@lobehub/chat 1.97.11 → 1.97.12

package/src/database/migrations/meta/_journal.json

@@ -182,6 +182,13 @@
       "when": 1749309388370,
       "tag": "0025_add_provider_config",
       "breakpoints": true
+    },
+    {
+      "idx": 26,
+      "version": "7",
+      "when": 1752212281564,
+      "tag": "0026_add_autovacuum_tuning",
+      "breakpoints": true
     }
   ],
   "version": "6"

package/src/libs/model-runtime/bedrock/index.ts

@@ -3,7 +3,6 @@ import {
   InvokeModelCommand,
   InvokeModelWithResponseStreamCommand,
 } from '@aws-sdk/client-bedrock-runtime';
-import { experimental_buildLlama2Prompt } from 'ai/prompts';
 
 import { LobeRuntimeAI } from '../BaseAI';
 import { AgentRuntimeErrorType } from '../error';
@@ -25,6 +24,38 @@ import {
   createBedrockStream,
 } from '../utils/streams';
 
+/**
+ * A prompt constructor for HuggingFace LLama 2 chat models.
+ * Does not support `function` messages.
+ * @see https://huggingface.co/meta-llama/Llama-2-70b-chat-hf and https://huggingface.co/blog/llama2#how-to-prompt-llama-2
+ */
+export function experimental_buildLlama2Prompt(messages: { content: string; role: string }[]) {
+  const startPrompt = `<s>[INST] `;
+  const endPrompt = ` [/INST]`;
+  const conversation = messages.map(({ content, role }, index) => {
+    switch (role) {
+      case 'user': {
+        return content.trim();
+      }
+      case 'assistant': {
+        return ` [/INST] ${content}</s><s>[INST] `;
+      }
+      case 'function': {
+        throw new Error('Llama 2 does not support function calls.');
+      }
+      default: {
+        if (role === 'system' && index === 0) {
+          return `<<SYS>>\n${content}\n<</SYS>>\n\n`;
+        } else {
+          throw new Error(`Invalid message role: ${role}`);
+        }
+      }
+    }
+  });
+
+  return startPrompt + conversation.join('') + endPrompt;
+}
+
 export interface LobeBedrockAIParams {
   accessKeyId?: string;
   accessKeySecret?: string;

package/src/libs/model-runtime/utils/streams/bedrock/common.ts

@@ -2,7 +2,8 @@ import {
   InvokeModelWithResponseStreamResponse,
   ResponseStream,
 } from '@aws-sdk/client-bedrock-runtime';
-import { readableFromAsyncIterable } from 'ai';
+
+import { readableFromAsyncIterable } from '../protocol';
 
 const chatStreamable = async function* (stream: AsyncIterable<ResponseStream>) {
   for await (const response of stream) {

package/src/libs/model-runtime/utils/streams/protocol.ts

@@ -108,6 +108,22 @@ const chatStreamable = async function* <T>(stream: AsyncIterable<T>) {
 };
 
 const ERROR_CHUNK_PREFIX = '%FIRST_CHUNK_ERROR%: ';
+
+export function readableFromAsyncIterable<T>(iterable: AsyncIterable<T>) {
+  let it = iterable[Symbol.asyncIterator]();
+  return new ReadableStream<T>({
+    async cancel(reason) {
+      await it.return?.(reason);
+    },
+
+    async pull(controller) {
+      const { done, value } = await it.next();
+      if (done) controller.close();
+      else controller.enqueue(value);
+    },
+  });
+}
+
 // make the response to the streamable format
 export const convertIterableToStream = <T>(stream: AsyncIterable<T>) => {
   const iterable = chatStreamable(stream);

package/src/libs/model-runtime/utils/streams/qwen.ts

@@ -1,6 +1,8 @@
-import { ChatCompletionContentPartText } from 'ai/prompts';
 import OpenAI from 'openai';
-import { ChatCompletionContentPart } from 'openai/resources/index.mjs';
+import {
+  ChatCompletionContentPart,
+  ChatCompletionContentPartText,
+} from 'openai/resources/index.mjs';
 import type { Stream } from 'openai/streaming';
 
 import { ChatStreamCallbacks } from '../../types';

package/src/libs/model-runtime/xai/index.ts

@@ -9,9 +9,12 @@ export interface XAIModelCard {
 
 export const GrokReasoningModels = new Set([
   'grok-3-mini',
-  'grok-4-0709',
+  'grok-4',
 ]);
 
+export const isGrokReasoningModel = (model: string) =>
+  Array.from(GrokReasoningModels).some((id) => model.includes(id));
+
 export const LobeXAI = createOpenAICompatibleRuntime({
   baseURL: 'https://api.x.ai/v1',
   chatCompletion: {
@@ -20,9 +23,9 @@ export const LobeXAI = createOpenAICompatibleRuntime({
 
       return {
         ...rest,
-        frequency_penalty: GrokReasoningModels.has(model) ? undefined : frequency_penalty,
+        frequency_penalty: isGrokReasoningModel(model) ? undefined : frequency_penalty,
         model,
-        presence_penalty: GrokReasoningModels.has(model) ? undefined : presence_penalty,
+        presence_penalty: isGrokReasoningModel(model) ? undefined : presence_penalty,
         stream: true,
         ...(enabledSearch && {
           search_parameters: {