@saber2pr/ai-assistant 0.0.13 → 0.0.15

package/ReadMe.md

@@ -49,6 +49,9 @@ Initializes and mounts the AI assistant. This assistant runs entirely in the bro
 | `containerId` | `string` | `'ai-assistant-root'` | The ID of the container element where the assistant will be mounted. |
 | `initialPosition` | `{ x: number; y: number }` | Bottom-right | The initial coordinates of the floating button. |
 | `onBeforeChat` | `Function` | - | A hook to intercept and modify messages before they are sent to the AI. |
+| `maxMessages` | `number` | `20` | Maximum number of messages to keep in context. Older messages will be truncated. Set to 0 to disable. |
+| `maxContextLength` | `number` | `2500` | Maximum total characters in the context. Messages will be truncated from the oldest if exceeded. Set to 0 to disable. |
+| `maxTokens` | `number` | `3200` | Maximum estimated tokens in the context. This prevents exceeding the model's context window. For 4096 context window models, safe value is ~3200 (leaves ~900 tokens buffer). Set to 0 to disable. |
 
 ### Local Development
 

package/lib/app.js

@@ -5,23 +5,14 @@ var __importDefault = (this && this.__importDefault) || function (mod) {
 Object.defineProperty(exports, "__esModule", { value: true });
 var chat_1 = __importDefault(require("./chat"));
 var h1 = document.createElement('h1');
-h1.textContent = 'Click the bottom right sidebar button to open the AI assistant';
+h1.textContent =
+    'Click the bottom right sidebar button to open the AI assistant';
 document.body.append(h1);
 // 默认执行初始化
 (0, chat_1.default)({
-    locale: 'en-US',
+    locale: 'zh-CN',
     // welcomeMessage: '有什么可以帮忙的？',
     // suggestions: ['如何用 Typescript 实现 Helloworld？', '物联网是什么？'],
     // placeholder: '给 GPT 发送消息',
     // emptyMessage: '我是AI，可以回答你的问题，请在下方输入框输入你的需求～',
-    // async onBeforeChat(messages) {
-    //   const knowledgeContent = await fetch('http://localhost:5001/HTML超文本标记语言/移动端禁用双指放大.md').then(res => res.text())
-    //   return [
-    //     {
-    //       role: "system",
-    //       content: `你是我的博客助手，根据我博客内容回答：移动端禁用双指放大的方法：\n${knowledgeContent}`
-    //     },
-    //     ...messages
-    //   ]
-    // },
 });

package/lib/components/MyRuntimeProvider/index.js

@@ -13,7 +13,7 @@ function MyRuntimeProvider(_a) {
     var children = _a.children;
     var llm = (0, context_1.useLLm)();
     var config = react_1.default.useContext(context_2.AIConfigContext);
-    var runtime = (0, react_2.useLocalRuntime)((0, myModelAdapterStream_1.MyModelAdapterStream)(llm, config.onBeforeChat), {
+    var runtime = (0, react_2.useLocalRuntime)((0, myModelAdapterStream_1.MyModelAdapterStream)(llm, config.onBeforeChat, config.maxMessages, config.maxContextLength, config.maxTokens), {
         adapters: {
             speech: new react_2.WebSpeechSynthesisAdapter(),
         },

package/lib/components/MyRuntimeProvider/myModelAdapterStream.d.ts

@@ -1,4 +1,4 @@
 import { ChatModelAdapter } from '@assistant-ui/react';
 import { MLCEngine } from '@mlc-ai/web-llm';
 import { ChatMessage } from '../../types/assistant';
-export declare const MyModelAdapterStream: (llm: MLCEngine, onBeforeChat?: (messages: ChatMessage[], llm: MLCEngine) => ChatMessage[] | Promise<ChatMessage[]>) => ChatModelAdapter;
+export declare const MyModelAdapterStream: (llm: MLCEngine, onBeforeChat?: (messages: ChatMessage[], llm: MLCEngine) => ChatMessage[] | Promise<ChatMessage[]>, maxMessages?: number, maxContextLength?: number, maxTokens?: number) => ChatModelAdapter;

package/lib/components/MyRuntimeProvider/myModelAdapterStream.js

@@ -1,4 +1,15 @@
 "use strict";
+var __assign = (this && this.__assign) || function () {
+    __assign = Object.assign || function(t) {
+        for (var s, i = 1, n = arguments.length; i < n; i++) {
+            s = arguments[i];
+            for (var p in s) if (Object.prototype.hasOwnProperty.call(s, p))
+                t[p] = s[p];
+        }
+        return t;
+    };
+    return __assign.apply(this, arguments);
+};
 var __generator = (this && this.__generator) || function (thisArg, body) {
     var _ = { label: 0, sent: function() { if (t[0] & 1) throw t[1]; return t[1]; }, trys: [], ops: [] }, f, y, t, g = Object.create((typeof Iterator === "function" ? Iterator : Object).prototype);
     return g.next = verb(0), g["throw"] = verb(1), g["return"] = verb(2), typeof Symbol === "function" && (g[Symbol.iterator] = function() { return this; }), g;
@@ -46,52 +57,282 @@ var __asyncGenerator = (this && this.__asyncGenerator) || function (thisArg, _ar
     function reject(value) { resume("throw", value); }
     function settle(f, v) { if (f(v), q.shift(), q.length) resume(q[0][0], q[0][1]); }
 };
+var __spreadArray = (this && this.__spreadArray) || function (to, from, pack) {
+    if (pack || arguments.length === 2) for (var i = 0, l = from.length, ar; i < l; i++) {
+        if (ar || !(i in from)) {
+            if (!ar) ar = Array.prototype.slice.call(from, 0, i);
+            ar[i] = from[i];
+        }
+    }
+    return to.concat(ar || Array.prototype.slice.call(from));
+};
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.MyModelAdapterStream = void 0;
 var event_1 = require("../../utils/event");
 var constants_1 = require("../../constants");
-var MyModelAdapterStream = function (llm, onBeforeChat) { return ({
-    run: function (_a) {
-        return __asyncGenerator(this, arguments, function run_1(_b) {
-            var chatMessages, chunks, reply, _c, chunks_1, chunks_1_1, chunk, e_1_1, summaryResponse, title, error_1;
-            var _d, e_1, _e, _f;
-            var _g, _h, _j;
-            var messages = _b.messages, abortSignal = _b.abortSignal;
-            return __generator(this, function (_k) {
-                switch (_k.label) {
-                    case 0:
-                        chatMessages = messages.map(function (item) { return ({
-                            role: item.role,
-                            content: item.content[0].text,
-                        }); });
-                        if (!onBeforeChat) return [3 /*break*/, 2];
-                        return [4 /*yield*/, __await(onBeforeChat(chatMessages, llm))];
-                    case 1:
-                        chatMessages = _k.sent();
-                        _k.label = 2;
-                    case 2: return [4 /*yield*/, __await(llm.chat.completions.create({
-                            messages: chatMessages,
-                            temperature: 1,
-                            stream: true,
-                        }))];
-                    case 3:
-                        chunks = _k.sent();
-                        reply = "";
-                        _k.label = 4;
-                    case 4:
-                        _k.trys.push([4, 11, 12, 17]);
-                        _c = true, chunks_1 = __asyncValues(chunks);
-                        _k.label = 5;
-                    case 5: return [4 /*yield*/, __await(chunks_1.next())];
-                    case 6:
-                        if (!(chunks_1_1 = _k.sent(), _d = chunks_1_1.done, !_d)) return [3 /*break*/, 10];
-                        _f = chunks_1_1.value;
-                        _c = false;
-                        chunk = _f;
-                        reply += ((_g = chunk.choices[0]) === null || _g === void 0 ? void 0 : _g.delta.content) || "";
-                        return [4 /*yield*/, __await({
+/**
+ * Rough estimation: 1 token ≈ 1.3-1.5 characters for English, 1.5-2 for Chinese
+ * Using 1.5 as a conservative estimate
+ */
+function estimateTokens(text) {
+    // Rough estimate: Chinese characters count more, English less
+    var chineseChars = (text.match(/[\u4e00-\u9fa5]/g) || []).length;
+    var otherChars = text.length - chineseChars;
+    // Chinese: ~1.8 tokens per char, English: ~1.3 tokens per char
+    return Math.ceil(chineseChars * 1.8 + otherChars * 1.3);
+}
+/**
+ * Truncate messages based on maxMessages and maxContextLength
+ * Also considers token estimation to avoid exceeding context window
+ */
+function truncateMessages(messages, maxMessages, maxContextLength, maxTokens) {
+    var result = __spreadArray([], messages, true);
+    // Truncate by message count (keep the most recent messages)
+    if (maxMessages && maxMessages > 0 && result.length > maxMessages) {
+        // Always keep the first message if it's a system message
+        var firstMessage = result[0];
+        var isSystemFirst = firstMessage.role === 'system';
+        if (isSystemFirst) {
+            result = __spreadArray([firstMessage], result.slice(-(maxMessages - 1)), true);
+        }
+        else {
+            result = result.slice(-maxMessages);
+        }
+    }
+    // Truncate by total character length
+    if (maxContextLength && maxContextLength > 0) {
+        var firstMessage = result[0];
+        var isSystemFirst = firstMessage.role === 'system';
+        var otherMessages = isSystemFirst ? result.slice(1) : result;
+        // Calculate length of other messages
+        var otherMessagesLength = otherMessages.reduce(function (sum, msg) { return sum + msg.content.length; }, 0);
+        var systemMessageLength = isSystemFirst ? firstMessage.content.length : 0;
+        var totalLength = otherMessagesLength + systemMessageLength;
+        if (totalLength > maxContextLength) {
+            // Reserve 20% of maxContextLength for system message, or at least 500 chars
+            var systemMessageReserve = Math.max(500, Math.floor(maxContextLength * 0.2));
+            var availableForOther = maxContextLength - systemMessageReserve;
+            // Truncate system message if it's too long
+            var systemMessage = null;
+            if (isSystemFirst) {
+                if (firstMessage.content.length > systemMessageReserve) {
+                    // Keep the beginning of system message (usually contains important instructions)
+                    systemMessage = __assign(__assign({}, firstMessage), { content: firstMessage.content.substring(0, systemMessageReserve) + '...' });
+                }
+                else {
+                    systemMessage = firstMessage;
+                }
+            }
+            // Truncate other messages to fit in remaining space
+            var truncated = __spreadArray([], otherMessages, true);
+            var currentLength = truncated.reduce(function (sum, msg) { return sum + msg.content.length; }, 0);
+            var systemLength = systemMessage ? systemMessage.content.length : 0;
+            // Remove oldest messages until under limit
+            while (truncated.length > 0 && (currentLength + systemLength) > maxContextLength) {
+                var removed = truncated.shift();
+                if (removed) {
+                    currentLength -= removed.content.length;
+                }
+            }
+            // If still too long, truncate system message further
+            if (systemMessage && (currentLength + systemMessage.content.length) > maxContextLength) {
+                var maxSystemLength = Math.max(200, maxContextLength - currentLength - 100); // Leave some buffer
+                if (systemMessage.content.length > maxSystemLength) {
+                    systemMessage = __assign(__assign({}, systemMessage), { content: systemMessage.content.substring(0, maxSystemLength) + '...' });
+                }
+            }
+            result = systemMessage ? __spreadArray([systemMessage], truncated, true) : truncated;
+        }
+    }
+    // Additional token-based truncation as a safety net
+    // Default to 3500 tokens (leaving ~600 tokens buffer for 4096 context window)
+    var tokenLimit = maxTokens || 3500;
+    if (tokenLimit > 0) {
+        var totalTokens = result.reduce(function (sum, msg) { return sum + estimateTokens(msg.content); }, 0);
+        if (totalTokens > tokenLimit) {
+            var firstMessage = result[0];
+            var isSystemFirst = firstMessage.role === 'system';
+            var systemMessage = isSystemFirst ? firstMessage : null;
+            var otherMessages = isSystemFirst ? result.slice(1) : result;
+            // Reserve tokens for system message (max 800 tokens)
+            var systemTokens = systemMessage ? Math.min(estimateTokens(systemMessage.content), 800) : 0;
+            var availableTokens = tokenLimit - systemTokens;
+            // Truncate other messages
+            var truncated = __spreadArray([], otherMessages, true);
+            var currentTokens = truncated.reduce(function (sum, msg) { return sum + estimateTokens(msg.content); }, 0);
+            // Remove oldest messages until under token limit
+            while (truncated.length > 0 && (currentTokens + systemTokens) > tokenLimit) {
+                var removed = truncated.shift();
+                if (removed) {
+                    currentTokens -= estimateTokens(removed.content);
+                }
+            }
+            // Truncate system message if still needed
+            var finalSystemMessage = systemMessage;
+            if (systemMessage && (currentTokens + systemTokens) > tokenLimit) {
+                var maxSystemTokens = Math.max(200, tokenLimit - currentTokens - 100);
+                var systemContent = systemMessage.content;
+                var systemContentTokens = estimateTokens(systemContent);
+                if (systemContentTokens > maxSystemTokens) {
+                    // Binary search for approximate length
+                    var low = 0;
+                    var high = systemContent.length;
+                    while (low < high) {
+                        var mid = Math.floor((low + high) / 2);
+                        var testContent = systemContent.substring(0, mid);
+                        if (estimateTokens(testContent) <= maxSystemTokens) {
+                            low = mid + 1;
+                        }
+                        else {
+                            high = mid;
+                        }
+                    }
+                    systemContent = systemContent.substring(0, Math.max(0, low - 1)) + '...';
+                }
+                finalSystemMessage = __assign(__assign({}, systemMessage), { content: systemContent });
+            }
+            result = finalSystemMessage ? __spreadArray([finalSystemMessage], truncated, true) : truncated;
+        }
+    }
+    return result;
+}
+var MyModelAdapterStream = function (llm, onBeforeChat, maxMessages, maxContextLength, maxTokens) {
+    if (maxMessages === void 0) { maxMessages = 20; }
+    if (maxContextLength === void 0) { maxContextLength = 2500; }
+    if (maxTokens === void 0) { maxTokens = 3200; }
+    return ({
+        run: function (_a) {
+            return __asyncGenerator(this, arguments, function run_1(_b) {
+                var chatMessages, finalTokenLimit, totalTokens, firstMessage, isSystemFirst, systemMessage, otherMessages, finalSystemMessage, systemTokens_1, low, high, mid, testContent, systemTokens, availableTokens, truncated, currentTokens, i, msg, msgTokens, chunks, reply, _c, chunks_1, chunks_1_1, chunk, e_1_1, summaryResponse, title, error_1;
+                var _d, e_1, _e, _f;
+                var _g, _h, _j;
+                var messages = _b.messages, abortSignal = _b.abortSignal;
+                return __generator(this, function (_k) {
+                    switch (_k.label) {
+                        case 0:
+                            chatMessages = messages.map(function (item) { return ({
+                                role: item.role,
+                                content: item.content[0].text,
+                            }); });
+                            // Apply truncation before onBeforeChat hook
+                            chatMessages = truncateMessages(chatMessages, maxMessages, maxContextLength, maxTokens);
+                            if (!onBeforeChat) return [3 /*break*/, 2];
+                            return [4 /*yield*/, __await(onBeforeChat(chatMessages, llm)
+                                // Re-apply truncation after onBeforeChat in case it added more content
+                                // Use stricter limits to ensure we don't exceed context window
+                                )];
+                        case 1:
+                            chatMessages = _k.sent();
+                            // Re-apply truncation after onBeforeChat in case it added more content
+                            // Use stricter limits to ensure we don't exceed context window
+                            chatMessages = truncateMessages(chatMessages, maxMessages, maxContextLength, maxTokens);
+                            _k.label = 2;
+                        case 2:
+                            finalTokenLimit = maxTokens || 3200;
+                            if (finalTokenLimit > 0) {
+                                totalTokens = chatMessages.reduce(function (sum, msg) { return sum + estimateTokens(msg.content); }, 0);
+                                if (totalTokens > finalTokenLimit) {
+                                    firstMessage = chatMessages[0];
+                                    isSystemFirst = (firstMessage === null || firstMessage === void 0 ? void 0 : firstMessage.role) === 'system';
+                                    systemMessage = isSystemFirst ? firstMessage : null;
+                                    otherMessages = isSystemFirst ? chatMessages.slice(1) : chatMessages;
+                                    finalSystemMessage = systemMessage;
+                                    if (systemMessage) {
+                                        systemTokens_1 = estimateTokens(systemMessage.content);
+                                        if (systemTokens_1 > 500) {
+                                            low = 0;
+                                            high = systemMessage.content.length;
+                                            while (low < high) {
+                                                mid = Math.floor((low + high) / 2);
+                                                testContent = systemMessage.content.substring(0, mid);
+                                                if (estimateTokens(testContent) <= 500) {
+                                                    low = mid + 1;
+                                                }
+                                                else {
+                                                    high = mid;
+                                                }
+                                            }
+                                            finalSystemMessage = __assign(__assign({}, systemMessage), { content: systemMessage.content.substring(0, Math.max(0, low - 1)) + '...' });
+                                        }
+                                    }
+                                    systemTokens = finalSystemMessage ? estimateTokens(finalSystemMessage.content) : 0;
+                                    availableTokens = finalTokenLimit - systemTokens;
+                                    truncated = [];
+                                    currentTokens = 0;
+                                    // Add messages from newest to oldest until we hit the limit
+                                    for (i = otherMessages.length - 1; i >= 0; i--) {
+                                        msg = otherMessages[i];
+                                        msgTokens = estimateTokens(msg.content);
+                                        if (currentTokens + msgTokens <= availableTokens) {
+                                            truncated.unshift(msg);
+                                            currentTokens += msgTokens;
+                                        }
+                                        else {
+                                            break;
+                                        }
+                                    }
+                                    chatMessages = finalSystemMessage ? __spreadArray([finalSystemMessage], truncated, true) : truncated;
+                                }
+                            }
+                            return [4 /*yield*/, __await(llm.chat.completions.create({
+                                    messages: chatMessages,
+                                    temperature: 1,
+                                    stream: true,
+                                }))];
+                        case 3:
+                            chunks = _k.sent();
+                            reply = "";
+                            _k.label = 4;
+                        case 4:
+                            _k.trys.push([4, 11, 12, 17]);
+                            _c = true, chunks_1 = __asyncValues(chunks);
+                            _k.label = 5;
+                        case 5: return [4 /*yield*/, __await(chunks_1.next())];
+                        case 6:
+                            if (!(chunks_1_1 = _k.sent(), _d = chunks_1_1.done, !_d)) return [3 /*break*/, 10];
+                            _f = chunks_1_1.value;
+                            _c = false;
+                            chunk = _f;
+                            reply += ((_g = chunk.choices[0]) === null || _g === void 0 ? void 0 : _g.delta.content) || "";
+                            return [4 /*yield*/, __await({
+                                    status: {
+                                        type: 'running',
+                                    },
+                                    content: [
+                                        {
+                                            text: reply,
+                                            type: 'text',
+                                        },
+                                    ],
+                                })];
+                        case 7: return [4 /*yield*/, _k.sent()];
+                        case 8:
+                            _k.sent();
+                            _k.label = 9;
+                        case 9:
+                            _c = true;
+                            return [3 /*break*/, 5];
+                        case 10: return [3 /*break*/, 17];
+                        case 11:
+                            e_1_1 = _k.sent();
+                            e_1 = { error: e_1_1 };
+                            return [3 /*break*/, 17];
+                        case 12:
+                            _k.trys.push([12, , 15, 16]);
+                            if (!(!_c && !_d && (_e = chunks_1.return))) return [3 /*break*/, 14];
+                            return [4 /*yield*/, __await(_e.call(chunks_1))];
+                        case 13:
+                            _k.sent();
+                            _k.label = 14;
+                        case 14: return [3 /*break*/, 16];
+                        case 15:
+                            if (e_1) throw e_1.error;
+                            return [7 /*endfinally*/];
+                        case 16: return [7 /*endfinally*/];
+                        case 17: return [4 /*yield*/, __await({
                                 status: {
-                                    type: 'running',
+                                    type: 'complete',
+                                    reason: 'stop',
                                 },
                                 content: [
                                     {
@@ -99,80 +340,45 @@ var MyModelAdapterStream = function (llm, onBeforeChat) { return ({
                                         type: 'text',
                                     },
                                 ],
-                            })];
-                    case 7: return [4 /*yield*/, _k.sent()];
-                    case 8:
-                        _k.sent();
-                        _k.label = 9;
-                    case 9:
-                        _c = true;
-                        return [3 /*break*/, 5];
-                    case 10: return [3 /*break*/, 17];
-                    case 11:
-                        e_1_1 = _k.sent();
-                        e_1 = { error: e_1_1 };
-                        return [3 /*break*/, 17];
-                    case 12:
-                        _k.trys.push([12, , 15, 16]);
-                        if (!(!_c && !_d && (_e = chunks_1.return))) return [3 /*break*/, 14];
-                        return [4 /*yield*/, __await(_e.call(chunks_1))];
-                    case 13:
-                        _k.sent();
-                        _k.label = 14;
-                    case 14: return [3 /*break*/, 16];
-                    case 15:
-                        if (e_1) throw e_1.error;
-                        return [7 /*endfinally*/];
-                    case 16: return [7 /*endfinally*/];
-                    case 17: return [4 /*yield*/, __await({
-                            status: {
-                                type: 'complete',
-                                reason: 'stop',
-                            },
-                            content: [
-                                {
-                                    text: reply,
-                                    type: 'text',
-                                },
-                            ],
-                        }
-                        // 对话完成后，如果这是第一轮对话，生成标题
-                        )];
-                    case 18: return [4 /*yield*/, _k.sent()];
-                    case 19:
-                        _k.sent();
-                        if (!(messages.length === 1)) return [3 /*break*/, 23];
-                        _k.label = 20;
-                    case 20:
-                        _k.trys.push([20, 22, , 23]);
-                        return [4 /*yield*/, __await(llm.chat.completions.create({
-                                messages: [
-                                    {
-                                        role: 'system',
-                                        content: 'You are a title generation assistant. Please summarize a short title (no more than 10 words) based on the user\'s input, without punctuation.'
-                                    },
-                                    {
-                                        role: 'user',
-                                        content: messages[0].content[0].text
-                                    }
-                                ],
-                                temperature: 0.5,
-                            }))];
-                    case 21:
-                        summaryResponse = _k.sent();
-                        title = (_j = (_h = summaryResponse.choices[0]) === null || _h === void 0 ? void 0 : _h.message.content) === null || _j === void 0 ? void 0 : _j.trim();
-                        if (title) {
-                            event_1.Dispatcher.instance.dispatch(constants_1.EVENT_THREAD_SET_TITLE, { data: title });
-                        }
-                        return [3 /*break*/, 23];
-                    case 22:
-                        error_1 = _k.sent();
-                        console.error('Failed to generate summary title:', error_1);
-                        return [3 /*break*/, 23];
-                    case 23: return [2 /*return*/];
-                }
+                            }
+                            // 对话完成后，如果这是第一轮对话，生成标题
+                            )];
+                        case 18: return [4 /*yield*/, _k.sent()];
+                        case 19:
+                            _k.sent();
+                            if (!(messages.length === 1)) return [3 /*break*/, 23];
+                            _k.label = 20;
+                        case 20:
+                            _k.trys.push([20, 22, , 23]);
+                            return [4 /*yield*/, __await(llm.chat.completions.create({
+                                    messages: [
+                                        {
+                                            role: 'system',
+                                            content: 'You are a title generation assistant. Please summarize a short title (no more than 10 words) based on the user\'s input, without punctuation.'
+                                        },
+                                        {
+                                            role: 'user',
+                                            content: messages[0].content[0].text
+                                        }
+                                    ],
+                                    temperature: 0.5,
+                                }))];
+                        case 21:
+                            summaryResponse = _k.sent();
+                            title = (_j = (_h = summaryResponse.choices[0]) === null || _h === void 0 ? void 0 : _h.message.content) === null || _j === void 0 ? void 0 : _j.trim();
+                            if (title) {
+                                event_1.Dispatcher.instance.dispatch(constants_1.EVENT_THREAD_SET_TITLE, { data: title });
+                            }
+                            return [3 /*break*/, 23];
+                        case 22:
+                            error_1 = _k.sent();
+                            console.error('Failed to generate summary title:', error_1);
+                            return [3 /*break*/, 23];
+                        case 23: return [2 /*return*/];
+                    }
+                });
             });
-        });
-    },
-}); };
+        },
+    });
+};
 exports.MyModelAdapterStream = MyModelAdapterStream;

package/lib/types/assistant.d.ts

@@ -55,4 +55,25 @@ export interface AIAssistantConfig {
      * }
      */
     onBeforeChat?: (messages: ChatMessage[], llm: MLCEngine) => ChatMessage[] | Promise<ChatMessage[]>;
+    /**
+     * Maximum number of messages to keep in context. Older messages will be truncated.
+     * Set to 0 or undefined to disable truncation.
+     * @default 20
+     */
+    maxMessages?: number;
+    /**
+     * Maximum total characters in the context. Messages will be truncated from the oldest if exceeded.
+     * Note: This is a rough estimate. Actual token count may vary. For models with 4096 context window,
+     * a safe value is around 2500-3000 characters.
+     * Set to 0 or undefined to disable truncation.
+     * @default 2500
+     */
+    maxContextLength?: number;
+    /**
+     * Maximum estimated tokens in the context. This is a safety net to prevent exceeding model's context window.
+     * For models with 4096 context window, a safe value is around 3200 tokens (leaving ~900 tokens buffer for response and overhead).
+     * Set to 0 or undefined to disable token-based truncation.
+     * @default 3200
+     */
+    maxTokens?: number;
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@saber2pr/ai-assistant",
-  "version": "0.0.13",
+  "version": "0.0.15",
   "description": "AI Assistant",
   "files": [
     "lib"