merlyn/server/utils/AiProviders/bedrock/index.js
Timothy Carambat 78c83383d8
Overhaul AWS Bedrock provider (#3537)
* Patch AWS Bedrock provider for newer models and performance

* patch prompt constructor
2025-03-25 15:58:16 -07:00

434 lines
14 KiB
JavaScript

const {
BedrockRuntimeClient,
ConverseCommand,
ConverseStreamCommand,
} = require("@aws-sdk/client-bedrock-runtime");
const {
writeResponseChunk,
clientAbortedHandler,
} = require("../../helpers/chat/responses");
const { NativeEmbedder } = require("../../EmbeddingEngines/native");
const {
LLMPerformanceMonitor,
} = require("../../helpers/chat/LLMPerformanceMonitor");
const { v4: uuidv4 } = require("uuid");
class AWSBedrockLLM {
/**
* These models do not support system prompts
* It is not explicitly stated but it is observed that they do not use the system prompt
* in their responses and will crash when a system prompt is provided.
* We can add more models to this list as we discover them or new models are added.
* We may want to extend this list or make a user-config if using custom bedrock models.
*/
noSystemPromptModels = [
"amazon.titan-text-express-v1",
"amazon.titan-text-lite-v1",
"cohere.command-text-v14",
"cohere.command-light-text-v14",
"us.deepseek.r1-v1:0",
];
constructor(embedder = null, modelPreference = null) {
if (!process.env.AWS_BEDROCK_LLM_ACCESS_KEY_ID)
throw new Error("No AWS Bedrock LLM profile id was set.");
if (!process.env.AWS_BEDROCK_LLM_ACCESS_KEY)
throw new Error("No AWS Bedrock LLM access key was set.");
if (!process.env.AWS_BEDROCK_LLM_REGION)
throw new Error("No AWS Bedrock LLM region was set.");
if (
process.env.AWS_BEDROCK_LLM_CONNECTION_METHOD === "sessionToken" &&
!process.env.AWS_BEDROCK_LLM_SESSION_TOKEN
)
throw new Error(
"No AWS Bedrock LLM session token was set while using session token as the authentication method."
);
this.model =
modelPreference || process.env.AWS_BEDROCK_LLM_MODEL_PREFERENCE;
this.limits = {
history: this.promptWindowLimit() * 0.15,
system: this.promptWindowLimit() * 0.15,
user: this.promptWindowLimit() * 0.7,
};
this.bedrockClient = new BedrockRuntimeClient({
region: process.env.AWS_BEDROCK_LLM_REGION,
credentials: {
accessKeyId: process.env.AWS_BEDROCK_LLM_ACCESS_KEY_ID,
secretAccessKey: process.env.AWS_BEDROCK_LLM_ACCESS_KEY,
...(this.authMethod === "sessionToken"
? { sessionToken: process.env.AWS_BEDROCK_LLM_SESSION_TOKEN }
: {}),
},
});
this.embedder = embedder ?? new NativeEmbedder();
this.defaultTemp = 0.7;
this.#log(
`Loaded with model: ${this.model}. Will communicate with AWS Bedrock using ${this.authMethod} authentication.`
);
}
/**
* Get the authentication method for the AWS Bedrock LLM.
* There are only two valid values for this setting - anything else will default to "iam".
* @returns {"iam"|"sessionToken"}
*/
get authMethod() {
const method = process.env.AWS_BEDROCK_LLM_CONNECTION_METHOD || "iam";
if (!["iam", "sessionToken"].includes(method)) return "iam";
return method;
}
#appendContext(contextTexts = []) {
if (!contextTexts || !contextTexts.length) return "";
return (
"\nContext:\n" +
contextTexts
.map((text, i) => {
return `[CONTEXT ${i}]:\n${text}\n[END CONTEXT ${i}]\n\n`;
})
.join("")
);
}
#log(text, ...args) {
console.log(`\x1b[32m[AWSBedrock]\x1b[0m ${text}`, ...args);
}
streamingEnabled() {
return "streamGetChatCompletion" in this;
}
static promptWindowLimit(_modelName) {
const limit = process.env.AWS_BEDROCK_LLM_MODEL_TOKEN_LIMIT || 8191;
if (!limit || isNaN(Number(limit)))
throw new Error("No valid token context limit was set.");
return Number(limit);
}
// Ensure the user set a value for the token limit
// and if undefined - assume 4096 window.
promptWindowLimit() {
const limit = process.env.AWS_BEDROCK_LLM_MODEL_TOKEN_LIMIT || 8191;
if (!limit || isNaN(Number(limit)))
throw new Error("No valid token context limit was set.");
return Number(limit);
}
async isValidChatCompletionModel(_ = "") {
return true;
}
/**
* Generates appropriate content array for a message + attachments.
* TODO: Implement this - attachments are not supported yet for Bedrock.
* @param {{userPrompt:string, attachments: import("../../helpers").Attachment[]}}
* @returns {string|object[]}
*/
#generateContent({ userPrompt, attachments = [] }) {
if (!attachments.length) return [{ text: userPrompt }];
// const content = [{ type: "text", text: userPrompt }];
// for (let attachment of attachments) {
// content.push({
// type: "image_url",
// image_url: attachment.contentString,
// });
// }
// return { content: content.flat() };
}
/**
* Construct the user prompt for this model.
* @param {{attachments: import("../../helpers").Attachment[]}} param0
* @returns
*/
constructPrompt({
systemPrompt = "",
contextTexts = [],
chatHistory = [],
userPrompt = "",
_attachments = [],
}) {
let prompt = [
{
role: "system",
content: [
{ text: `${systemPrompt}${this.#appendContext(contextTexts)}` },
],
},
];
// If the model does not support system prompts, we need to add a user message and assistant message
if (this.noSystemPromptModels.includes(this.model)) {
prompt = [
{
role: "user",
content: [
{ text: `${systemPrompt}${this.#appendContext(contextTexts)}` },
],
},
{
role: "assistant",
content: [{ text: "Okay." }],
},
];
}
return [
...prompt,
...chatHistory.map((msg) => ({
role: msg.role,
content: this.#generateContent({
userPrompt: msg.content,
attachments: msg.attachments,
}),
})),
{
role: "user",
content: this.#generateContent({
userPrompt: userPrompt,
attachments: [],
}),
},
];
}
/**
* Parses and prepends reasoning from the response and returns the full text response.
* @param {Object} response
* @returns {string}
*/
#parseReasoningFromResponse({ content = [] }) {
let textResponse = content[0]?.text;
if (
!!content?.[1]?.reasoningContent &&
content?.[1]?.reasoningContent?.reasoningText?.text?.trim().length > 0
)
textResponse = `<think>${content?.[1]?.reasoningContent?.reasoningText?.text}</think>${textResponse}`;
return textResponse;
}
async getChatCompletion(messages = null, { temperature = 0.7 }) {
const hasSystem = messages[0]?.role === "system";
const [system, ...history] = hasSystem ? messages : [null, ...messages];
const result = await LLMPerformanceMonitor.measureAsyncFunction(
this.bedrockClient
.send(
new ConverseCommand({
modelId: this.model,
messages: history,
inferenceConfig: {
maxTokens: this.promptWindowLimit(),
temperature,
},
system: !!system ? system.content : undefined,
})
)
.catch((e) => {
throw new Error(
`AWSBedrock::getChatCompletion failed to communicate with Bedrock client. ${e.message}`
);
}),
messages,
false
);
const response = result.output;
if (!response || !response?.output) return null;
return {
textResponse: this.#parseReasoningFromResponse(response.output?.message),
metrics: {
prompt_tokens: response?.usage?.inputTokens,
completion_tokens: response?.usage?.outputTokens,
total_tokens: response?.usage?.totalTokens,
outputTps:
response?.usage?.outputTokens / (response?.metrics?.latencyMs / 1000),
duration: result.duration,
},
};
}
async streamGetChatCompletion(messages = null, { temperature = 0.7 }) {
const hasSystem = messages[0]?.role === "system";
const [system, ...history] = hasSystem ? messages : [null, ...messages];
const measuredStreamRequest = await LLMPerformanceMonitor.measureStream(
this.bedrockClient.send(
new ConverseStreamCommand({
modelId: this.model,
messages: history,
inferenceConfig: { maxTokens: this.promptWindowLimit(), temperature },
system: !!system ? system.content : undefined,
})
),
messages,
false
);
return measuredStreamRequest;
}
/**
* Handles the stream response from the AWS Bedrock API.
* Bedrock does not support usage metrics in the stream response so we need to estimate them.
* @param {Object} response - the response object
* @param {import('../../helpers/chat/LLMPerformanceMonitor').MonitoredStream} stream - the stream response from the AWS Bedrock API w/tracking
* @param {Object} responseProps - the response properties
* @returns {Promise<string>}
*/
handleStream(response, stream, responseProps) {
const { uuid = uuidv4(), sources = [] } = responseProps;
let hasUsageMetrics = false;
let usage = {
prompt_tokens: 0,
completion_tokens: 0,
};
return new Promise(async (resolve) => {
let fullText = "";
let reasoningText = "";
// Establish listener to early-abort a streaming response
// in case things go sideways or the user does not like the response.
// We preserve the generated text but continue as if chat was completed
// to preserve previously generated content.
const handleAbort = () => {
stream?.endMeasurement(usage);
clientAbortedHandler(resolve, fullText);
};
response.on("close", handleAbort);
try {
for await (const chunk of stream.stream) {
if (chunk === undefined)
throw new Error(
"Stream returned undefined chunk. Aborting reply - check model provider logs."
);
const action = Object.keys(chunk)[0];
if (action === "metadata") {
hasUsageMetrics = true;
usage.prompt_tokens = chunk.metadata?.usage?.inputTokens ?? 0;
usage.completion_tokens = chunk.metadata?.usage?.outputTokens ?? 0;
usage.total_tokens = chunk.metadata?.usage?.totalTokens ?? 0;
}
if (action === "contentBlockDelta") {
const token = chunk.contentBlockDelta?.delta?.text;
const reasoningToken =
chunk.contentBlockDelta?.delta?.reasoningContent?.text;
// Reasoning models will always return the reasoning text before the token text.
if (reasoningToken) {
// If the reasoning text is empty (''), we need to initialize it
// and send the first chunk of reasoning text.
if (reasoningText.length === 0) {
writeResponseChunk(response, {
uuid,
sources: [],
type: "textResponseChunk",
textResponse: `<think>${reasoningToken}`,
close: false,
error: false,
});
reasoningText += `<think>${reasoningToken}`;
continue;
} else {
writeResponseChunk(response, {
uuid,
sources: [],
type: "textResponseChunk",
textResponse: reasoningToken,
close: false,
error: false,
});
reasoningText += reasoningToken;
}
}
// If the reasoning text is not empty, but the reasoning token is empty
// and the token text is not empty we need to close the reasoning text and begin sending the token text.
if (!!reasoningText && !reasoningToken && token) {
writeResponseChunk(response, {
uuid,
sources: [],
type: "textResponseChunk",
textResponse: `</think>`,
close: false,
error: false,
});
fullText += `${reasoningText}</think>`;
reasoningText = "";
}
if (token) {
fullText += token;
// If we never saw a usage metric, we can estimate them by number of completion chunks
if (!hasUsageMetrics) usage.completion_tokens++;
writeResponseChunk(response, {
uuid,
sources: [],
type: "textResponseChunk",
textResponse: token,
close: false,
error: false,
});
}
}
}
writeResponseChunk(response, {
uuid,
sources,
type: "textResponseChunk",
textResponse: "",
close: true,
error: false,
});
response.removeListener("close", handleAbort);
stream?.endMeasurement(usage);
resolve(fullText);
} catch (error) {
console.log(`\x1b[43m\x1b[34m[STREAMING ERROR]\x1b[0m ${e.message}`);
writeResponseChunk(response, {
uuid,
type: "abort",
textResponse: null,
sources: [],
close: true,
error: `AWSBedrock:streaming - could not stream chat. ${error?.cause ?? error.message}`,
});
response.removeListener("close", handleAbort);
stream?.endMeasurement(usage);
resolve(fullText); // Return what we currently have - if anything.
}
});
}
// Simple wrapper for dynamic embedder & normalize interface for all LLM implementations
async embedTextInput(textInput) {
return await this.embedder.embedTextInput(textInput);
}
async embedChunks(textChunks = []) {
return await this.embedder.embedChunks(textChunks);
}
async compressMessages(promptArgs = {}, rawHistory = []) {
const { messageArrayCompressor } = require("../../helpers/chat");
const messageArray = this.constructPrompt(promptArgs);
return await messageArrayCompressor(this, messageArray, rawHistory);
}
}
module.exports = {
AWSBedrockLLM,
};