opengauge 0.1.0

package/dist/core/providers/openai.js

@@ -0,0 +1,169 @@
+"use strict";
+var __createBinding = (this && this.__createBinding) || (Object.create ? (function(o, m, k, k2) {
+    if (k2 === undefined) k2 = k;
+    var desc = Object.getOwnPropertyDescriptor(m, k);
+    if (!desc || ("get" in desc ? !m.__esModule : desc.writable || desc.configurable)) {
+      desc = { enumerable: true, get: function() { return m[k]; } };
+    }
+    Object.defineProperty(o, k2, desc);
+}) : (function(o, m, k, k2) {
+    if (k2 === undefined) k2 = k;
+    o[k2] = m[k];
+}));
+var __setModuleDefault = (this && this.__setModuleDefault) || (Object.create ? (function(o, v) {
+    Object.defineProperty(o, "default", { enumerable: true, value: v });
+}) : function(o, v) {
+    o["default"] = v;
+});
+var __importStar = (this && this.__importStar) || (function () {
+    var ownKeys = function(o) {
+        ownKeys = Object.getOwnPropertyNames || function (o) {
+            var ar = [];
+            for (var k in o) if (Object.prototype.hasOwnProperty.call(o, k)) ar[ar.length] = k;
+            return ar;
+        };
+        return ownKeys(o);
+    };
+    return function (mod) {
+        if (mod && mod.__esModule) return mod;
+        var result = {};
+        if (mod != null) for (var k = ownKeys(mod), i = 0; i < k.length; i++) if (k[i] !== "default") __createBinding(result, mod, k[i]);
+        __setModuleDefault(result, mod);
+        return result;
+    };
+})();
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.OpenAIProvider = void 0;
+class OpenAIProvider {
+    name = 'openai';
+    defaultModel;
+    apiKey;
+    baseUrl;
+    constructor(config) {
+        this.apiKey = config.api_key || '';
+        this.baseUrl = config.base_url || 'https://api.openai.com';
+        this.defaultModel = config.default_model || 'gpt-4o';
+    }
+    async chat(request) {
+        const model = request.model || this.defaultModel;
+        const body = {
+            model,
+            messages: request.messages.map((m) => ({
+                role: m.role,
+                content: m.content,
+            })),
+            max_tokens: request.maxTokens || 4096,
+        };
+        if (request.temperature !== undefined) {
+            body.temperature = request.temperature;
+        }
+        const response = await fetch(`${this.baseUrl}/v1/chat/completions`, {
+            method: 'POST',
+            headers: {
+                'Content-Type': 'application/json',
+                Authorization: `Bearer ${this.apiKey}`,
+            },
+            body: JSON.stringify(body),
+        });
+        if (!response.ok) {
+            const err = await response.text();
+            throw new Error(`OpenAI API error (${response.status}): ${err}`);
+        }
+        const data = await response.json();
+        const content = data.choices?.[0]?.message?.content || '';
+        return {
+            content,
+            tokensIn: data.usage?.prompt_tokens || 0,
+            tokensOut: data.usage?.completion_tokens || 0,
+            model,
+            provider: this.name,
+        };
+    }
+    async *chatStream(request) {
+        const model = request.model || this.defaultModel;
+        const body = {
+            model,
+            messages: request.messages.map((m) => ({
+                role: m.role,
+                content: m.content,
+            })),
+            max_tokens: request.maxTokens || 4096,
+            stream: true,
+            stream_options: { include_usage: true },
+        };
+        if (request.temperature !== undefined) {
+            body.temperature = request.temperature;
+        }
+        const response = await fetch(`${this.baseUrl}/v1/chat/completions`, {
+            method: 'POST',
+            headers: {
+                'Content-Type': 'application/json',
+                Authorization: `Bearer ${this.apiKey}`,
+            },
+            body: JSON.stringify(body),
+        });
+        if (!response.ok) {
+            const err = await response.text();
+            throw new Error(`OpenAI API error (${response.status}): ${err}`);
+        }
+        const reader = response.body?.getReader();
+        if (!reader)
+            throw new Error('No response body');
+        const decoder = new TextDecoder();
+        let buffer = '';
+        let tokensIn = 0;
+        let tokensOut = 0;
+        try {
+            while (true) {
+                const { done, value } = await reader.read();
+                if (done)
+                    break;
+                buffer += decoder.decode(value, { stream: true });
+                const lines = buffer.split('\n');
+                buffer = lines.pop() || '';
+                for (const line of lines) {
+                    if (!line.startsWith('data: '))
+                        continue;
+                    const data = line.slice(6).trim();
+                    if (data === '[DONE]')
+                        continue;
+                    try {
+                        const event = JSON.parse(data);
+                        const delta = event.choices?.[0]?.delta?.content;
+                        if (delta) {
+                            yield { content: delta, done: false };
+                        }
+                        if (event.usage) {
+                            tokensIn = event.usage.prompt_tokens || 0;
+                            tokensOut = event.usage.completion_tokens || 0;
+                        }
+                    }
+                    catch {
+                        // Skip malformed JSON
+                    }
+                }
+            }
+        }
+        finally {
+            reader.releaseLock();
+        }
+        yield { content: '', done: true, tokensIn, tokensOut };
+    }
+    async countTokens(text) {
+        // Use tiktoken for accurate OpenAI token counting
+        try {
+            const { encoding_for_model } = await Promise.resolve().then(() => __importStar(require('tiktoken')));
+            const enc = encoding_for_model('gpt-4o');
+            const tokens = enc.encode(text);
+            const count = tokens.length;
+            enc.free();
+            return count;
+        }
+        catch {
+            // Fallback: ~4 chars per token
+            return Math.ceil(text.length / 4);
+        }
+    }
+}
+exports.OpenAIProvider = OpenAIProvider;
+//# sourceMappingURL=openai.js.map

package/dist/core/providers/openai.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"openai.js","sourceRoot":"","sources":["../../../src/core/providers/openai.ts"],"names":[],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAQA,MAAa,cAAc;IACzB,IAAI,GAAG,QAAQ,CAAC;IAChB,YAAY,CAAS;IACb,MAAM,CAAS;IACf,OAAO,CAAS;IAExB,YAAY,MAAsB;QAChC,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC,OAAO,IAAI,EAAE,CAAC;QACnC,IAAI,CAAC,OAAO,GAAG,MAAM,CAAC,QAAQ,IAAI,wBAAwB,CAAC;QAC3D,IAAI,CAAC,YAAY,GAAG,MAAM,CAAC,aAAa,IAAI,QAAQ,CAAC;IACvD,CAAC;IAED,KAAK,CAAC,IAAI,CAAC,OAAoB;QAC7B,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,IAAI,IAAI,CAAC,YAAY,CAAC;QAEjD,MAAM,IAAI,GAAQ;YAChB,KAAK;YACL,QAAQ,EAAE,OAAO,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;gBACrC,IAAI,EAAE,CAAC,CAAC,IAAI;gBACZ,OAAO,EAAE,CAAC,CAAC,OAAO;aACnB,CAAC,CAAC;YACH,UAAU,EAAE,OAAO,CAAC,SAAS,IAAI,IAAI;SACtC,CAAC;QAEF,IAAI,OAAO,CAAC,WAAW,KAAK,SAAS,EAAE,CAAC;YACtC,IAAI,CAAC,WAAW,GAAG,OAAO,CAAC,WAAW,CAAC;QACzC,CAAC;QAED,MAAM,QAAQ,GAAG,MAAM,KAAK,CAAC,GAAG,IAAI,CAAC,OAAO,sBAAsB,EAAE;YAClE,MAAM,EAAE,MAAM;YACd,OAAO,EAAE;gBACP,cAAc,EAAE,kBAAkB;gBAClC,aAAa,EAAE,UAAU,IAAI,CAAC,MAAM,EAAE;aACvC;YACD,IAAI,EAAE,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC;SAC3B,CAAC,CAAC;QAEH,IAAI,CAAC,QAAQ,CAAC,EAAE,EAAE,CAAC;YACjB,MAAM,GAAG,GAAG,MAAM,QAAQ,CAAC,IAAI,EAAE,CAAC;YAClC,MAAM,IAAI,KAAK,CAAC,qBAAqB,QAAQ,CAAC,MAAM,MAAM,GAAG,EAAE,CAAC,CAAC;QACnE,CAAC;QAED,MAAM,IAAI,GAAG,MAAM,QAAQ,CAAC,IAAI,EAAS,CAAC;QAC1C,MAAM,OAAO,GAAG,IAAI,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,OAAO,EAAE,OAAO,IAAI,EAAE,CAAC;QAE1D,OAAO;YACL,OAAO;YACP,QAAQ,EAAE,IAAI,CAAC,KAAK,EAAE,aAAa,IAAI,CAAC;YACxC,SAAS,EAAE,IAAI,CAAC,KAAK,EAAE,iBAAiB,IAAI,CAAC;YAC7C,KAAK;YACL,QAAQ,EAAE,IAAI,CAAC,IAAI;SACpB,CAAC;IACJ,CAAC;IAED,KAAK,CAAC,CAAC,UAAU,CAAC,OAAoB;QACpC,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,IAAI,IAAI,CAAC,YAAY,CAAC;QAEjD,MAAM,IAAI,GAAQ;YAChB,KAAK;YACL,QAAQ,EAAE,OAAO,CAAC,QAAQ,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;gBACrC,IAAI,EAAE,CAAC,CAAC,IAAI;gBACZ,OAAO,EAAE,CAAC,CAAC,OAAO;aACnB,CAAC,CAAC;YACH,UAAU,EAAE,OAAO,CAAC,SAAS,IAAI,IAAI;YACrC,MAAM,EAAE,IAAI;YACZ,cAAc,EAAE,EAAE,aAAa,EAAE,IAAI,EAAE;SACxC,CAAC;QAEF,IAAI,OAAO,CAAC,WAAW,KAAK,SAAS,EAAE,CAAC;YACtC,IAAI,CAAC,WAAW,GAAG,OAAO,CAAC,WAAW,CAAC;QACzC,CAAC;QAED,MAAM,QAAQ,GAAG,MAAM,KAAK,CAAC,GAAG,IAAI,CAAC,OAAO,sBAAsB,EAAE;YAClE,MAAM,EAAE,MAAM;YACd,OAAO,EAAE;gBACP,cAAc,EAAE,kBAAkB;gBAClC,aAAa,EAAE,UAAU,IAAI,CAAC,MAAM,EAAE;aACvC;YACD,IAAI,EAAE,IAAI,CAAC,SAAS,CAAC,IAAI,CAAC;SAC3B,CAAC,CAAC;QAEH,IAAI,CAAC,QAAQ,CAAC,EAAE,EAAE,CAAC;YACjB,MAAM,GAAG,GAAG,MAAM,QAAQ,CAAC,IAAI,EAAE,CAAC;YAClC,MAAM,IAAI,KAAK,CAAC,qBAAqB,QAAQ,CAAC,MAAM,MAAM,GAAG,EAAE,CAAC,CAAC;QACnE,CAAC;QAED,MAAM,MAAM,GAAG,QAAQ,CAAC,IAAI,EAAE,SAAS,EAAE,CAAC;QAC1C,IAAI,CAAC,MAAM;YAAE,MAAM,IAAI,KAAK,CAAC,kBAAkB,CAAC,CAAC;QAEjD,MAAM,OAAO,GAAG,IAAI,WAAW,EAAE,CAAC;QAClC,IAAI,MAAM,GAAG,EAAE,CAAC;QAChB,IAAI,QAAQ,GAAG,CAAC,CAAC;QACjB,IAAI,SAAS,GAAG,CAAC,CAAC;QAElB,IAAI,CAAC;YACH,OAAO,IAAI,EAAE,CAAC;gBACZ,MAAM,EAAE,IAAI,EAAE,KAAK,EAAE,GAAG,MAAM,MAAM,CAAC,IAAI,EAAE,CAAC;gBAC5C,IAAI,IAAI;oBAAE,MAAM;gBAEhB,MAAM,IAAI,OAAO,CAAC,MAAM,CAAC,KAAK,EAAE,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC;gBAClD,MAAM,KAAK,GAAG,MAAM,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;gBACjC,MAAM,GAAG,KAAK,CAAC,GAAG,EAAE,IAAI,EAAE,CAAC;gBAE3B,KAAK,MAAM,IAAI,IAAI,KAAK,EAAE,CAAC;oBACzB,IAAI,CAAC,IAAI,CAAC,UAAU,CAAC,QAAQ,CAAC;wBAAE,SAAS;oBACzC,MAAM,IAAI,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC;oBAClC,IAAI,IAAI,KAAK,QAAQ;wBAAE,SAAS;oBAEhC,IAAI,CAAC;wBACH,MAAM,KAAK,GAAG,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;wBAC/B,MAAM,KAAK,GAAG,KAAK,CAAC,OAAO,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,OAAO,CAAC;wBAEjD,IAAI,KAAK,EAAE,CAAC;4BACV,MAAM,EAAE,OAAO,EAAE,KAAK,EAAE,IAAI,EAAE,KAAK,EAAE,CAAC;wBACxC,CAAC;wBAED,IAAI,KAAK,CAAC,KAAK,EAAE,CAAC;4BAChB,QAAQ,GAAG,KAAK,CAAC,KAAK,CAAC,aAAa,IAAI,CAAC,CAAC;4BAC1C,SAAS,GAAG,KAAK,CAAC,KAAK,CAAC,iBAAiB,IAAI,CAAC,CAAC;wBACjD,CAAC;oBACH,CAAC;oBAAC,MAAM,CAAC;wBACP,sBAAsB;oBACxB,CAAC;gBACH,CAAC;YACH,CAAC;QACH,CAAC;gBAAS,CAAC;YACT,MAAM,CAAC,WAAW,EAAE,CAAC;QACvB,CAAC;QAED,MAAM,EAAE,OAAO,EAAE,EAAE,EAAE,IAAI,EAAE,IAAI,EAAE,QAAQ,EAAE,SAAS,EAAE,CAAC;IACzD,CAAC;IAED,KAAK,CAAC,WAAW,CAAC,IAAY;QAC5B,kDAAkD;QAClD,IAAI,CAAC;YACH,MAAM,EAAE,kBAAkB,EAAE,GAAG,wDAAa,UAAU,GAAC,CAAC;YACxD,MAAM,GAAG,GAAG,kBAAkB,CAAC,QAAe,CAAC,CAAC;YAChD,MAAM,MAAM,GAAG,GAAG,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC;YAChC,MAAM,KAAK,GAAG,MAAM,CAAC,MAAM,CAAC;YAC5B,GAAG,CAAC,IAAI,EAAE,CAAC;YACX,OAAO,KAAK,CAAC;QACf,CAAC;QAAC,MAAM,CAAC;YACP,+BAA+B;YAC/B,OAAO,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;QACpC,CAAC;IACH,CAAC;CACF;AAlJD,wCAkJC"}

package/dist/core/rag/assembler.d.ts

@@ -0,0 +1,47 @@
+/**
+ * Context Assembler — Builds the optimal context window for LLM requests
+ *
+ * Combines:
+ * 1. System prompt
+ * 2. Retrieved relevant context (RAG)
+ * 3. Last N verbatim turns
+ * 4. Compressed checkpoint (if exists)
+ *
+ * Applies token budget to ensure we fit within the model's context window.
+ */
+import { ChatMessage } from '../providers/adapter';
+import { Queries } from '../../db/queries';
+export interface AssemblerConfig {
+    /** Maximum tokens for the context window. Default: 8000 */
+    maxContextTokens: number;
+    /** Tokens reserved for the model's response. Default: 2000 */
+    responseReserve: number;
+    /** Number of recent messages to keep verbatim. Default: 10 */
+    recentMessageCount: number;
+    /** Number of RAG results to retrieve. Default: 15 */
+    ragTopK: number;
+    /** Compression aggressiveness (0-1). Default: 0.3 */
+    compressionLevel: number;
+    /** Desired savings percentage against raw prompt tokens. Default: 50 */
+    targetSavingsPercent: number;
+    /** Minimum token budget to preserve quality/context intelligence. Default: 320 */
+    qualityFloorTokens: number;
+}
+export declare const DEFAULT_ASSEMBLER_CONFIG: AssemblerConfig;
+/**
+ * Assemble the context window for an LLM request.
+ *
+ * @param userMessage - The current user message.
+ * @param conversationId - The conversation ID.
+ * @param systemPrompt - The system prompt.
+ * @param queries - Database queries instance.
+ * @param checkpointSummary - Latest checkpoint summary, if any.
+ * @param config - Assembler configuration.
+ */
+export declare function assembleContext(userMessage: string, conversationId: string, systemPrompt: string | null, queries: Queries, checkpointSummary: string | null, config?: AssemblerConfig): Promise<{
+    messages: ChatMessage[];
+    tokensRaw: number;
+    tokensSent: number;
+    ragResultCount: number;
+}>;
+//# sourceMappingURL=assembler.d.ts.map

package/dist/core/rag/assembler.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"assembler.d.ts","sourceRoot":"","sources":["../../../src/core/rag/assembler.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,OAAO,EAAE,WAAW,EAAE,MAAM,sBAAsB,CAAC;AACnD,OAAO,EAAE,OAAO,EAAE,MAAM,kBAAkB,CAAC;AAK3C,MAAM,WAAW,eAAe;IAC9B,2DAA2D;IAC3D,gBAAgB,EAAE,MAAM,CAAC;IACzB,8DAA8D;IAC9D,eAAe,EAAE,MAAM,CAAC;IACxB,8DAA8D;IAC9D,kBAAkB,EAAE,MAAM,CAAC;IAC3B,qDAAqD;IACrD,OAAO,EAAE,MAAM,CAAC;IAChB,qDAAqD;IACrD,gBAAgB,EAAE,MAAM,CAAC;IACzB,wEAAwE;IACxE,oBAAoB,EAAE,MAAM,CAAC;IAC7B,kFAAkF;IAClF,kBAAkB,EAAE,MAAM,CAAC;CAC5B;AAED,eAAO,MAAM,wBAAwB,EAAE,eAQtC,CAAC;AAMF;;;;;;;;;GASG;AACH,wBAAsB,eAAe,CACnC,WAAW,EAAE,MAAM,EACnB,cAAc,EAAE,MAAM,EACtB,YAAY,EAAE,MAAM,GAAG,IAAI,EAC3B,OAAO,EAAE,OAAO,EAChB,iBAAiB,EAAE,MAAM,GAAG,IAAI,EAChC,MAAM,GAAE,eAA0C,GACjD,OAAO,CAAC;IACT,QAAQ,EAAE,WAAW,EAAE,CAAC;IACxB,SAAS,EAAE,MAAM,CAAC;IAClB,UAAU,EAAE,MAAM,CAAC;IACnB,cAAc,EAAE,MAAM,CAAC;CACxB,CAAC,CAuLD"}

package/dist/core/rag/assembler.js

@@ -0,0 +1,178 @@
+"use strict";
+/**
+ * Context Assembler — Builds the optimal context window for LLM requests
+ *
+ * Combines:
+ * 1. System prompt
+ * 2. Retrieved relevant context (RAG)
+ * 3. Last N verbatim turns
+ * 4. Compressed checkpoint (if exists)
+ *
+ * Applies token budget to ensure we fit within the model's context window.
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.DEFAULT_ASSEMBLER_CONFIG = void 0;
+exports.assembleContext = assembleContext;
+const retriever_1 = require("./retriever");
+const compressor_1 = require("../optimizer/compressor");
+const dedup_1 = require("../optimizer/dedup");
+exports.DEFAULT_ASSEMBLER_CONFIG = {
+    maxContextTokens: 8000,
+    responseReserve: 2000,
+    recentMessageCount: 6,
+    ragTopK: 15,
+    compressionLevel: 0.3,
+    targetSavingsPercent: 50,
+    qualityFloorTokens: 320,
+};
+function estimateTokens(text) {
+    return Math.ceil(text.split(/\s+/).length * 1.3);
+}
+/**
+ * Assemble the context window for an LLM request.
+ *
+ * @param userMessage - The current user message.
+ * @param conversationId - The conversation ID.
+ * @param systemPrompt - The system prompt.
+ * @param queries - Database queries instance.
+ * @param checkpointSummary - Latest checkpoint summary, if any.
+ * @param config - Assembler configuration.
+ */
+async function assembleContext(userMessage, conversationId, systemPrompt, queries, checkpointSummary, config = exports.DEFAULT_ASSEMBLER_CONFIG) {
+    const tokenBudget = config.maxContextTokens - config.responseReserve;
+    let tokensUsed = 0;
+    let tokensRaw = 0;
+    const assembled = [];
+    // 1. System prompt (always included)
+    if (systemPrompt) {
+        const sysTokens = estimateTokens(systemPrompt);
+        assembled.push({ role: 'system', content: systemPrompt });
+        tokensUsed += sysTokens;
+        tokensRaw += sysTokens;
+    }
+    // 2. Checkpoint summary (if exists)
+    if (checkpointSummary) {
+        const cpTokens = estimateTokens(checkpointSummary);
+        assembled.push({
+            role: 'system',
+            content: `[Conversation Summary]\n${checkpointSummary}\n[End Summary]`,
+        });
+        tokensUsed += cpTokens;
+        tokensRaw += cpTokens;
+    }
+    // 3. Retrieve relevant past context via RAG
+    let ragResults = [];
+    try {
+        ragResults = await (0, retriever_1.retrieveSimilar)(userMessage, conversationId, queries, config.ragTopK);
+    }
+    catch {
+        // RAG may not be available
+    }
+    // 4. Get all messages and calculate raw token count
+    const allMessages = queries.getMessages(conversationId);
+    for (const msg of allMessages) {
+        tokensRaw += estimateTokens(msg.content);
+    }
+    const historyAlreadyHasCurrentUser = allMessages.length > 0 &&
+        allMessages[allMessages.length - 1].role === 'user' &&
+        allMessages[allMessages.length - 1].content.trim() === userMessage.trim();
+    if (!historyAlreadyHasCurrentUser) {
+        tokensRaw += estimateTokens(userMessage);
+    }
+    const targetSentBySavings = Math.floor(tokensRaw * (1 - Math.min(95, Math.max(0, config.targetSavingsPercent)) / 100));
+    const effectiveBudget = Math.min(tokenBudget, Math.max(config.qualityFloorTokens, targetSentBySavings));
+    // 5. Deduplicate messages (Stage 2 of optimizer pipeline)
+    const dedupResult = (0, dedup_1.deduplicateByText)(allMessages.map((m) => ({
+        id: m.id,
+        role: m.role,
+        content: m.content,
+        created_at: m.created_at,
+    })), 0.7);
+    const dedupedMessages = dedupResult.kept;
+    const latestHistoryMessage = dedupedMessages[dedupedMessages.length - 1];
+    const hasCurrentUserInHistory = latestHistoryMessage?.role === 'user' &&
+        latestHistoryMessage?.content.trim() === userMessage.trim();
+    // Split into recent (verbatim) and older (compressible)
+    const verbatimCount = Math.min(2, config.recentMessageCount);
+    const recentMessages = dedupedMessages.slice(-verbatimCount);
+    const olderMessages = dedupedMessages.slice(0, -verbatimCount || undefined);
+    const recentIds = new Set(recentMessages.map((m) => m.id));
+    // 6. Compress older messages (Stage 1 of optimizer pipeline)
+    const compressedOlder = (0, compressor_1.compressMessages)(olderMessages.map((m) => ({ role: m.role, content: m.content })), 0, // don't preserve any recent within this set — they're already separated
+    config.compressionLevel);
+    // 7. Add RAG results (that aren't already in recent or older messages)
+    const ragTokenBudget = Math.floor((effectiveBudget - tokensUsed) * 0.2);
+    let ragTokensUsed = 0;
+    const allKeptIds = new Set([...recentIds, ...olderMessages.map((m) => m.id)]);
+    if (ragResults.length > 0) {
+        const relevantRag = ragResults.filter((r) => !allKeptIds.has(r.id));
+        if (relevantRag.length > 0) {
+            const ragContextParts = [];
+            for (const rag of relevantRag) {
+                const compressed = (0, compressor_1.compressPrompt)(rag.content, config.compressionLevel);
+                const tokens = estimateTokens(compressed.compressed);
+                if (ragTokensUsed + tokens <= ragTokenBudget) {
+                    ragContextParts.push(`[${rag.role}]: ${compressed.compressed}`);
+                    ragTokensUsed += tokens;
+                }
+            }
+            if (ragContextParts.length > 0) {
+                assembled.push({
+                    role: 'system',
+                    content: `[Relevant Earlier Context]\n${ragContextParts.join('\n')}\n[End Context]`,
+                });
+                tokensUsed += ragTokensUsed;
+            }
+        }
+    }
+    // 8. Add compressed older messages
+    const olderBudget = Math.floor((effectiveBudget - tokensUsed) * 0.45);
+    let olderTokensUsed = 0;
+    for (const msg of compressedOlder) {
+        const tokens = estimateTokens(msg.content);
+        if (olderTokensUsed + tokens <= olderBudget) {
+            assembled.push({
+                role: msg.role,
+                content: msg.content,
+            });
+            olderTokensUsed += tokens;
+        }
+    }
+    tokensUsed += olderTokensUsed;
+    // 9. Add recent messages (adaptive compression in aggressive mode)
+    const remainingBudget = effectiveBudget - tokensUsed;
+    let recentTokensUsed = 0;
+    const aggressiveMode = config.targetSavingsPercent >= 70;
+    for (let index = 0; index < recentMessages.length; index++) {
+        const msg = recentMessages[index];
+        const isNewestHistoryMessage = index === recentMessages.length - 1;
+        // Preserve newest user message from history verbatim for intent fidelity.
+        const shouldPreserveVerbatim = msg.role === 'system' || (isNewestHistoryMessage && msg.role === 'user');
+        let candidateContent = msg.content;
+        if (!shouldPreserveVerbatim && aggressiveMode && msg.role === 'assistant') {
+            candidateContent = (0, compressor_1.compressPrompt)(msg.content, Math.min(0.95, config.compressionLevel + 0.1)).compressed;
+        }
+        const tokens = estimateTokens(candidateContent);
+        if (recentTokensUsed + tokens <= remainingBudget) {
+            assembled.push({
+                role: msg.role,
+                content: candidateContent,
+            });
+            recentTokensUsed += tokens;
+        }
+    }
+    tokensUsed += recentTokensUsed;
+    // 10. Add the current user message if not already included from history
+    if (!hasCurrentUserInHistory) {
+        assembled.push({ role: 'user', content: userMessage });
+        tokensUsed += estimateTokens(userMessage);
+        tokensRaw += estimateTokens(userMessage);
+    }
+    return {
+        messages: assembled,
+        tokensRaw,
+        tokensSent: tokensUsed,
+        ragResultCount: ragResults.filter((r) => !allKeptIds.has(r.id)).length,
+    };
+}
+//# sourceMappingURL=assembler.js.map

package/dist/core/rag/assembler.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"assembler.js","sourceRoot":"","sources":["../../../src/core/rag/assembler.ts"],"names":[],"mappings":";AAAA;;;;;;;;;;GAUG;;;AAiDH,0CAmMC;AAhPD,2CAAgE;AAChE,wDAA2E;AAC3E,8CAAuD;AAmB1C,QAAA,wBAAwB,GAAoB;IACvD,gBAAgB,EAAE,IAAI;IACtB,eAAe,EAAE,IAAI;IACrB,kBAAkB,EAAE,CAAC;IACrB,OAAO,EAAE,EAAE;IACX,gBAAgB,EAAE,GAAG;IACrB,oBAAoB,EAAE,EAAE;IACxB,kBAAkB,EAAE,GAAG;CACxB,CAAC;AAEF,SAAS,cAAc,CAAC,IAAY;IAClC,OAAO,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC,MAAM,GAAG,GAAG,CAAC,CAAC;AACnD,CAAC;AAED;;;;;;;;;GASG;AACI,KAAK,UAAU,eAAe,CACnC,WAAmB,EACnB,cAAsB,EACtB,YAA2B,EAC3B,OAAgB,EAChB,iBAAgC,EAChC,SAA0B,gCAAwB;IAOlD,MAAM,WAAW,GAAG,MAAM,CAAC,gBAAgB,GAAG,MAAM,CAAC,eAAe,CAAC;IACrE,IAAI,UAAU,GAAG,CAAC,CAAC;IACnB,IAAI,SAAS,GAAG,CAAC,CAAC;IAClB,MAAM,SAAS,GAAkB,EAAE,CAAC;IAEpC,qCAAqC;IACrC,IAAI,YAAY,EAAE,CAAC;QACjB,MAAM,SAAS,GAAG,cAAc,CAAC,YAAY,CAAC,CAAC;QAC/C,SAAS,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,YAAY,EAAE,CAAC,CAAC;QAC1D,UAAU,IAAI,SAAS,CAAC;QACxB,SAAS,IAAI,SAAS,CAAC;IACzB,CAAC;IAED,oCAAoC;IACpC,IAAI,iBAAiB,EAAE,CAAC;QACtB,MAAM,QAAQ,GAAG,cAAc,CAAC,iBAAiB,CAAC,CAAC;QACnD,SAAS,CAAC,IAAI,CAAC;YACb,IAAI,EAAE,QAAQ;YACd,OAAO,EAAE,2BAA2B,iBAAiB,iBAAiB;SACvE,CAAC,CAAC;QACH,UAAU,IAAI,QAAQ,CAAC;QACvB,SAAS,IAAI,QAAQ,CAAC;IACxB,CAAC;IAED,4CAA4C;IAC5C,IAAI,UAAU,GAAuB,EAAE,CAAC;IACxC,IAAI,CAAC;QACH,UAAU,GAAG,MAAM,IAAA,2BAAe,EAChC,WAAW,EACX,cAAc,EACd,OAAO,EACP,MAAM,CAAC,OAAO,CACf,CAAC;IACJ,CAAC;IAAC,MAAM,CAAC;QACP,2BAA2B;IAC7B,CAAC;IAED,oDAAoD;IACpD,MAAM,WAAW,GAAG,OAAO,CAAC,WAAW,CAAC,cAAc,CAAC,CAAC;IACxD,KAAK,MAAM,GAAG,IAAI,WAAW,EAAE,CAAC;QAC9B,SAAS,IAAI,cAAc,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC;IAC3C,CAAC;IAED,MAAM,4BAA4B,GAChC,WAAW,CAAC,MAAM,GAAG,CAAC;QACtB,WAAW,CAAC,WAAW,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,IAAI,KAAK,MAAM;QACnD,WAAW,CAAC,WAAW,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,OAAO,CAAC,IAAI,EAAE,KAAK,WAAW,CAAC,IAAI,EAAE,CAAC;IAE5E,IAAI,CAAC,4BAA4B,EAAE,CAAC;QAClC,SAAS,IAAI,cAAc,CAAC,WAAW,CAAC,CAAC;IAC3C,CAAC;IAED,MAAM,mBAAmB,GAAG,IAAI,CAAC,KAAK,CACpC,SAAS,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,EAAE,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,MAAM,CAAC,oBAAoB,CAAC,CAAC,GAAG,GAAG,CAAC,CAC/E,CAAC;IAEF,MAAM,eAAe,GAAG,IAAI,CAAC,GAAG,CAC9B,WAAW,EACX,IAAI,CAAC,GAAG,CAAC,MAAM,CAAC,kBAAkB,EAAE,mBAAmB,CAAC,CACzD,CAAC;IAEF,0DAA0D;IAC1D,MAAM,WAAW,GAAG,IAAA,yBAAiB,EACnC,WAAW,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC;QACtB,EAAE,EAAE,CAAC,CAAC,EAAE;QACR,IAAI,EAAE,CAAC,CAAC,IAAI;QACZ,OAAO,EAAE,CAAC,CAAC,OAAO;QAClB,UAAU,EAAE,CAAC,CAAC,UAAU;KACzB,CAAC,CAAC,EACH,GAAG,CACJ,CAAC;IACF,MAAM,eAAe,GAAG,WAAW,CAAC,IAAI,CAAC;IACzC,MAAM,oBAAoB,GAAG,eAAe,CAAC,eAAe,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;IACzE,MAAM,uBAAuB,GAC3B,oBAAoB,EAAE,IAAI,KAAK,MAAM;QACrC,oBAAoB,EAAE,OAAO,CAAC,IAAI,EAAE,KAAK,WAAW,CAAC,IAAI,EAAE,CAAC;IAE9D,wDAAwD;IACxD,MAAM,aAAa,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,MAAM,CAAC,kBAAkB,CAAC,CAAC;IAC7D,MAAM,cAAc,GAAG,eAAe,CAAC,KAAK,CAAC,CAAC,aAAa,CAAC,CAAC;IAC7D,MAAM,aAAa,GAAG,eAAe,CAAC,KAAK,CAAC,CAAC,EAAE,CAAC,aAAa,IAAI,SAAS,CAAC,CAAC;IAC5E,MAAM,SAAS,GAAG,IAAI,GAAG,CAAC,cAAc,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;IAE3D,6DAA6D;IAC7D,MAAM,eAAe,GAAG,IAAA,6BAAgB,EACtC,aAAa,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,OAAO,EAAE,CAAC,CAAC,OAAO,EAAE,CAAC,CAAC,EAChE,CAAC,EAAE,wEAAwE;IAC3E,MAAM,CAAC,gBAAgB,CACxB,CAAC;IAEF,uEAAuE;IACvE,MAAM,cAAc,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,eAAe,GAAG,UAAU,CAAC,GAAG,GAAG,CAAC,CAAC;IACxE,IAAI,aAAa,GAAG,CAAC,CAAC;IACtB,MAAM,UAAU,GAAG,IAAI,GAAG,CAAC,CAAC,GAAG,SAAS,EAAE,GAAG,aAAa,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC;IAE9E,IAAI,UAAU,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;QAC1B,MAAM,WAAW,GAAG,UAAU,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,UAAU,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;QAEpE,IAAI,WAAW,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YAC3B,MAAM,eAAe,GAAa,EAAE,CAAC;YAErC,KAAK,MAAM,GAAG,IAAI,WAAW,EAAE,CAAC;gBAC9B,MAAM,UAAU,GAAG,IAAA,2BAAc,EAAC,GAAG,CAAC,OAAO,EAAE,MAAM,CAAC,gBAAgB,CAAC,CAAC;gBACxE,MAAM,MAAM,GAAG,cAAc,CAAC,UAAU,CAAC,UAAU,CAAC,CAAC;gBAErD,IAAI,aAAa,GAAG,MAAM,IAAI,cAAc,EAAE,CAAC;oBAC7C,eAAe,CAAC,IAAI,CAAC,IAAI,GAAG,CAAC,IAAI,MAAM,UAAU,CAAC,UAAU,EAAE,CAAC,CAAC;oBAChE,aAAa,IAAI,MAAM,CAAC;gBAC1B,CAAC;YACH,CAAC;YAED,IAAI,eAAe,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;gBAC/B,SAAS,CAAC,IAAI,CAAC;oBACb,IAAI,EAAE,QAAQ;oBACd,OAAO,EAAE,+BAA+B,eAAe,CAAC,IAAI,CAAC,IAAI,CAAC,iBAAiB;iBACpF,CAAC,CAAC;gBACH,UAAU,IAAI,aAAa,CAAC;YAC9B,CAAC;QACH,CAAC;IACH,CAAC;IAED,mCAAmC;IACnC,MAAM,WAAW,GAAG,IAAI,CAAC,KAAK,CAAC,CAAC,eAAe,GAAG,UAAU,CAAC,GAAG,IAAI,CAAC,CAAC;IACtE,IAAI,eAAe,GAAG,CAAC,CAAC;IAExB,KAAK,MAAM,GAAG,IAAI,eAAe,EAAE,CAAC;QAClC,MAAM,MAAM,GAAG,cAAc,CAAC,GAAG,CAAC,OAAO,CAAC,CAAC;QAC3C,IAAI,eAAe,GAAG,MAAM,IAAI,WAAW,EAAE,CAAC;YAC5C,SAAS,CAAC,IAAI,CAAC;gBACb,IAAI,EAAE,GAAG,CAAC,IAAuC;gBACjD,OAAO,EAAE,GAAG,CAAC,OAAO;aACrB,CAAC,CAAC;YACH,eAAe,IAAI,MAAM,CAAC;QAC5B,CAAC;IACH,CAAC;IACD,UAAU,IAAI,eAAe,CAAC;IAE9B,mEAAmE;IACnE,MAAM,eAAe,GAAG,eAAe,GAAG,UAAU,CAAC;IACrD,IAAI,gBAAgB,GAAG,CAAC,CAAC;IACzB,MAAM,cAAc,GAAG,MAAM,CAAC,oBAAoB,IAAI,EAAE,CAAC;IAEzD,KAAK,IAAI,KAAK,GAAG,CAAC,EAAE,KAAK,GAAG,cAAc,CAAC,MAAM,EAAE,KAAK,EAAE,EAAE,CAAC;QAC3D,MAAM,GAAG,GAAG,cAAc,CAAC,KAAK,CAAC,CAAC;QAClC,MAAM,sBAAsB,GAAG,KAAK,KAAK,cAAc,CAAC,MAAM,GAAG,CAAC,CAAC;QAEnE,0EAA0E;QAC1E,MAAM,sBAAsB,GAC1B,GAAG,CAAC,IAAI,KAAK,QAAQ,IAAI,CAAC,sBAAsB,IAAI,GAAG,CAAC,IAAI,KAAK,MAAM,CAAC,CAAC;QAE3E,IAAI,gBAAgB,GAAG,GAAG,CAAC,OAAO,CAAC;QACnC,IAAI,CAAC,sBAAsB,IAAI,cAAc,IAAI,GAAG,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YAC1E,gBAAgB,GAAG,IAAA,2BAAc,EAC/B,GAAG,CAAC,OAAO,EACX,IAAI,CAAC,GAAG,CAAC,IAAI,EAAE,MAAM,CAAC,gBAAgB,GAAG,GAAG,CAAC,CAC9C,CAAC,UAAU,CAAC;QACf,CAAC;QAED,MAAM,MAAM,GAAG,cAAc,CAAC,gBAAgB,CAAC,CAAC;QAChD,IAAI,gBAAgB,GAAG,MAAM,IAAI,eAAe,EAAE,CAAC;YACjD,SAAS,CAAC,IAAI,CAAC;gBACb,IAAI,EAAE,GAAG,CAAC,IAAuC;gBACjD,OAAO,EAAE,gBAAgB;aAC1B,CAAC,CAAC;YACH,gBAAgB,IAAI,MAAM,CAAC;QAC7B,CAAC;IACH,CAAC;IACD,UAAU,IAAI,gBAAgB,CAAC;IAE/B,wEAAwE;IACxE,IAAI,CAAC,uBAAuB,EAAE,CAAC;QAC7B,SAAS,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,WAAW,EAAE,CAAC,CAAC;QACvD,UAAU,IAAI,cAAc,CAAC,WAAW,CAAC,CAAC;QAC1C,SAAS,IAAI,cAAc,CAAC,WAAW,CAAC,CAAC;IAC3C,CAAC;IAED,OAAO;QACL,QAAQ,EAAE,SAAS;QACnB,SAAS;QACT,UAAU,EAAE,UAAU;QACtB,cAAc,EAAE,UAAU,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,UAAU,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,MAAM;KACvE,CAAC;AACJ,CAAC"}

package/dist/core/rag/embedder.d.ts

@@ -0,0 +1,16 @@
+/**
+ * Embedder — Local embedding using all-MiniLM-L6-v2 via onnxruntime-node
+ *
+ * Zero API cost, ~23ms per embedding, fully offline.
+ * Output dimension: 384
+ */
+/**
+ * Generate embedding for a text using the ONNX model.
+ * Returns null if the model is not available.
+ */
+export declare function embed(text: string): Promise<Float32Array | null>;
+/**
+ * Batch embed multiple texts.
+ */
+export declare function embedBatch(texts: string[]): Promise<(Float32Array | null)[]>;
+//# sourceMappingURL=embedder.d.ts.map

package/dist/core/rag/embedder.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"embedder.d.ts","sourceRoot":"","sources":["../../../src/core/rag/embedder.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAmGH;;;GAGG;AACH,wBAAsB,KAAK,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,YAAY,GAAG,IAAI,CAAC,CAuDtE;AA2CD;;GAEG;AACH,wBAAsB,UAAU,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,OAAO,CAAC,CAAC,YAAY,GAAG,IAAI,CAAC,EAAE,CAAC,CAMlF"}

package/dist/core/rag/embedder.js

@@ -0,0 +1,223 @@
+"use strict";
+/**
+ * Embedder — Local embedding using all-MiniLM-L6-v2 via onnxruntime-node
+ *
+ * Zero API cost, ~23ms per embedding, fully offline.
+ * Output dimension: 384
+ */
+var __createBinding = (this && this.__createBinding) || (Object.create ? (function(o, m, k, k2) {
+    if (k2 === undefined) k2 = k;
+    var desc = Object.getOwnPropertyDescriptor(m, k);
+    if (!desc || ("get" in desc ? !m.__esModule : desc.writable || desc.configurable)) {
+      desc = { enumerable: true, get: function() { return m[k]; } };
+    }
+    Object.defineProperty(o, k2, desc);
+}) : (function(o, m, k, k2) {
+    if (k2 === undefined) k2 = k;
+    o[k2] = m[k];
+}));
+var __setModuleDefault = (this && this.__setModuleDefault) || (Object.create ? (function(o, v) {
+    Object.defineProperty(o, "default", { enumerable: true, value: v });
+}) : function(o, v) {
+    o["default"] = v;
+});
+var __importStar = (this && this.__importStar) || (function () {
+    var ownKeys = function(o) {
+        ownKeys = Object.getOwnPropertyNames || function (o) {
+            var ar = [];
+            for (var k in o) if (Object.prototype.hasOwnProperty.call(o, k)) ar[ar.length] = k;
+            return ar;
+        };
+        return ownKeys(o);
+    };
+    return function (mod) {
+        if (mod && mod.__esModule) return mod;
+        var result = {};
+        if (mod != null) for (var k = ownKeys(mod), i = 0; i < k.length; i++) if (k[i] !== "default") __createBinding(result, mod, k[i]);
+        __setModuleDefault(result, mod);
+        return result;
+    };
+})();
+var __importDefault = (this && this.__importDefault) || function (mod) {
+    return (mod && mod.__esModule) ? mod : { "default": mod };
+};
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.embed = embed;
+exports.embedBatch = embedBatch;
+const path_1 = __importDefault(require("path"));
+const fs_1 = __importDefault(require("fs"));
+const os_1 = __importDefault(require("os"));
+let session = null;
+let sessionChecked = false;
+let tokenizer = null;
+// Simple word-piece tokenizer for MiniLM
+// In production, you'd use the actual tokenizer from the model
+// For now, we use a simplified approach
+const MAX_SEQ_LENGTH = 256;
+/**
+ * Initialize the ONNX runtime session with MiniLM-L6-v2.
+ * Caches the result (including "not available") to avoid repeated checks.
+ */
+async function getSession() {
+    if (session)
+        return session;
+    if (sessionChecked)
+        return null; // Already checked, model not available
+    sessionChecked = true;
+    try {
+        const ort = await Promise.resolve().then(() => __importStar(require('onnxruntime-node')));
+        // Check for model in ~/.opengauge/models/
+        const modelDir = path_1.default.join(os_1.default.homedir(), '.opengauge', 'models');
+        const modelPath = path_1.default.join(modelDir, 'all-MiniLM-L6-v2.onnx');
+        if (!fs_1.default.existsSync(modelPath)) {
+            console.log('[OpenGauge] Embedding model not found. Using fallback hash embeddings.');
+            console.log(`  To enable full semantic search, place all-MiniLM-L6-v2.onnx in ${modelDir}`);
+            return null;
+        }
+        session = await ort.InferenceSession.create(modelPath, {
+            executionProviders: ['cpu'],
+        });
+        console.log('[OpenGauge] ONNX embedding model loaded successfully.');
+        return session;
+    }
+    catch (error) {
+        console.warn('[OpenGauge] Failed to initialize ONNX embedding model. Using fallback.', error);
+        return null;
+    }
+}
+/**
+ * Simple tokenization: split on whitespace and punctuation,
+ * convert to indices. This is a fallback; a proper tokenizer
+ * would use the model's vocabulary.
+ */
+function simpleTokenize(text) {
+    // Simple whitespace tokenization + padding/truncation
+    const words = text
+        .toLowerCase()
+        .replace(/[^\w\s]/g, ' ')
+        .split(/\s+/)
+        .filter((w) => w.length > 0)
+        .slice(0, MAX_SEQ_LENGTH - 2);
+    const seqLength = words.length + 2; // [CLS] + words + [SEP]
+    const inputIds = new BigInt64Array(seqLength);
+    const attentionMask = new BigInt64Array(seqLength);
+    const tokenTypeIds = new BigInt64Array(seqLength);
+    // [CLS] token = 101
+    inputIds[0] = BigInt(101);
+    attentionMask[0] = BigInt(1);
+    // Map words to pseudo token IDs using hash
+    for (let i = 0; i < words.length; i++) {
+        inputIds[i + 1] = BigInt(hashWord(words[i]));
+        attentionMask[i + 1] = BigInt(1);
+    }
+    // [SEP] token = 102
+    inputIds[seqLength - 1] = BigInt(102);
+    attentionMask[seqLength - 1] = BigInt(1);
+    return { inputIds, attentionMask, tokenTypeIds };
+}
+function hashWord(word) {
+    let hash = 0;
+    for (let i = 0; i < word.length; i++) {
+        hash = ((hash << 5) - hash + word.charCodeAt(i)) & 0x7fff;
+    }
+    return (hash % 30000) + 1000; // Keep in vocab range
+}
+/**
+ * Generate embedding for a text using the ONNX model.
+ * Returns null if the model is not available.
+ */
+async function embed(text) {
+    const sess = await getSession();
+    if (!sess) {
+        // Fallback: generate a simple TF-IDF-like hash embedding
+        return fallbackEmbed(text);
+    }
+    try {
+        const ort = await Promise.resolve().then(() => __importStar(require('onnxruntime-node')));
+        const { inputIds, attentionMask, tokenTypeIds } = simpleTokenize(text);
+        const feeds = {
+            input_ids: new ort.Tensor('int64', inputIds, [1, inputIds.length]),
+            attention_mask: new ort.Tensor('int64', attentionMask, [1, attentionMask.length]),
+            token_type_ids: new ort.Tensor('int64', tokenTypeIds, [1, tokenTypeIds.length]),
+        };
+        const results = await sess.run(feeds);
+        // Mean pooling over token embeddings
+        const output = results['last_hidden_state'] || results['token_embeddings'] || Object.values(results)[0];
+        const data = output.data;
+        const hiddenSize = 384;
+        const numTokens = inputIds.length;
+        const embedding = new Float32Array(hiddenSize);
+        for (let i = 0; i < numTokens; i++) {
+            for (let j = 0; j < hiddenSize; j++) {
+                embedding[j] += data[i * hiddenSize + j];
+            }
+        }
+        // Average
+        for (let j = 0; j < hiddenSize; j++) {
+            embedding[j] /= numTokens;
+        }
+        // L2 normalize
+        let norm = 0;
+        for (let j = 0; j < hiddenSize; j++) {
+            norm += embedding[j] * embedding[j];
+        }
+        norm = Math.sqrt(norm);
+        if (norm > 0) {
+            for (let j = 0; j < hiddenSize; j++) {
+                embedding[j] /= norm;
+            }
+        }
+        return embedding;
+    }
+    catch (error) {
+        console.warn('Embedding failed, using fallback:', error);
+        return fallbackEmbed(text);
+    }
+}
+/**
+ * Fallback embedding using character n-gram hashing.
+ * Produces a 384-dim vector. Not as good as MiniLM but enables
+ * basic similarity search without the ONNX model.
+ */
+function fallbackEmbed(text) {
+    const DIM = 384;
+    const embedding = new Float32Array(DIM);
+    const normalized = text.toLowerCase().replace(/[^\w\s]/g, '');
+    const words = normalized.split(/\s+/);
+    // Hash words into the embedding dimensions
+    for (const word of words) {
+        for (let n = 1; n <= 3; n++) {
+            for (let i = 0; i <= word.length - n; i++) {
+                const ngram = word.substring(i, i + n);
+                let hash = 0;
+                for (let c = 0; c < ngram.length; c++) {
+                    hash = (hash * 31 + ngram.charCodeAt(c)) % DIM;
+                }
+                embedding[hash] += 1;
+            }
+        }
+    }
+    // L2 normalize
+    let norm = 0;
+    for (let i = 0; i < DIM; i++) {
+        norm += embedding[i] * embedding[i];
+    }
+    norm = Math.sqrt(norm);
+    if (norm > 0) {
+        for (let i = 0; i < DIM; i++) {
+            embedding[i] /= norm;
+        }
+    }
+    return embedding;
+}
+/**
+ * Batch embed multiple texts.
+ */
+async function embedBatch(texts) {
+    const results = [];
+    for (const text of texts) {
+        results.push(await embed(text));
+    }
+    return results;
+}
+//# sourceMappingURL=embedder.js.map