowerCase())
|
||||
.join(";");
|
||||
|
||||
const canonicalRequest = [
|
||||
method.toUpperCase(),
|
||||
canonicalUri,
|
||||
canonicalQueryString,
|
||||
canonicalHeaders,
|
||||
signedHeaders,
|
||||
payloadHash,
|
||||
].join("\n");
|
||||
|
||||
const algorithm = "AWS4-HMAC-SHA256";
|
||||
const credentialScope = `${dateStamp}/${region}/${service}/aws4_request`;
|
||||
|
||||
const canonicalRequestHash = Array.from(
|
||||
new Uint8Array(
|
||||
await crypto.subtle.digest("SHA-256", encoder.encode(canonicalRequest)),
|
||||
),
|
||||
)
|
||||
.map((b) => b.toString(16).padStart(2, "0"))
|
||||
.join("");
|
||||
|
||||
const stringToSign = [
|
||||
algorithm,
|
||||
amzDate,
|
||||
credentialScope,
|
||||
canonicalRequestHash,
|
||||
].join("\n");
|
||||
|
||||
const signingKey = await getSigningKey(
|
||||
secretAccessKey,
|
||||
dateStamp,
|
||||
region,
|
||||
service,
|
||||
);
|
||||
|
||||
const signature = Array.from(
|
||||
new Uint8Array(await createHmac(signingKey, stringToSign)),
|
||||
)
|
||||
.map((b) => b.toString(16).padStart(2, "0"))
|
||||
.join("");
|
||||
|
||||
const authorization = [
|
||||
`${algorithm} Credential=${accessKeyId}/${credentialScope}`,
|
||||
`SignedHeaders=${signedHeaders}`,
|
||||
`Signature=${signature}`,
|
||||
].join(", ");
|
||||
|
||||
return {
|
||||
...headers,
|
||||
Authorization: authorization,
|
||||
};
|
||||
} catch (error) {
|
||||
console.error("[AWS Signing Error]: Failed to sign request");
|
||||
throw new Error("Failed to sign AWS request");
|
||||
}
|
||||
}
|
||||
|
||||
// Bedrock utilities
|
||||
function decodeBase64(base64String: string): string {
|
||||
try {
|
||||
const bytes = Buffer.from(base64String, "base64");
|
||||
const decoder = new TextDecoder("utf-8");
|
||||
return decoder.decode(bytes);
|
||||
} catch (e) {
|
||||
console.error("[Base64 Decode Error]:", e);
|
||||
return "";
|
||||
}
|
||||
}
|
||||
|
||||
export function parseEventData(chunk: Uint8Array): EventResult {
|
||||
const decoder = new TextDecoder("utf-8");
|
||||
const text = decoder.decode(chunk);
|
||||
const results: EventResult = [];
|
||||
|
||||
try {
|
||||
// First try to parse as regular JSON
|
||||
const parsed = JSON.parse(text);
|
||||
if (parsed.bytes) {
|
||||
const decoded = decodeBase64(parsed.bytes);
|
||||
try {
|
||||
const decodedJson = JSON.parse(decoded);
|
||||
results.push(decodedJson);
|
||||
} catch (e) {
|
||||
results.push({ output: decoded });
|
||||
}
|
||||
return results;
|
||||
}
|
||||
|
||||
if (typeof parsed.body === "string") {
|
||||
try {
|
||||
const parsedBody = JSON.parse(parsed.body);
|
||||
results.push(parsedBody);
|
||||
} catch (e) {
|
||||
results.push({ output: parsed.body });
|
||||
}
|
||||
return results;
|
||||
}
|
||||
|
||||
results.push(parsed.body || parsed);
|
||||
return results;
|
||||
} catch (e) {
|
||||
// If regular JSON parse fails, try to extract event content
|
||||
const eventRegex = /:event-type[^\{]+(\{[^\}]+\})/g;
|
||||
let match;
|
||||
|
||||
while ((match = eventRegex.exec(text)) !== null) {
|
||||
try {
|
||||
const eventData = match[1];
|
||||
const parsed = JSON.parse(eventData);
|
||||
|
||||
if (parsed.bytes) {
|
||||
const decoded = decodeBase64(parsed.bytes);
|
||||
try {
|
||||
const decodedJson = JSON.parse(decoded);
|
||||
if (decodedJson.choices?.[0]?.message?.content) {
|
||||
results.push({ output: decodedJson.choices[0].message.content });
|
||||
} else {
|
||||
results.push(decodedJson);
|
||||
}
|
||||
} catch (e) {
|
||||
results.push({ output: decoded });
|
||||
}
|
||||
} else {
|
||||
results.push(parsed);
|
||||
}
|
||||
} catch (e) {
|
||||
console.debug("[Event Parse Warning]:", e);
|
||||
}
|
||||
}
|
||||
|
||||
// If no events were found, try to extract clean text
|
||||
if (results.length === 0) {
|
||||
// Remove event metadata markers and clean the text
|
||||
const cleanText = text
|
||||
.replace(/\{KG[^:]+:event-type[^}]+\}/g, "") // Remove event markers
|
||||
.replace(/[\x00-\x1F\x7F-\x9F\uFEFF]/g, "") // Remove control characters
|
||||
.trim();
|
||||
|
||||
if (cleanText) {
|
||||
results.push({ output: cleanText });
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return results;
|
||||
}
|
||||
|
||||
export function processMessage(
|
||||
data: ParsedEvent,
|
||||
remainText: string,
|
||||
runTools: any[],
|
||||
index: number,
|
||||
): { remainText: string; index: number } {
|
||||
if (!data) return { remainText, index };
|
||||
|
||||
try {
|
||||
// Handle Nova's tool calls with exact schema match
|
||||
// console.log("processMessage data=========================",data);
|
||||
if (data.contentBlockStart?.start?.toolUse) {
|
||||
const toolUse = data.contentBlockStart.start.toolUse;
|
||||
index += 1;
|
||||
runTools.push({
|
||||
id: toolUse.toolUseId,
|
||||
type: "function",
|
||||
function: {
|
||||
name: toolUse.name || "", // Ensure name is always present
|
||||
arguments: "{}", // Initialize empty arguments
|
||||
},
|
||||
});
|
||||
return { remainText, index };
|
||||
}
|
||||
|
||||
// Handle Nova's tool input in contentBlockDelta
|
||||
if (data.contentBlockDelta?.delta?.toolUse?.input) {
|
||||
if (runTools[index]) {
|
||||
runTools[index].function.arguments =
|
||||
data.contentBlockDelta.delta.toolUse.input;
|
||||
}
|
||||
return { remainText, index };
|
||||
}
|
||||
|
||||
// Handle Nova's text content
|
||||
if (data.output?.message?.content?.[0]?.text) {
|
||||
remainText += data.output.message.content[0].text;
|
||||
return { remainText, index };
|
||||
}
|
||||
|
||||
// Handle Nova's messageStart event
|
||||
if (data.messageStart) {
|
||||
return { remainText, index };
|
||||
}
|
||||
|
||||
// Handle Nova's text delta
|
||||
if (data.contentBlockDelta?.delta?.text) {
|
||||
remainText += data.contentBlockDelta.delta.text;
|
||||
return { remainText, index };
|
||||
}
|
||||
|
||||
// Handle Nova's contentBlockStop event
|
||||
if (data.contentBlockStop) {
|
||||
return { remainText, index };
|
||||
}
|
||||
|
||||
// Handle Nova's messageStop event
|
||||
if (data.messageStop) {
|
||||
return { remainText, index };
|
||||
}
|
||||
|
||||
// Handle message_start event (for other models)
|
||||
if (data.type === "message_start") {
|
||||
return { remainText, index };
|
||||
}
|
||||
|
||||
// Handle content_block_start event (for other models)
|
||||
if (data.type === "content_block_start") {
|
||||
if (data.content_block?.type === "tool_use") {
|
||||
index += 1;
|
||||
runTools.push({
|
||||
id: data.content_block.id,
|
||||
type: "function",
|
||||
function: {
|
||||
name: data.content_block.name || "", // Ensure name is always present
|
||||
arguments: "",
|
||||
},
|
||||
});
|
||||
}
|
||||
return { remainText, index };
|
||||
}
|
||||
|
||||
// Handle content_block_delta event (for other models)
|
||||
if (data.type === "content_block_delta") {
|
||||
if (data.delta?.type === "input_json_delta" && runTools[index]) {
|
||||
runTools[index].function.arguments += data.delta.partial_json;
|
||||
} else if (data.delta?.type === "text_delta") {
|
||||
const newText = data.delta.text || "";
|
||||
remainText += newText;
|
||||
}
|
||||
return { remainText, index };
|
||||
}
|
||||
|
||||
// Handle tool calls for other models
|
||||
if (data.choices?.[0]?.message?.tool_calls) {
|
||||
for (const toolCall of data.choices[0].message.tool_calls) {
|
||||
index += 1;
|
||||
runTools.push({
|
||||
id: toolCall.id || `tool-${Date.now()}`,
|
||||
type: "function",
|
||||
function: {
|
||||
name: toolCall.function?.name || "", // Ensure name is always present
|
||||
arguments: toolCall.function?.arguments || "",
|
||||
},
|
||||
});
|
||||
}
|
||||
return { remainText, index };
|
||||
}
|
||||
|
||||
// Handle various response formats
|
||||
let newText = "";
|
||||
if (data.delta?.text) {
|
||||
newText = data.delta.text;
|
||||
} else if (data.choices?.[0]?.message?.content) {
|
||||
newText = data.choices[0].message.content;
|
||||
} else if (data.content?.[0]?.text) {
|
||||
newText = data.content[0].text;
|
||||
} else if (data.generation) {
|
||||
newText = data.generation;
|
||||
} else if (data.outputText) {
|
||||
newText = data.outputText;
|
||||
} else if (data.response) {
|
||||
newText = data.response;
|
||||
} else if (data.output) {
|
||||
newText = data.output;
|
||||
}
|
||||
|
||||
// Only append if we have new text
|
||||
if (newText) {
|
||||
remainText += newText;
|
||||
}
|
||||
} catch (e) {
|
||||
console.warn("Failed to process Bedrock message:");
|
||||
}
|
||||
|
||||
return { remainText, index };
|
||||
}
|
||||
|
||||
export function processChunks(
|
||||
chunks: Uint8Array[],
|
||||
pendingChunk: Uint8Array | null,
|
||||
remainText: string,
|
||||
runTools: any[],
|
||||
index: number,
|
||||
): {
|
||||
chunks: Uint8Array[];
|
||||
pendingChunk: Uint8Array | null;
|
||||
remainText: string;
|
||||
index: number;
|
||||
} {
|
||||
let currentText = remainText;
|
||||
let currentIndex = index;
|
||||
|
||||
while (chunks.length > 0) {
|
||||
const chunk = chunks[0];
|
||||
try {
|
||||
// If there's a pending chunk, try to merge it with the current chunk
|
||||
let chunkToProcess = chunk;
|
||||
if (pendingChunk) {
|
||||
const mergedChunk = new Uint8Array(pendingChunk.length + chunk.length);
|
||||
mergedChunk.set(pendingChunk);
|
||||
mergedChunk.set(chunk, pendingChunk.length);
|
||||
chunkToProcess = mergedChunk;
|
||||
pendingChunk = null;
|
||||
}
|
||||
|
||||
// Try to process the chunk
|
||||
const parsedEvents = parseEventData(chunkToProcess);
|
||||
if (parsedEvents.length > 0) {
|
||||
// Process each event in the chunk
|
||||
for (const parsed of parsedEvents) {
|
||||
const result = processMessage(
|
||||
parsed,
|
||||
currentText,
|
||||
runTools,
|
||||
currentIndex,
|
||||
);
|
||||
currentText = result.remainText;
|
||||
currentIndex = result.index;
|
||||
}
|
||||
chunks.shift(); // Remove processed chunk
|
||||
} else {
|
||||
// If parsing fails, it might be an incomplete chunk
|
||||
pendingChunk = chunkToProcess;
|
||||
chunks.shift();
|
||||
}
|
||||
} catch (e) {
|
||||
// console.error("[Chunk Process Error]:", e);
|
||||
// chunks.shift(); // Remove error chunk
|
||||
// pendingChunk = null; // Reset pending chunk on error
|
||||
console.warn("Failed to process chunk, attempting recovery");
|
||||
// Attempt to recover by processing the next chunk
|
||||
if (chunks.length > 1) {
|
||||
chunks.shift();
|
||||
pendingChunk = null;
|
||||
} else {
|
||||
// If this is the last chunk, throw to prevent data loss
|
||||
throw new Error("Failed to process final chunk");
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return {
|
||||
chunks,
|
||||
pendingChunk,
|
||||
remainText: currentText,
|
||||
index: currentIndex,
|
||||
};
|
||||
}
|
||||
|
||||
export function getBedrockEndpoint(
|
||||
region: string,
|
||||
modelId: string,
|
||||
shouldStream: boolean,
|
||||
): string {
|
||||
if (!region || !modelId) {
|
||||
throw new Error("Region and model ID are required for Bedrock endpoint");
|
||||
}
|
||||
const baseEndpoint = `https://bedrock-runtime.${region}.amazonaws.com`;
|
||||
const endpoint =
|
||||
shouldStream === false
|
||||
? `${baseEndpoint}/model/${modelId}/invoke`
|
||||
: `${baseEndpoint}/model/${modelId}/invoke-with-response-stream`;
|
||||
return endpoint;
|
||||
}
|
||||
|
||||
export function extractMessage(res: any, modelId: string = ""): string {
|
||||
if (!res) {
|
||||
throw new Error("Empty response received");
|
||||
}
|
||||
|
||||
let message = "";
|
||||
|
||||
// Handle Nova model response format
|
||||
if (modelId.toLowerCase().includes("nova")) {
|
||||
if (res.output?.message?.content?.[0]?.text) {
|
||||
message = res.output.message.content[0].text;
|
||||
} else {
|
||||
message = res.output || "";
|
||||
}
|
||||
}
|
||||
// Handle Mistral model response format
|
||||
else if (modelId.toLowerCase().includes("mistral")) {
|
||||
if (res.choices?.[0]?.message?.content) {
|
||||
message = res.choices[0].message.content;
|
||||
} else {
|
||||
message = res.output || "";
|
||||
}
|
||||
}
|
||||
// Handle Llama model response format
|
||||
else if (modelId.toLowerCase().includes("llama")) {
|
||||
message = res?.generation || "";
|
||||
}
|
||||
// Handle Titan model response format
|
||||
else if (modelId.toLowerCase().includes("titan")) {
|
||||
message = res?.outputText || "";
|
||||
}
|
||||
// Handle Claude and other models
|
||||
else if (res.content?.[0]?.text) {
|
||||
message = res.content[0].text;
|
||||
}
|
||||
// Handle other response formats
|
||||
else {
|
||||
message = res.output || res.response || res.message || "";
|
||||
}
|
||||
|
||||
return message;
|
||||
}
|
||||
258
docs/bedrock-response-format.md
Normal file
258
docs/bedrock-response-format.md
Normal file
@@ -0,0 +1,258 @@
|
||||
# Understanding Bedrock Response Format
|
||||
|
||||
The AWS Bedrock streaming response format consists of multiple Server-Sent Events (SSE) chunks. Each chunk follows this structure:
|
||||
|
||||
```
|
||||
:event-type chunk
|
||||
:content-type application/json
|
||||
:message-type event
|
||||
{"bytes":"base64_encoded_data","p":"signature"}
|
||||
```
|
||||
|
||||
## Model-Specific Response Formats
|
||||
|
||||
### Claude 3 Format
|
||||
|
||||
When using Claude 3 models (e.g., claude-3-haiku-20240307), the decoded messages include:
|
||||
|
||||
1. **message_start**
|
||||
```json
|
||||
{
|
||||
"type": "message_start",
|
||||
"message": {
|
||||
"id": "msg_bdrk_01A6sahWac4XVTR9sX3rgvsZ",
|
||||
"type": "message",
|
||||
"role": "assistant",
|
||||
"model": "claude-3-haiku-20240307",
|
||||
"content": [],
|
||||
"stop_reason": null,
|
||||
"stop_sequence": null,
|
||||
"usage": {
|
||||
"input_tokens": 8,
|
||||
"output_tokens": 1
|
||||
}
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
2. **content_block_start**
|
||||
```json
|
||||
{
|
||||
"type": "content_block_start",
|
||||
"index": 0,
|
||||
"content_block": {
|
||||
"type": "text",
|
||||
"text": ""
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
3. **content_block_delta**
|
||||
```json
|
||||
{
|
||||
"type": "content_block_delta",
|
||||
"index": 0,
|
||||
"delta": {
|
||||
"type": "text_delta",
|
||||
"text": "Hello"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### Mistral Format
|
||||
|
||||
When using Mistral models (e.g., mistral-large-2407), the decoded messages have a different structure:
|
||||
|
||||
```json
|
||||
{
|
||||
"id": "b0098812-0ad9-42da-9f17-a5e2f554eb6b",
|
||||
"object": "chat.completion.chunk",
|
||||
"created": 1732582566,
|
||||
"model": "mistral-large-2407",
|
||||
"choices": [{
|
||||
"index": 0,
|
||||
"logprobs": null,
|
||||
"context_logits": null,
|
||||
"generation_logits": null,
|
||||
"message": {
|
||||
"role": null,
|
||||
"content": "Hello",
|
||||
"tool_calls": null,
|
||||
"index": null,
|
||||
"tool_call_id": null
|
||||
},
|
||||
"stop_reason": null
|
||||
}],
|
||||
"usage": null,
|
||||
"p": null
|
||||
}
|
||||
```
|
||||
|
||||
### Llama Format
|
||||
|
||||
When using Llama models (3.1 or 3.2), the decoded messages use a simpler structure focused on generation tokens:
|
||||
|
||||
```json
|
||||
{
|
||||
"generation": "Hello",
|
||||
"prompt_token_count": null,
|
||||
"generation_token_count": 2,
|
||||
"stop_reason": null
|
||||
}
|
||||
```
|
||||
|
||||
Each chunk contains:
|
||||
- generation: The generated text piece
|
||||
- prompt_token_count: Token count of the input (only present in first chunk)
|
||||
- generation_token_count: Running count of generated tokens
|
||||
- stop_reason: Indicates completion (null until final chunk)
|
||||
|
||||
First chunk example (includes prompt_token_count):
|
||||
```json
|
||||
{
|
||||
"generation": "\n\n",
|
||||
"prompt_token_count": 10,
|
||||
"generation_token_count": 1,
|
||||
"stop_reason": null
|
||||
}
|
||||
```
|
||||
|
||||
### Titan Text Format
|
||||
|
||||
When using Amazon's Titan models (text or TG1), the response comes as a single chunk with complete text and metrics:
|
||||
|
||||
```json
|
||||
{
|
||||
"outputText": "\nBot: Hello! How can I help you today?",
|
||||
"index": 0,
|
||||
"totalOutputTextTokenCount": 13,
|
||||
"completionReason": "FINISH",
|
||||
"inputTextTokenCount": 3,
|
||||
"amazon-bedrock-invocationMetrics": {
|
||||
"inputTokenCount": 3,
|
||||
"outputTokenCount": 13,
|
||||
"invocationLatency": 833,
|
||||
"firstByteLatency": 833
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
Both Titan text and Titan TG1 use the same response format, with only minor differences in token counts and latency values. For example, here's a TG1 response:
|
||||
|
||||
```json
|
||||
{
|
||||
"outputText": "\nBot: Hello! How can I help you?",
|
||||
"index": 0,
|
||||
"totalOutputTextTokenCount": 12,
|
||||
"completionReason": "FINISH",
|
||||
"inputTextTokenCount": 3,
|
||||
"amazon-bedrock-invocationMetrics": {
|
||||
"inputTokenCount": 3,
|
||||
"outputTokenCount": 12,
|
||||
"invocationLatency": 845,
|
||||
"firstByteLatency": 845
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
Key fields:
|
||||
- outputText: The complete generated response
|
||||
- totalOutputTextTokenCount: Total tokens in the response
|
||||
- completionReason: Reason for completion (e.g., "FINISH")
|
||||
- inputTextTokenCount: Number of input tokens
|
||||
- amazon-bedrock-invocationMetrics: Detailed performance metrics
|
||||
|
||||
## Model-Specific Completion Metrics
|
||||
|
||||
### Mistral
|
||||
```json
|
||||
{
|
||||
"usage": {
|
||||
"prompt_tokens": 5,
|
||||
"total_tokens": 29,
|
||||
"completion_tokens": 24
|
||||
},
|
||||
"amazon-bedrock-invocationMetrics": {
|
||||
"inputTokenCount": 5,
|
||||
"outputTokenCount": 24,
|
||||
"invocationLatency": 719,
|
||||
"firstByteLatency": 148
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### Claude 3
|
||||
Included in the message_delta with stop_reason.
|
||||
|
||||
### Llama
|
||||
Included in the final chunk with stop_reason "stop":
|
||||
```json
|
||||
{
|
||||
"amazon-bedrock-invocationMetrics": {
|
||||
"inputTokenCount": 10,
|
||||
"outputTokenCount": 11,
|
||||
"invocationLatency": 873,
|
||||
"firstByteLatency": 550
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### Titan
|
||||
Both Titan text and TG1 include metrics in the single response chunk:
|
||||
```json
|
||||
{
|
||||
"amazon-bedrock-invocationMetrics": {
|
||||
"inputTokenCount": 3,
|
||||
"outputTokenCount": 12,
|
||||
"invocationLatency": 845,
|
||||
"firstByteLatency": 845
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
## How the Response is Processed
|
||||
|
||||
1. The raw response is first split into chunks based on SSE format
|
||||
2. For each chunk:
|
||||
- The base64 encoded data is decoded
|
||||
- The JSON is parsed to extract the message content
|
||||
- Based on the model type and message type, different processing is applied:
|
||||
|
||||
### Claude 3 Processing
|
||||
- message_start: Initializes a new message with model info and usage stats
|
||||
- content_block_start: Starts a new content block (text, tool use, etc.)
|
||||
- content_block_delta: Adds incremental content to the current block
|
||||
- message_delta: Updates message metadata
|
||||
|
||||
### Mistral Processing
|
||||
- Each chunk contains a complete message object with choices array
|
||||
- The content is streamed through the message.content field
|
||||
- Final chunk includes token usage and invocation metrics
|
||||
|
||||
### Llama Processing
|
||||
- Each chunk contains a generation field with the text piece
|
||||
- First chunk includes prompt_token_count
|
||||
- Tracks generation progress through generation_token_count
|
||||
- Simple streaming format focused on text generation
|
||||
- Final chunk includes complete metrics
|
||||
|
||||
### Titan Processing
|
||||
- Single chunk response with complete text
|
||||
- No streaming - returns full response at once
|
||||
- Includes comprehensive metrics in the same chunk
|
||||
|
||||
## Handling in Code
|
||||
|
||||
The response is processed by the `transformBedrockStream` function in `app/utils/aws.ts`, which:
|
||||
|
||||
1. Reads the stream chunks
|
||||
2. Parses each chunk using `parseEventData`
|
||||
3. Handles model-specific formats:
|
||||
- For Claude: Processes message_start, content_block_start, content_block_delta
|
||||
- For Mistral: Extracts content from choices[0].message.content
|
||||
- For Llama: Uses the generation field directly
|
||||
- For Titan: Uses the outputText field from the single response
|
||||
4. Transforms the parsed data into a consistent format for the client
|
||||
5. Yields the transformed data as SSE events
|
||||
|
||||
This allows for real-time streaming of the model's response while maintaining a consistent format for the client application, regardless of which model is being used.
|
||||
@@ -94,4 +94,4 @@
|
||||
"lint-staged/yaml": "^2.2.2"
|
||||
},
|
||||
"packageManager": "yarn@1.22.19"
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user