agentic-flow 2.0.7 → 3.0.0-alpha.2

package/dist/agentic-flow/src/router/providers/onnx-local.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"onnx-local.d.ts","sourceRoot":"","sources":["../../../../../src/router/providers/onnx-local.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAOH,OAAO,KAAK,EACV,WAAW,EACX,UAAU,EACV,YAAY,EACZ,WAAW,EAIZ,MAAM,aAAa,CAAC;AAErB,MAAM,WAAW,eAAe;IAC9B,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,kBAAkB,CAAC,EAAE,MAAM,EAAE,CAAC;IAC9B,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,WAAW,CAAC,EAAE,MAAM,CAAC;CACtB;AAED,qBAAa,iBAAkB,YAAW,WAAW;IACnD,IAAI,SAAgB;IACpB,IAAI,EAAG,QAAQ,CAAU;IACzB,iBAAiB,UAAS;IAC1B,aAAa,UAAS;IACtB,WAAW,UAAS;IAEpB,OAAO,CAAC,OAAO,CAAqC;IACpD,OAAO,CAAC,MAAM,CAA4B;IAC1C,OAAO,CAAC,SAAS,CAAa;IAC9B,OAAO,CAAC,QAAQ,CAAa;gBAEjB,MAAM,GAAE,eAAoB;IASxC;;OAEG;YACW,aAAa;IAc3B;;OAEG;IACH,OAAO,CAAC,MAAM;IAId;;OAEG;IACH,OAAO,CAAC,MAAM;IAgBd;;OAEG;YACW,iBAAiB;IA6C/B;;OAEG;IACH,OAAO,CAAC,cAAc;IAqBtB;;;OAGG;IACH,OAAO,CAAC,iBAAiB;IA2BzB;;OAEG;IACG,IAAI,CAAC,MAAM,EAAE,UAAU,GAAG,OAAO,CAAC,YAAY,CAAC;IA2IrD;;OAEG;IACI,MAAM,CAAC,MAAM,EAAE,UAAU,GAAG,cAAc,CAAC,WAAW,CAAC;IAI9D;;OAEG;IACH,oBAAoB,CAAC,QAAQ,EAAE,MAAM,EAAE,GAAG,OAAO;IAKjD;;OAEG;IACH,YAAY;;;;;;IASZ;;OAEG;IACG,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC;CAU/B"}

package/dist/agentic-flow/src/router/providers/onnx-local.js

@@ -0,0 +1,295 @@
+/**
+ * ONNX Runtime Local Inference Provider for Phi-4
+ *
+ * Uses onnxruntime-node for true local CPU/GPU inference
+ */
+import * as ort from 'onnxruntime-node';
+import { get_encoding } from 'tiktoken';
+import { ensurePhi4Model, ModelDownloader } from '../../utils/model-downloader.js';
+export class ONNXLocalProvider {
+    name = 'onnx-local';
+    type = 'custom';
+    supportsStreaming = false; // Streaming requires complex token generation loop
+    supportsTools = false;
+    supportsMCP = false;
+    session = null;
+    config;
+    tokenizer = null;
+    tiktoken = null;
+    constructor(config = {}) {
+        this.config = {
+            modelPath: config.modelPath || './models/phi-4/cpu_and_mobile/cpu-int4-rtn-block-32-acc-level-4/model.onnx',
+            executionProviders: config.executionProviders || ['cpu'],
+            maxTokens: config.maxTokens || 100,
+            temperature: config.temperature || 0.7
+        };
+    }
+    /**
+     * Load optimized tiktoken tokenizer (cl100k_base for Phi-4)
+     */
+    async loadTokenizer() {
+        if (this.tiktoken)
+            return;
+        try {
+            // Use cl100k_base encoding (GPT-4, similar to Phi-4)
+            this.tiktoken = get_encoding('cl100k_base');
+            console.log('✅ Tokenizer loaded (tiktoken cl100k_base)');
+        }
+        catch (error) {
+            console.error('❌ Failed to load tiktoken:', error);
+            throw new Error(`Tokenizer loading failed: ${error}`);
+        }
+    }
+    /**
+     * Encode text using tiktoken (fast BPE)
+     */
+    encode(text) {
+        return Array.from(this.tiktoken.encode(text));
+    }
+    /**
+     * Decode tokens using tiktoken
+     */
+    decode(ids) {
+        try {
+            const decoded = this.tiktoken.decode(new Uint32Array(ids));
+            // tiktoken returns buffer, convert to string
+            if (typeof decoded === 'string') {
+                return decoded;
+            }
+            else if (decoded instanceof Uint8Array || decoded instanceof Buffer) {
+                return new TextDecoder().decode(decoded);
+            }
+            return String(decoded);
+        }
+        catch (error) {
+            console.warn('Decode error, returning raw IDs:', error);
+            return ids.join(',');
+        }
+    }
+    /**
+     * Initialize ONNX session (with automatic model download)
+     */
+    async initializeSession() {
+        if (this.session)
+            return;
+        try {
+            // Ensure model is downloaded
+            console.log(`🔍 Checking for Phi-4 ONNX model...`);
+            const modelPath = await ensurePhi4Model((progress) => {
+                if (progress.percentage % 10 < 1) { // Log every ~10%
+                    console.log(`   📥 Downloading: ${ModelDownloader.formatProgress(progress)}`);
+                }
+            });
+            // Update config with actual model path
+            this.config.modelPath = modelPath;
+            console.log(`📦 Loading ONNX model: ${this.config.modelPath}`);
+            this.session = await ort.InferenceSession.create(this.config.modelPath, {
+                executionProviders: this.config.executionProviders,
+                graphOptimizationLevel: 'all',
+                enableCpuMemArena: true,
+                enableMemPattern: true
+            });
+            console.log(`✅ ONNX model loaded`);
+            console.log(`🔧 Execution providers: ${this.config.executionProviders.join(', ')}`);
+            // Load tokenizer
+            await this.loadTokenizer();
+        }
+        catch (error) {
+            const providerError = {
+                name: 'ONNXInitError',
+                message: `Failed to initialize ONNX model: ${error}`,
+                provider: 'onnx-local',
+                retryable: false
+            };
+            throw providerError;
+        }
+    }
+    /**
+     * Format messages for Phi-4 chat template
+     */
+    formatMessages(messages) {
+        let prompt = '';
+        for (const msg of messages) {
+            const content = typeof msg.content === 'string'
+                ? msg.content
+                : msg.content.map(c => c.type === 'text' ? c.text : '').join('');
+            if (msg.role === 'system') {
+                prompt += `<|system|>\n${content}<|end|>\n`;
+            }
+            else if (msg.role === 'user') {
+                prompt += `<|user|>\n${content}<|end|>\n`;
+            }
+            else if (msg.role === 'assistant') {
+                prompt += `<|assistant|>\n${content}<|end|>\n`;
+            }
+        }
+        prompt += '<|assistant|>\n';
+        return prompt;
+    }
+    /**
+     * Initialize KV cache tensors for all 32 layers
+     * Phi-4 architecture: 32 layers, 8 KV heads, 128 head_dim
+     */
+    initializeKVCache(batchSize, sequenceLength) {
+        const numLayers = 32;
+        const numKVHeads = 8;
+        const headDim = 128; // 3072 / 24 = 128
+        const kvCache = {};
+        // Initialize empty cache for each layer (key and value)
+        for (let i = 0; i < numLayers; i++) {
+            // Empty cache: [batch_size, num_kv_heads, 0, head_dim]
+            const emptyCache = new Float32Array(0);
+            kvCache[`past_key_values.${i}.key`] = new ort.Tensor('float32', emptyCache, [batchSize, numKVHeads, 0, headDim]);
+            kvCache[`past_key_values.${i}.value`] = new ort.Tensor('float32', emptyCache, [batchSize, numKVHeads, 0, headDim]);
+        }
+        return kvCache;
+    }
+    /**
+     * Chat completion using ONNX with KV cache
+     */
+    async chat(params) {
+        await this.initializeSession();
+        const startTime = Date.now();
+        const prompt = this.formatMessages(params.messages);
+        try {
+            // Tokenize input using optimized tiktoken
+            const inputIds = this.encode(prompt);
+            console.log(`📝 Input tokens: ${inputIds.length}`);
+            // Initialize KV cache (reusable for batch)
+            let pastKVCache = this.initializeKVCache(1, 0);
+            // Track all generated tokens
+            const allTokenIds = [...inputIds];
+            const outputIds = [];
+            // Pre-allocate tensor buffers for performance
+            const maxSeqLen = inputIds.length + (params.maxTokens || this.config.maxTokens);
+            // Autoregressive generation loop
+            const maxNewTokens = params.maxTokens || this.config.maxTokens;
+            for (let step = 0; step < maxNewTokens; step++) {
+                // For first step, use all input tokens; for subsequent steps, use only last token
+                const currentInputIds = step === 0 ? inputIds : [outputIds[outputIds.length - 1]];
+                const currentSeqLen = currentInputIds.length;
+                // Create input tensor for current step
+                const inputTensor = new ort.Tensor('int64', BigInt64Array.from(currentInputIds.map(BigInt)), [1, currentSeqLen]);
+                // Create attention mask for current step
+                const totalSeqLen = allTokenIds.length;
+                const attentionMask = new ort.Tensor('int64', BigInt64Array.from(Array(totalSeqLen).fill(1n)), [1, totalSeqLen]);
+                // Build feeds with input, attention mask, and KV cache
+                const feeds = {
+                    input_ids: inputTensor,
+                    attention_mask: attentionMask,
+                    ...pastKVCache
+                };
+                // Run inference
+                const results = await this.session.run(feeds);
+                // Get logits for next token (last position)
+                const logits = results.logits.data;
+                const vocabSize = results.logits.dims[results.logits.dims.length - 1];
+                // Extract logits for last token
+                const lastTokenLogitsOffset = (currentSeqLen - 1) * vocabSize;
+                // Apply temperature and get next token
+                let nextToken = 0;
+                let maxVal = -Infinity;
+                for (let i = 0; i < vocabSize; i++) {
+                    const logit = logits[lastTokenLogitsOffset + i] / (params.temperature || this.config.temperature);
+                    if (logit > maxVal) {
+                        maxVal = logit;
+                        nextToken = i;
+                    }
+                }
+                // Add to output
+                outputIds.push(nextToken);
+                allTokenIds.push(nextToken);
+                // Check for end token (2 is typical EOS for Phi models)
+                if (nextToken === 2 || nextToken === 0) {
+                    console.log(`🛑 Stop token detected: ${nextToken}`);
+                    break;
+                }
+                // Update KV cache from outputs for next iteration
+                pastKVCache = {};
+                for (let i = 0; i < 32; i++) {
+                    pastKVCache[`past_key_values.${i}.key`] = results[`present.${i}.key`];
+                    pastKVCache[`past_key_values.${i}.value`] = results[`present.${i}.value`];
+                }
+                // Progress indicator
+                if ((step + 1) % 10 === 0) {
+                    console.log(`🔄 Generated ${step + 1} tokens...`);
+                }
+            }
+            // Decode output using optimized tiktoken
+            const generatedText = this.decode(outputIds);
+            const latency = Date.now() - startTime;
+            const tokensPerSecond = (outputIds.length / (latency / 1000)).toFixed(1);
+            console.log(`✅ Generated: ${generatedText}`);
+            console.log(`⏱️  Latency: ${latency}ms (${tokensPerSecond} tokens/sec)`);
+            const content = [{
+                    type: 'text',
+                    text: generatedText.trim()
+                }];
+            return {
+                id: `onnx-local-${Date.now()}`,
+                model: this.config.modelPath,
+                content,
+                stopReason: 'end_turn',
+                usage: {
+                    inputTokens: inputIds.length,
+                    outputTokens: outputIds.length
+                },
+                metadata: {
+                    provider: 'onnx-local',
+                    model: 'Phi-4-mini-instruct-onnx',
+                    latency,
+                    cost: 0, // Local inference is free
+                    executionProviders: this.config.executionProviders,
+                    tokensPerSecond: parseFloat(tokensPerSecond)
+                }
+            };
+        }
+        catch (error) {
+            const providerError = {
+                name: 'ONNXInferenceError',
+                message: `ONNX inference failed: ${error}`,
+                provider: 'onnx-local',
+                retryable: true
+            };
+            throw providerError;
+        }
+    }
+    /**
+     * Streaming not implemented (requires complex generation loop)
+     */
+    async *stream(params) {
+        throw new Error('Streaming not yet implemented for ONNX local inference');
+    }
+    /**
+     * Validate capabilities
+     */
+    validateCapabilities(features) {
+        const supported = ['chat'];
+        return features.every(f => supported.includes(f));
+    }
+    /**
+     * Get model info
+     */
+    getModelInfo() {
+        return {
+            modelPath: this.config.modelPath,
+            executionProviders: this.config.executionProviders,
+            initialized: this.session !== null,
+            tokenizerLoaded: this.tiktoken !== null
+        };
+    }
+    /**
+     * Cleanup resources
+     */
+    async dispose() {
+        if (this.session) {
+            // ONNX Runtime sessions don't have explicit disposal in Node.js
+            this.session = null;
+        }
+        if (this.tiktoken) {
+            this.tiktoken.free();
+            this.tiktoken = null;
+        }
+    }
+}
+//# sourceMappingURL=onnx-local.js.map

package/dist/agentic-flow/src/router/providers/onnx-local.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"onnx-local.js","sourceRoot":"","sources":["../../../../../src/router/providers/onnx-local.ts"],"names":[],"mappings":"AAAA;;;;GAIG;AAEH,OAAO,KAAK,GAAG,MAAM,kBAAkB,CAAC;AAGxC,OAAO,EAAE,YAAY,EAAE,MAAM,UAAU,CAAC;AACxC,OAAO,EAAE,eAAe,EAAE,eAAe,EAAE,MAAM,iCAAiC,CAAC;AAkBnF,MAAM,OAAO,iBAAiB;IAC5B,IAAI,GAAG,YAAY,CAAC;IACpB,IAAI,GAAG,QAAiB,CAAC;IACzB,iBAAiB,GAAG,KAAK,CAAC,CAAC,mDAAmD;IAC9E,aAAa,GAAG,KAAK,CAAC;IACtB,WAAW,GAAG,KAAK,CAAC;IAEZ,OAAO,GAAgC,IAAI,CAAC;IAC5C,MAAM,CAA4B;IAClC,SAAS,GAAQ,IAAI,CAAC;IACtB,QAAQ,GAAQ,IAAI,CAAC;IAE7B,YAAY,SAA0B,EAAE;QACtC,IAAI,CAAC,MAAM,GAAG;YACZ,SAAS,EAAE,MAAM,CAAC,SAAS,IAAI,4EAA4E;YAC3G,kBAAkB,EAAE,MAAM,CAAC,kBAAkB,IAAI,CAAC,KAAK,CAAC;YACxD,SAAS,EAAE,MAAM,CAAC,SAAS,IAAI,GAAG;YAClC,WAAW,EAAE,MAAM,CAAC,WAAW,IAAI,GAAG;SACvC,CAAC;IACJ,CAAC;IAED;;OAEG;IACK,KAAK,CAAC,aAAa;QACzB,IAAI,IAAI,CAAC,QAAQ;YAAE,OAAO;QAE1B,IAAI,CAAC;YACH,qDAAqD;YACrD,IAAI,CAAC,QAAQ,GAAG,YAAY,CAAC,aAAa,CAAC,CAAC;YAE5C,OAAO,CAAC,GAAG,CAAC,2CAA2C,CAAC,CAAC;QAC3D,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,OAAO,CAAC,KAAK,CAAC,4BAA4B,EAAE,KAAK,CAAC,CAAC;YACnD,MAAM,IAAI,KAAK,CAAC,6BAA6B,KAAK,EAAE,CAAC,CAAC;QACxD,CAAC;IACH,CAAC;IAED;;OAEG;IACK,MAAM,CAAC,IAAY;QACzB,OAAO,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC;IAChD,CAAC;IAED;;OAEG;IACK,MAAM,CAAC,GAAa;QAC1B,IAAI,CAAC;YACH,MAAM,OAAO,GAAG,IAAI,CAAC,QAAQ,CAAC,MAAM,CAAC,IAAI,WAAW,CAAC,GAAG,CAAC,CAAC,CAAC;YAC3D,6CAA6C;YAC7C,IAAI,OAAO,OAAO,KAAK,QAAQ,EAAE,CAAC;gBAChC,OAAO,OAAO,CAAC;YACjB,CAAC;iBAAM,IAAI,OAAO,YAAY,UAAU,IAAI,OAAO,YAAY,MAAM,EAAE,CAAC;gBACtE,OAAO,IAAI,WAAW,EAAE,CAAC,MAAM,CAAC,OAAO,CAAC,CAAC;YAC3C,CAAC;YACD,OAAO,MAAM,CAAC,OAAO,CAAC,CAAC;QACzB,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,OAAO,CAAC,IAAI,CAAC,kCAAkC,EAAE,KAAK,CAAC,CAAC;YACxD,OAAO,GAAG,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;QACvB,CAAC;IACH,CAAC;IAED;;OAEG;IACK,KAAK,CAAC,iBAAiB;QAC7B,IAAI,IAAI,CAAC,OAAO;YAAE,OAAO;QAEzB,IAAI,CAAC;YACH,6BAA6B;YAC7B,OAAO,CAAC,GAAG,CAAC,qCAAqC,CAAC,CAAC;YAEnD,MAAM,SAAS,GAAG,MAAM,eAAe,CAAC,CAAC,QAAQ,EAAE,EAAE;gBACnD,IAAI,QAAQ,CAAC,UAAU,GAAG,EAAE,GAAG,CAAC,EAAE,CAAC,CAAC,iBAAiB;oBACnD,OAAO,CAAC,GAAG,CAAC,sBAAsB,eAAe,CAAC,cAAc,CAAC,QAAQ,CAAC,EAAE,CAAC,CAAC;gBAChF,CAAC;YACH,CAAC,CAAC,CAAC;YAEH,uCAAuC;YACvC,IAAI,CAAC,MAAM,CAAC,SAAS,GAAG,SAAS,CAAC;YAElC,OAAO,CAAC,GAAG,CAAC,0BAA0B,IAAI,CAAC,MAAM,CAAC,SAAS,EAAE,CAAC,CAAC;YAE/D,IAAI,CAAC,OAAO,GAAG,MAAM,GAAG,CAAC,gBAAgB,CAAC,MAAM,CAC9C,IAAI,CAAC,MAAM,CAAC,SAAS,EACrB;gBACE,kBAAkB,EAAE,IAAI,CAAC,MAAM,CAAC,kBAAyB;gBACzD,sBAAsB,EAAE,KAAK;gBAC7B,iBAAiB,EAAE,IAAI;gBACvB,gBAAgB,EAAE,IAAI;aACvB,CACF,CAAC;YAEF,OAAO,CAAC,GAAG,CAAC,qBAAqB,CAAC,CAAC;YACnC,OAAO,CAAC,GAAG,CAAC,2BAA2B,IAAI,CAAC,MAAM,CAAC,kBAAkB,CAAC,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YAEpF,iBAAiB;YACjB,MAAM,IAAI,CAAC,aAAa,EAAE,CAAC;QAE7B,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,MAAM,aAAa,GAAkB;gBACnC,IAAI,EAAE,eAAe;gBACrB,OAAO,EAAE,oCAAoC,KAAK,EAAE;gBACpD,QAAQ,EAAE,YAAY;gBACtB,SAAS,EAAE,KAAK;aACjB,CAAC;YACF,MAAM,aAAa,CAAC;QACtB,CAAC;IACH,CAAC;IAED;;OAEG;IACK,cAAc,CAAC,QAAmB;QACxC,IAAI,MAAM,GAAG,EAAE,CAAC;QAEhB,KAAK,MAAM,GAAG,IAAI,QAAQ,EAAE,CAAC;YAC3B,MAAM,OAAO,GAAG,OAAO,GAAG,CAAC,OAAO,KAAK,QAAQ;gBAC7C,CAAC,CAAC,GAAG,CAAC,OAAO;gBACb,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YAEnE,IAAI,GAAG,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;gBAC1B,MAAM,IAAI,eAAe,OAAO,WAAW,CAAC;YAC9C,CAAC;iBAAM,IAAI,GAAG,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;gBAC/B,MAAM,IAAI,aAAa,OAAO,WAAW,CAAC;YAC5C,CAAC;iBAAM,IAAI,GAAG,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;gBACpC,MAAM,IAAI,kBAAkB,OAAO,WAAW,CAAC;YACjD,CAAC;QACH,CAAC;QAED,MAAM,IAAI,iBAAiB,CAAC;QAC5B,OAAO,MAAM,CAAC;IAChB,CAAC;IAED;;;OAGG;IACK,iBAAiB,CAAC,SAAiB,EAAE,cAAsB;QACjE,MAAM,SAAS,GAAG,EAAE,CAAC;QACrB,MAAM,UAAU,GAAG,CAAC,CAAC;QACrB,MAAM,OAAO,GAAG,GAAG,CAAC,CAAC,kBAAkB;QACvC,MAAM,OAAO,GAAwB,EAAE,CAAC;QAExC,wDAAwD;QACxD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,SAAS,EAAE,CAAC,EAAE,EAAE,CAAC;YACnC,uDAAuD;YACvD,MAAM,UAAU,GAAG,IAAI,YAAY,CAAC,CAAC,CAAC,CAAC;YAEvC,OAAO,CAAC,mBAAmB,CAAC,MAAM,CAAC,GAAG,IAAK,GAAG,CAAC,MAAc,CAC3D,SAAS,EACT,UAAU,EACV,CAAC,SAAS,EAAE,UAAU,EAAE,CAAC,EAAE,OAAO,CAAC,CACpC,CAAC;YAEF,OAAO,CAAC,mBAAmB,CAAC,QAAQ,CAAC,GAAG,IAAK,GAAG,CAAC,MAAc,CAC7D,SAAS,EACT,UAAU,EACV,CAAC,SAAS,EAAE,UAAU,EAAE,CAAC,EAAE,OAAO,CAAC,CACpC,CAAC;QACJ,CAAC;QAED,OAAO,OAAO,CAAC;IACjB,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,IAAI,CAAC,MAAkB;QAC3B,MAAM,IAAI,CAAC,iBAAiB,EAAE,CAAC;QAE/B,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QAC7B,MAAM,MAAM,GAAG,IAAI,CAAC,cAAc,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;QAEpD,IAAI,CAAC;YACH,0CAA0C;YAC1C,MAAM,QAAQ,GAAG,IAAI,CAAC,MAAM,CAAC,MAAM,CAAC,CAAC;YACrC,OAAO,CAAC,GAAG,CAAC,oBAAoB,QAAQ,CAAC,MAAM,EAAE,CAAC,CAAC;YAEnD,2CAA2C;YAC3C,IAAI,WAAW,GAAG,IAAI,CAAC,iBAAiB,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC;YAE/C,6BAA6B;YAC7B,MAAM,WAAW,GAAG,CAAC,GAAG,QAAQ,CAAC,CAAC;YAClC,MAAM,SAAS,GAAa,EAAE,CAAC;YAE/B,8CAA8C;YAC9C,MAAM,SAAS,GAAG,QAAQ,CAAC,MAAM,GAAG,CAAC,MAAM,CAAC,SAAS,IAAI,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,CAAC;YAEhF,iCAAiC;YACjC,MAAM,YAAY,GAAG,MAAM,CAAC,SAAS,IAAI,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC;YAE/D,KAAK,IAAI,IAAI,GAAG,CAAC,EAAE,IAAI,GAAG,YAAY,EAAE,IAAI,EAAE,EAAE,CAAC;gBAC/C,kFAAkF;gBAClF,MAAM,eAAe,GAAG,IAAI,KAAK,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,SAAS,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC;gBAClF,MAAM,aAAa,GAAG,eAAe,CAAC,MAAM,CAAC;gBAE7C,uCAAuC;gBACvC,MAAM,WAAW,GAAG,IAAK,GAAG,CAAC,MAAc,CACzC,OAAO,EACP,aAAa,CAAC,IAAI,CAAC,eAAe,CAAC,GAAG,CAAC,MAAM,CAAC,CAAC,EAC/C,CAAC,CAAC,EAAE,aAAa,CAAC,CACnB,CAAC;gBAEF,yCAAyC;gBACzC,MAAM,WAAW,GAAG,WAAW,CAAC,MAAM,CAAC;gBACvC,MAAM,aAAa,GAAG,IAAK,GAAG,CAAC,MAAc,CAC3C,OAAO,EACP,aAAa,CAAC,IAAI,CAAC,KAAK,CAAC,WAAW,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC,EAC/C,CAAC,CAAC,EAAE,WAAW,CAAC,CACjB,CAAC;gBAEF,uDAAuD;gBACvD,MAAM,KAAK,GAAwB;oBACjC,SAAS,EAAE,WAAW;oBACtB,cAAc,EAAE,aAAa;oBAC7B,GAAG,WAAW;iBACf,CAAC;gBAEF,gBAAgB;gBAChB,MAAM,OAAO,GAAG,MAAM,IAAI,CAAC,OAAQ,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC;gBAE/C,4CAA4C;gBAC5C,MAAM,MAAM,GAAG,OAAO,CAAC,MAAM,CAAC,IAAoB,CAAC;gBACnD,MAAM,SAAS,GAAG,OAAO,CAAC,MAAM,CAAC,IAAI,CAAC,OAAO,CAAC,MAAM,CAAC,IAAI,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;gBAEtE,gCAAgC;gBAChC,MAAM,qBAAqB,GAAG,CAAC,aAAa,GAAG,CAAC,CAAC,GAAG,SAAS,CAAC;gBAE9D,uCAAuC;gBACvC,IAAI,SAAS,GAAG,CAAC,CAAC;gBAClB,IAAI,MAAM,GAAG,CAAC,QAAQ,CAAC;gBAEvB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,SAAS,EAAE,CAAC,EAAE,EAAE,CAAC;oBACnC,MAAM,KAAK,GAAG,MAAM,CAAC,qBAAqB,GAAG,CAAC,CAAC,GAAG,CAAC,MAAM,CAAC,WAAW,IAAI,IAAI,CAAC,MAAM,CAAC,WAAW,CAAC,CAAC;oBAClG,IAAI,KAAK,GAAG,MAAM,EAAE,CAAC;wBACnB,MAAM,GAAG,KAAK,CAAC;wBACf,SAAS,GAAG,CAAC,CAAC;oBAChB,CAAC;gBACH,CAAC;gBAED,gBAAgB;gBAChB,SAAS,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;gBAC1B,WAAW,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;gBAE5B,wDAAwD;gBACxD,IAAI,SAAS,KAAK,CAAC,IAAI,SAAS,KAAK,CAAC,EAAE,CAAC;oBACvC,OAAO,CAAC,GAAG,CAAC,2BAA2B,SAAS,EAAE,CAAC,CAAC;oBACpD,MAAM;gBACR,CAAC;gBAED,kDAAkD;gBAClD,WAAW,GAAG,EAAE,CAAC;gBACjB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,EAAE,EAAE,CAAC,EAAE,EAAE,CAAC;oBAC5B,WAAW,CAAC,mBAAmB,CAAC,MAAM,CAAC,GAAG,OAAO,CAAC,WAAW,CAAC,MAAM,CAAC,CAAC;oBACtE,WAAW,CAAC,mBAAmB,CAAC,QAAQ,CAAC,GAAG,OAAO,CAAC,WAAW,CAAC,QAAQ,CAAC,CAAC;gBAC5E,CAAC;gBAED,qBAAqB;gBACrB,IAAI,CAAC,IAAI,GAAG,CAAC,CAAC,GAAG,EAAE,KAAK,CAAC,EAAE,CAAC;oBAC1B,OAAO,CAAC,GAAG,CAAC,gBAAgB,IAAI,GAAG,CAAC,YAAY,CAAC,CAAC;gBACpD,CAAC;YACH,CAAC;YAED,yCAAyC;YACzC,MAAM,aAAa,GAAG,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,CAAC;YAC7C,MAAM,OAAO,GAAG,IAAI,CAAC,GAAG,EAAE,GAAG,SAAS,CAAC;YACvC,MAAM,eAAe,GAAG,CAAC,SAAS,CAAC,MAAM,GAAG,CAAC,OAAO,GAAG,IAAI,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC;YAEzE,OAAO,CAAC,GAAG,CAAC,gBAAgB,aAAa,EAAE,CAAC,CAAC;YAC7C,OAAO,CAAC,GAAG,CAAC,gBAAgB,OAAO,OAAO,eAAe,cAAc,CAAC,CAAC;YAEzE,MAAM,OAAO,GAAmB,CAAC;oBAC/B,IAAI,EAAE,MAAM;oBACZ,IAAI,EAAE,aAAa,CAAC,IAAI,EAAE;iBAC3B,CAAC,CAAC;YAEH,OAAO;gBACL,EAAE,EAAE,cAAc,IAAI,CAAC,GAAG,EAAE,EAAE;gBAC9B,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,SAAS;gBAC5B,OAAO;gBACP,UAAU,EAAE,UAAU;gBACtB,KAAK,EAAE;oBACL,WAAW,EAAE,QAAQ,CAAC,MAAM;oBAC5B,YAAY,EAAE,SAAS,CAAC,MAAM;iBAC/B;gBACD,QAAQ,EAAE;oBACR,QAAQ,EAAE,YAAY;oBACtB,KAAK,EAAE,0BAA0B;oBACjC,OAAO;oBACP,IAAI,EAAE,CAAC,EAAE,0BAA0B;oBACnC,kBAAkB,EAAE,IAAI,CAAC,MAAM,CAAC,kBAAkB;oBAClD,eAAe,EAAE,UAAU,CAAC,eAAe,CAAC;iBAC7C;aACF,CAAC;QAEJ,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,MAAM,aAAa,GAAkB;gBACnC,IAAI,EAAE,oBAAoB;gBAC1B,OAAO,EAAE,0BAA0B,KAAK,EAAE;gBAC1C,QAAQ,EAAE,YAAY;gBACtB,SAAS,EAAE,IAAI;aAChB,CAAC;YACF,MAAM,aAAa,CAAC;QACtB,CAAC;IACH,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,CAAC,MAAM,CAAC,MAAkB;QAC9B,MAAM,IAAI,KAAK,CAAC,wDAAwD,CAAC,CAAC;IAC5E,CAAC;IAED;;OAEG;IACH,oBAAoB,CAAC,QAAkB;QACrC,MAAM,SAAS,GAAG,CAAC,MAAM,CAAC,CAAC;QAC3B,OAAO,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC,SAAS,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;IACpD,CAAC;IAED;;OAEG;IACH,YAAY;QACV,OAAO;YACL,SAAS,EAAE,IAAI,CAAC,MAAM,CAAC,SAAS;YAChC,kBAAkB,EAAE,IAAI,CAAC,MAAM,CAAC,kBAAkB;YAClD,WAAW,EAAE,IAAI,CAAC,OAAO,KAAK,IAAI;YAClC,eAAe,EAAE,IAAI,CAAC,QAAQ,KAAK,IAAI;SACxC,CAAC;IACJ,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,OAAO;QACX,IAAI,IAAI,CAAC,OAAO,EAAE,CAAC;YACjB,gEAAgE;YAChE,IAAI,CAAC,OAAO,GAAG,IAAI,CAAC;QACtB,CAAC;QACD,IAAI,IAAI,CAAC,QAAQ,EAAE,CAAC;YAClB,IAAI,CAAC,QAAQ,CAAC,IAAI,EAAE,CAAC;YACrB,IAAI,CAAC,QAAQ,GAAG,IAAI,CAAC;QACvB,CAAC;IACH,CAAC;CACF","sourcesContent":["/**\n * ONNX Runtime Local Inference Provider for Phi-4\n *\n * Uses onnxruntime-node for true local CPU/GPU inference\n */\n\nimport * as ort from 'onnxruntime-node';\nimport * as fs from 'fs';\nimport * as path from 'path';\nimport { get_encoding } from 'tiktoken';\nimport { ensurePhi4Model, ModelDownloader } from '../../utils/model-downloader.js';\nimport type {\n  LLMProvider,\n  ChatParams,\n  ChatResponse,\n  StreamChunk,\n  ProviderError,\n  Message,\n  ContentBlock\n} from '../types.js';\n\nexport interface ONNXLocalConfig {\n  modelPath?: string;\n  executionProviders?: string[];\n  maxTokens?: number;\n  temperature?: number;\n}\n\nexport class ONNXLocalProvider implements LLMProvider {\n  name = 'onnx-local';\n  type = 'custom' as const;\n  supportsStreaming = false; // Streaming requires complex token generation loop\n  supportsTools = false;\n  supportsMCP = false;\n\n  private session: ort.InferenceSession | null = null;\n  private config: Required<ONNXLocalConfig>;\n  private tokenizer: any = null;\n  private tiktoken: any = null;\n\n  constructor(config: ONNXLocalConfig = {}) {\n    this.config = {\n      modelPath: config.modelPath || './models/phi-4/cpu_and_mobile/cpu-int4-rtn-block-32-acc-level-4/model.onnx',\n      executionProviders: config.executionProviders || ['cpu'],\n      maxTokens: config.maxTokens || 100,\n      temperature: config.temperature || 0.7\n    };\n  }\n\n  /**\n   * Load optimized tiktoken tokenizer (cl100k_base for Phi-4)\n   */\n  private async loadTokenizer(): Promise<void> {\n    if (this.tiktoken) return;\n\n    try {\n      // Use cl100k_base encoding (GPT-4, similar to Phi-4)\n      this.tiktoken = get_encoding('cl100k_base');\n\n      console.log('✅ Tokenizer loaded (tiktoken cl100k_base)');\n    } catch (error) {\n      console.error('❌ Failed to load tiktoken:', error);\n      throw new Error(`Tokenizer loading failed: ${error}`);\n    }\n  }\n\n  /**\n   * Encode text using tiktoken (fast BPE)\n   */\n  private encode(text: string): number[] {\n    return Array.from(this.tiktoken.encode(text));\n  }\n\n  /**\n   * Decode tokens using tiktoken\n   */\n  private decode(ids: number[]): string {\n    try {\n      const decoded = this.tiktoken.decode(new Uint32Array(ids));\n      // tiktoken returns buffer, convert to string\n      if (typeof decoded === 'string') {\n        return decoded;\n      } else if (decoded instanceof Uint8Array || decoded instanceof Buffer) {\n        return new TextDecoder().decode(decoded);\n      }\n      return String(decoded);\n    } catch (error) {\n      console.warn('Decode error, returning raw IDs:', error);\n      return ids.join(',');\n    }\n  }\n\n  /**\n   * Initialize ONNX session (with automatic model download)\n   */\n  private async initializeSession(): Promise<void> {\n    if (this.session) return;\n\n    try {\n      // Ensure model is downloaded\n      console.log(`🔍 Checking for Phi-4 ONNX model...`);\n\n      const modelPath = await ensurePhi4Model((progress) => {\n        if (progress.percentage % 10 < 1) { // Log every ~10%\n          console.log(`   📥 Downloading: ${ModelDownloader.formatProgress(progress)}`);\n        }\n      });\n\n      // Update config with actual model path\n      this.config.modelPath = modelPath;\n\n      console.log(`📦 Loading ONNX model: ${this.config.modelPath}`);\n\n      this.session = await ort.InferenceSession.create(\n        this.config.modelPath,\n        {\n          executionProviders: this.config.executionProviders as any,\n          graphOptimizationLevel: 'all',\n          enableCpuMemArena: true,\n          enableMemPattern: true\n        }\n      );\n\n      console.log(`✅ ONNX model loaded`);\n      console.log(`🔧 Execution providers: ${this.config.executionProviders.join(', ')}`);\n\n      // Load tokenizer\n      await this.loadTokenizer();\n\n    } catch (error) {\n      const providerError: ProviderError = {\n        name: 'ONNXInitError',\n        message: `Failed to initialize ONNX model: ${error}`,\n        provider: 'onnx-local',\n        retryable: false\n      };\n      throw providerError;\n    }\n  }\n\n  /**\n   * Format messages for Phi-4 chat template\n   */\n  private formatMessages(messages: Message[]): string {\n    let prompt = '';\n\n    for (const msg of messages) {\n      const content = typeof msg.content === 'string'\n        ? msg.content\n        : msg.content.map(c => c.type === 'text' ? c.text : '').join('');\n\n      if (msg.role === 'system') {\n        prompt += `<|system|>\\n${content}<|end|>\\n`;\n      } else if (msg.role === 'user') {\n        prompt += `<|user|>\\n${content}<|end|>\\n`;\n      } else if (msg.role === 'assistant') {\n        prompt += `<|assistant|>\\n${content}<|end|>\\n`;\n      }\n    }\n\n    prompt += '<|assistant|>\\n';\n    return prompt;\n  }\n\n  /**\n   * Initialize KV cache tensors for all 32 layers\n   * Phi-4 architecture: 32 layers, 8 KV heads, 128 head_dim\n   */\n  private initializeKVCache(batchSize: number, sequenceLength: number) {\n    const numLayers = 32;\n    const numKVHeads = 8;\n    const headDim = 128; // 3072 / 24 = 128\n    const kvCache: Record<string, any> = {};\n\n    // Initialize empty cache for each layer (key and value)\n    for (let i = 0; i < numLayers; i++) {\n      // Empty cache: [batch_size, num_kv_heads, 0, head_dim]\n      const emptyCache = new Float32Array(0);\n\n      kvCache[`past_key_values.${i}.key`] = new (ort.Tensor as any)(\n        'float32',\n        emptyCache,\n        [batchSize, numKVHeads, 0, headDim]\n      );\n\n      kvCache[`past_key_values.${i}.value`] = new (ort.Tensor as any)(\n        'float32',\n        emptyCache,\n        [batchSize, numKVHeads, 0, headDim]\n      );\n    }\n\n    return kvCache;\n  }\n\n  /**\n   * Chat completion using ONNX with KV cache\n   */\n  async chat(params: ChatParams): Promise<ChatResponse> {\n    await this.initializeSession();\n\n    const startTime = Date.now();\n    const prompt = this.formatMessages(params.messages);\n\n    try {\n      // Tokenize input using optimized tiktoken\n      const inputIds = this.encode(prompt);\n      console.log(`📝 Input tokens: ${inputIds.length}`);\n\n      // Initialize KV cache (reusable for batch)\n      let pastKVCache = this.initializeKVCache(1, 0);\n\n      // Track all generated tokens\n      const allTokenIds = [...inputIds];\n      const outputIds: number[] = [];\n\n      // Pre-allocate tensor buffers for performance\n      const maxSeqLen = inputIds.length + (params.maxTokens || this.config.maxTokens);\n\n      // Autoregressive generation loop\n      const maxNewTokens = params.maxTokens || this.config.maxTokens;\n\n      for (let step = 0; step < maxNewTokens; step++) {\n        // For first step, use all input tokens; for subsequent steps, use only last token\n        const currentInputIds = step === 0 ? inputIds : [outputIds[outputIds.length - 1]];\n        const currentSeqLen = currentInputIds.length;\n\n        // Create input tensor for current step\n        const inputTensor = new (ort.Tensor as any)(\n          'int64',\n          BigInt64Array.from(currentInputIds.map(BigInt)),\n          [1, currentSeqLen]\n        );\n\n        // Create attention mask for current step\n        const totalSeqLen = allTokenIds.length;\n        const attentionMask = new (ort.Tensor as any)(\n          'int64',\n          BigInt64Array.from(Array(totalSeqLen).fill(1n)),\n          [1, totalSeqLen]\n        );\n\n        // Build feeds with input, attention mask, and KV cache\n        const feeds: Record<string, any> = {\n          input_ids: inputTensor,\n          attention_mask: attentionMask,\n          ...pastKVCache\n        };\n\n        // Run inference\n        const results = await this.session!.run(feeds);\n\n        // Get logits for next token (last position)\n        const logits = results.logits.data as Float32Array;\n        const vocabSize = results.logits.dims[results.logits.dims.length - 1];\n\n        // Extract logits for last token\n        const lastTokenLogitsOffset = (currentSeqLen - 1) * vocabSize;\n\n        // Apply temperature and get next token\n        let nextToken = 0;\n        let maxVal = -Infinity;\n\n        for (let i = 0; i < vocabSize; i++) {\n          const logit = logits[lastTokenLogitsOffset + i] / (params.temperature || this.config.temperature);\n          if (logit > maxVal) {\n            maxVal = logit;\n            nextToken = i;\n          }\n        }\n\n        // Add to output\n        outputIds.push(nextToken);\n        allTokenIds.push(nextToken);\n\n        // Check for end token (2 is typical EOS for Phi models)\n        if (nextToken === 2 || nextToken === 0) {\n          console.log(`🛑 Stop token detected: ${nextToken}`);\n          break;\n        }\n\n        // Update KV cache from outputs for next iteration\n        pastKVCache = {};\n        for (let i = 0; i < 32; i++) {\n          pastKVCache[`past_key_values.${i}.key`] = results[`present.${i}.key`];\n          pastKVCache[`past_key_values.${i}.value`] = results[`present.${i}.value`];\n        }\n\n        // Progress indicator\n        if ((step + 1) % 10 === 0) {\n          console.log(`🔄 Generated ${step + 1} tokens...`);\n        }\n      }\n\n      // Decode output using optimized tiktoken\n      const generatedText = this.decode(outputIds);\n      const latency = Date.now() - startTime;\n      const tokensPerSecond = (outputIds.length / (latency / 1000)).toFixed(1);\n\n      console.log(`✅ Generated: ${generatedText}`);\n      console.log(`⏱️  Latency: ${latency}ms (${tokensPerSecond} tokens/sec)`);\n\n      const content: ContentBlock[] = [{\n        type: 'text',\n        text: generatedText.trim()\n      }];\n\n      return {\n        id: `onnx-local-${Date.now()}`,\n        model: this.config.modelPath,\n        content,\n        stopReason: 'end_turn',\n        usage: {\n          inputTokens: inputIds.length,\n          outputTokens: outputIds.length\n        },\n        metadata: {\n          provider: 'onnx-local',\n          model: 'Phi-4-mini-instruct-onnx',\n          latency,\n          cost: 0, // Local inference is free\n          executionProviders: this.config.executionProviders,\n          tokensPerSecond: parseFloat(tokensPerSecond)\n        }\n      };\n\n    } catch (error) {\n      const providerError: ProviderError = {\n        name: 'ONNXInferenceError',\n        message: `ONNX inference failed: ${error}`,\n        provider: 'onnx-local',\n        retryable: true\n      };\n      throw providerError;\n    }\n  }\n\n  /**\n   * Streaming not implemented (requires complex generation loop)\n   */\n  async *stream(params: ChatParams): AsyncGenerator<StreamChunk> {\n    throw new Error('Streaming not yet implemented for ONNX local inference');\n  }\n\n  /**\n   * Validate capabilities\n   */\n  validateCapabilities(features: string[]): boolean {\n    const supported = ['chat'];\n    return features.every(f => supported.includes(f));\n  }\n\n  /**\n   * Get model info\n   */\n  getModelInfo() {\n    return {\n      modelPath: this.config.modelPath,\n      executionProviders: this.config.executionProviders,\n      initialized: this.session !== null,\n      tokenizerLoaded: this.tiktoken !== null\n    };\n  }\n\n  /**\n   * Cleanup resources\n   */\n  async dispose(): Promise<void> {\n    if (this.session) {\n      // ONNX Runtime sessions don't have explicit disposal in Node.js\n      this.session = null;\n    }\n    if (this.tiktoken) {\n      this.tiktoken.free();\n      this.tiktoken = null;\n    }\n  }\n}\n"]}

package/dist/agentic-flow/src/router/providers/onnx-phi4.d.ts

@@ -0,0 +1,64 @@
+/**
+ * ONNX Runtime Provider for Phi-4 Model
+ *
+ * Hybrid implementation with fallback to HuggingFace Inference API
+ * when local ONNX model is not available
+ */
+import type { LLMProvider, ChatParams, ChatResponse, StreamChunk } from '../types.js';
+export interface ONNXPhi4Config {
+    modelId?: string;
+    useLocalONNX?: boolean;
+    huggingfaceApiKey?: string;
+    maxTokens?: number;
+    temperature?: number;
+}
+export declare class ONNXPhi4Provider implements LLMProvider {
+    name: string;
+    type: "custom";
+    supportsStreaming: boolean;
+    supportsTools: boolean;
+    supportsMCP: boolean;
+    private config;
+    private hf;
+    private modelPath;
+    constructor(config?: ONNXPhi4Config);
+    /**
+     * Format messages for Phi-4 chat template
+     */
+    private formatMessages;
+    /**
+     * Chat completion via HuggingFace Inference API
+     */
+    private chatViaAPI;
+    /**
+     * Chat completion via local ONNX (not yet implemented)
+     */
+    private chatViaONNX;
+    /**
+     * Chat completion (uses API or local ONNX based on config)
+     */
+    chat(params: ChatParams): Promise<ChatResponse>;
+    /**
+     * Streaming generation via HuggingFace API
+     */
+    stream(params: ChatParams): AsyncGenerator<StreamChunk>;
+    /**
+     * Validate capabilities
+     */
+    validateCapabilities(features: string[]): boolean;
+    /**
+     * Get model info
+     */
+    getModelInfo(): {
+        modelId: string;
+        mode: string;
+        supportsLocalInference: boolean;
+        modelPath: string;
+        apiKey: string;
+    };
+    /**
+     * Switch between API and local ONNX
+     */
+    setMode(useLocalONNX: boolean): void;
+}
+//# sourceMappingURL=onnx-phi4.d.ts.map

package/dist/agentic-flow/src/router/providers/onnx-phi4.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"onnx-phi4.d.ts","sourceRoot":"","sources":["../../../../../src/router/providers/onnx-phi4.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAGH,OAAO,KAAK,EACV,WAAW,EACX,UAAU,EACV,YAAY,EACZ,WAAW,EAIZ,MAAM,aAAa,CAAC;AAErB,MAAM,WAAW,cAAc;IAC7B,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,YAAY,CAAC,EAAE,OAAO,CAAC;IACvB,iBAAiB,CAAC,EAAE,MAAM,CAAC;IAC3B,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,WAAW,CAAC,EAAE,MAAM,CAAC;CACtB;AAED,qBAAa,gBAAiB,YAAW,WAAW;IAClD,IAAI,SAAe;IACnB,IAAI,EAAG,QAAQ,CAAU;IACzB,iBAAiB,UAAQ;IACzB,aAAa,UAAS;IACtB,WAAW,UAAS;IAEpB,OAAO,CAAC,MAAM,CAA2B;IACzC,OAAO,CAAC,EAAE,CAAc;IACxB,OAAO,CAAC,SAAS,CAAgF;gBAErF,MAAM,GAAE,cAAmB;IAYvC;;OAEG;IACH,OAAO,CAAC,cAAc;IAqBtB;;OAEG;YACW,UAAU;IA+DxB;;OAEG;YACW,WAAW;IAIzB;;OAEG;IACG,IAAI,CAAC,MAAM,EAAE,UAAU,GAAG,OAAO,CAAC,YAAY,CAAC;IAQrD;;OAEG;IACI,MAAM,CAAC,MAAM,EAAE,UAAU,GAAG,cAAc,CAAC,WAAW,CAAC;IAyC9D;;OAEG;IACH,oBAAoB,CAAC,QAAQ,EAAE,MAAM,EAAE,GAAG,OAAO;IAKjD;;OAEG;IACH,YAAY;;;;;;;IAUZ;;OAEG;IACH,OAAO,CAAC,YAAY,EAAE,OAAO;CAG9B"}

package/dist/agentic-flow/src/router/providers/onnx-phi4.js

@@ -0,0 +1,191 @@
+/**
+ * ONNX Runtime Provider for Phi-4 Model
+ *
+ * Hybrid implementation with fallback to HuggingFace Inference API
+ * when local ONNX model is not available
+ */
+import { HfInference } from '@huggingface/inference';
+export class ONNXPhi4Provider {
+    name = 'onnx-phi4';
+    type = 'custom';
+    supportsStreaming = true;
+    supportsTools = false;
+    supportsMCP = false;
+    config;
+    hf;
+    modelPath = './models/phi-4/cpu_and_mobile/cpu-int4-rtn-block-32-acc-level-4/model.onnx';
+    constructor(config = {}) {
+        this.config = {
+            modelId: config.modelId || 'microsoft/Phi-3-mini-4k-instruct',
+            useLocalONNX: config.useLocalONNX ?? false, // Default to API until local model downloaded
+            huggingfaceApiKey: config.huggingfaceApiKey || process.env.HUGGINGFACE_API_KEY || '',
+            maxTokens: config.maxTokens || 512,
+            temperature: config.temperature || 0.7
+        };
+        this.hf = new HfInference(this.config.huggingfaceApiKey);
+    }
+    /**
+     * Format messages for Phi-4 chat template
+     */
+    formatMessages(messages) {
+        let prompt = '';
+        for (const msg of messages) {
+            const content = typeof msg.content === 'string'
+                ? msg.content
+                : msg.content.map(c => c.type === 'text' ? c.text : '').join('');
+            if (msg.role === 'system') {
+                prompt += `<|system|>\n${content}<|end|>\n`;
+            }
+            else if (msg.role === 'user') {
+                prompt += `<|user|>\n${content}<|end|>\n`;
+            }
+            else if (msg.role === 'assistant') {
+                prompt += `<|assistant|>\n${content}<|end|>\n`;
+            }
+        }
+        prompt += '<|assistant|>\n';
+        return prompt;
+    }
+    /**
+     * Chat completion via HuggingFace Inference API
+     */
+    async chatViaAPI(params) {
+        const startTime = Date.now();
+        const prompt = this.formatMessages(params.messages);
+        try {
+            const result = await this.hf.textGeneration({
+                model: this.config.modelId,
+                inputs: prompt,
+                parameters: {
+                    max_new_tokens: params.maxTokens || this.config.maxTokens,
+                    temperature: params.temperature || this.config.temperature,
+                    return_full_text: false,
+                    do_sample: true,
+                    top_p: 0.9
+                }
+            });
+            const latency = Date.now() - startTime;
+            const generatedText = result.generated_text;
+            // Clean up the response
+            const assistantResponse = generatedText
+                .split('<|end|>')[0]
+                .trim();
+            // Estimate token counts
+            const inputTokens = Math.ceil(prompt.length / 4);
+            const outputTokens = Math.ceil(assistantResponse.length / 4);
+            const content = [{
+                    type: 'text',
+                    text: assistantResponse
+                }];
+            return {
+                id: `onnx-phi4-${Date.now()}`,
+                model: this.config.modelId,
+                content,
+                stopReason: 'end_turn',
+                usage: {
+                    inputTokens,
+                    outputTokens
+                },
+                metadata: {
+                    provider: 'onnx-phi4',
+                    model: this.config.modelId,
+                    latency,
+                    cost: outputTokens * 0.000002, // Rough estimate
+                    mode: 'api'
+                }
+            };
+        }
+        catch (error) {
+            const providerError = {
+                name: 'ONNXPhi4APIError',
+                message: `HuggingFace API inference failed: ${error}`,
+                provider: 'onnx-phi4',
+                retryable: true
+            };
+            throw providerError;
+        }
+    }
+    /**
+     * Chat completion via local ONNX (not yet implemented)
+     */
+    async chatViaONNX(params) {
+        throw new Error('Local ONNX inference not yet implemented. Download model.onnx.data first.');
+    }
+    /**
+     * Chat completion (uses API or local ONNX based on config)
+     */
+    async chat(params) {
+        if (this.config.useLocalONNX) {
+            return this.chatViaONNX(params);
+        }
+        else {
+            return this.chatViaAPI(params);
+        }
+    }
+    /**
+     * Streaming generation via HuggingFace API
+     */
+    async *stream(params) {
+        const prompt = this.formatMessages(params.messages);
+        try {
+            const stream = this.hf.textGenerationStream({
+                model: this.config.modelId,
+                inputs: prompt,
+                parameters: {
+                    max_new_tokens: params.maxTokens || this.config.maxTokens,
+                    temperature: params.temperature || this.config.temperature,
+                    return_full_text: false
+                }
+            });
+            for await (const chunk of stream) {
+                if (chunk.token.text) {
+                    yield {
+                        type: 'content_block_delta',
+                        delta: {
+                            type: 'text_delta',
+                            text: chunk.token.text
+                        }
+                    };
+                }
+            }
+            yield {
+                type: 'message_stop'
+            };
+        }
+        catch (error) {
+            const providerError = {
+                name: 'ONNXPhi4StreamError',
+                message: `Streaming failed: ${error}`,
+                provider: 'onnx-phi4',
+                retryable: true
+            };
+            throw providerError;
+        }
+    }
+    /**
+     * Validate capabilities
+     */
+    validateCapabilities(features) {
+        const supported = ['chat', 'stream'];
+        return features.every(f => supported.includes(f));
+    }
+    /**
+     * Get model info
+     */
+    getModelInfo() {
+        return {
+            modelId: this.config.modelId,
+            mode: this.config.useLocalONNX ? 'local-onnx' : 'api',
+            supportsLocalInference: false, // Will be true when model.onnx.data downloaded
+            modelPath: this.modelPath,
+            apiKey: this.config.huggingfaceApiKey ? '***' : undefined
+        };
+    }
+    /**
+     * Switch between API and local ONNX
+     */
+    setMode(useLocalONNX) {
+        this.config.useLocalONNX = useLocalONNX;
+    }
+}
+//# sourceMappingURL=onnx-phi4.js.map

package/dist/agentic-flow/src/router/providers/onnx-phi4.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"onnx-phi4.js","sourceRoot":"","sources":["../../../../../src/router/providers/onnx-phi4.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,EAAE,WAAW,EAAE,MAAM,wBAAwB,CAAC;AAmBrD,MAAM,OAAO,gBAAgB;IAC3B,IAAI,GAAG,WAAW,CAAC;IACnB,IAAI,GAAG,QAAiB,CAAC;IACzB,iBAAiB,GAAG,IAAI,CAAC;IACzB,aAAa,GAAG,KAAK,CAAC;IACtB,WAAW,GAAG,KAAK,CAAC;IAEZ,MAAM,CAA2B;IACjC,EAAE,CAAc;IAChB,SAAS,GAAG,4EAA4E,CAAC;IAEjG,YAAY,SAAyB,EAAE;QACrC,IAAI,CAAC,MAAM,GAAG;YACZ,OAAO,EAAE,MAAM,CAAC,OAAO,IAAI,kCAAkC;YAC7D,YAAY,EAAE,MAAM,CAAC,YAAY,IAAI,KAAK,EAAE,8CAA8C;YAC1F,iBAAiB,EAAE,MAAM,CAAC,iBAAiB,IAAI,OAAO,CAAC,GAAG,CAAC,mBAAmB,IAAI,EAAE;YACpF,SAAS,EAAE,MAAM,CAAC,SAAS,IAAI,GAAG;YAClC,WAAW,EAAE,MAAM,CAAC,WAAW,IAAI,GAAG;SACvC,CAAC;QAEF,IAAI,CAAC,EAAE,GAAG,IAAI,WAAW,CAAC,IAAI,CAAC,MAAM,CAAC,iBAAiB,CAAC,CAAC;IAC3D,CAAC;IAED;;OAEG;IACK,cAAc,CAAC,QAAmB;QACxC,IAAI,MAAM,GAAG,EAAE,CAAC;QAEhB,KAAK,MAAM,GAAG,IAAI,QAAQ,EAAE,CAAC;YAC3B,MAAM,OAAO,GAAG,OAAO,GAAG,CAAC,OAAO,KAAK,QAAQ;gBAC7C,CAAC,CAAC,GAAG,CAAC,OAAO;gBACb,CAAC,CAAC,GAAG,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YAEnE,IAAI,GAAG,CAAC,IAAI,KAAK,QAAQ,EAAE,CAAC;gBAC1B,MAAM,IAAI,eAAe,OAAO,WAAW,CAAC;YAC9C,CAAC;iBAAM,IAAI,GAAG,CAAC,IAAI,KAAK,MAAM,EAAE,CAAC;gBAC/B,MAAM,IAAI,aAAa,OAAO,WAAW,CAAC;YAC5C,CAAC;iBAAM,IAAI,GAAG,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;gBACpC,MAAM,IAAI,kBAAkB,OAAO,WAAW,CAAC;YACjD,CAAC;QACH,CAAC;QAED,MAAM,IAAI,iBAAiB,CAAC;QAC5B,OAAO,MAAM,CAAC;IAChB,CAAC;IAED;;OAEG;IACK,KAAK,CAAC,UAAU,CAAC,MAAkB;QACzC,MAAM,SAAS,GAAG,IAAI,CAAC,GAAG,EAAE,CAAC;QAC7B,MAAM,MAAM,GAAG,IAAI,CAAC,cAAc,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;QAEpD,IAAI,CAAC;YACH,MAAM,MAAM,GAAG,MAAM,IAAI,CAAC,EAAE,CAAC,cAAc,CAAC;gBAC1C,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;gBAC1B,MAAM,EAAE,MAAM;gBACd,UAAU,EAAE;oBACV,cAAc,EAAE,MAAM,CAAC,SAAS,IAAI,IAAI,CAAC,MAAM,CAAC,SAAS;oBACzD,WAAW,EAAE,MAAM,CAAC,WAAW,IAAI,IAAI,CAAC,MAAM,CAAC,WAAW;oBAC1D,gBAAgB,EAAE,KAAK;oBACvB,SAAS,EAAE,IAAI;oBACf,KAAK,EAAE,GAAG;iBACX;aACF,CAAC,CAAC;YAEH,MAAM,OAAO,GAAG,IAAI,CAAC,GAAG,EAAE,GAAG,SAAS,CAAC;YACvC,MAAM,aAAa,GAAG,MAAM,CAAC,cAAc,CAAC;YAE5C,wBAAwB;YACxB,MAAM,iBAAiB,GAAG,aAAa;iBACpC,KAAK,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC;iBACnB,IAAI,EAAE,CAAC;YAEV,wBAAwB;YACxB,MAAM,WAAW,GAAG,IAAI,CAAC,IAAI,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;YACjD,MAAM,YAAY,GAAG,IAAI,CAAC,IAAI,CAAC,iBAAiB,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;YAE7D,MAAM,OAAO,GAAmB,CAAC;oBAC/B,IAAI,EAAE,MAAM;oBACZ,IAAI,EAAE,iBAAiB;iBACxB,CAAC,CAAC;YAEH,OAAO;gBACL,EAAE,EAAE,aAAa,IAAI,CAAC,GAAG,EAAE,EAAE;gBAC7B,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;gBAC1B,OAAO;gBACP,UAAU,EAAE,UAAU;gBACtB,KAAK,EAAE;oBACL,WAAW;oBACX,YAAY;iBACb;gBACD,QAAQ,EAAE;oBACR,QAAQ,EAAE,WAAW;oBACrB,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;oBAC1B,OAAO;oBACP,IAAI,EAAE,YAAY,GAAG,QAAQ,EAAE,iBAAiB;oBAChD,IAAI,EAAE,KAAK;iBACZ;aACF,CAAC;QAEJ,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,MAAM,aAAa,GAAkB;gBACnC,IAAI,EAAE,kBAAkB;gBACxB,OAAO,EAAE,qCAAqC,KAAK,EAAE;gBACrD,QAAQ,EAAE,WAAW;gBACrB,SAAS,EAAE,IAAI;aAChB,CAAC;YACF,MAAM,aAAa,CAAC;QACtB,CAAC;IACH,CAAC;IAED;;OAEG;IACK,KAAK,CAAC,WAAW,CAAC,MAAkB;QAC1C,MAAM,IAAI,KAAK,CAAC,2EAA2E,CAAC,CAAC;IAC/F,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,IAAI,CAAC,MAAkB;QAC3B,IAAI,IAAI,CAAC,MAAM,CAAC,YAAY,EAAE,CAAC;YAC7B,OAAO,IAAI,CAAC,WAAW,CAAC,MAAM,CAAC,CAAC;QAClC,CAAC;aAAM,CAAC;YACN,OAAO,IAAI,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC;QACjC,CAAC;IACH,CAAC;IAED;;OAEG;IACH,KAAK,CAAC,CAAC,MAAM,CAAC,MAAkB;QAC9B,MAAM,MAAM,GAAG,IAAI,CAAC,cAAc,CAAC,MAAM,CAAC,QAAQ,CAAC,CAAC;QAEpD,IAAI,CAAC;YACH,MAAM,MAAM,GAAG,IAAI,CAAC,EAAE,CAAC,oBAAoB,CAAC;gBAC1C,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;gBAC1B,MAAM,EAAE,MAAM;gBACd,UAAU,EAAE;oBACV,cAAc,EAAE,MAAM,CAAC,SAAS,IAAI,IAAI,CAAC,MAAM,CAAC,SAAS;oBACzD,WAAW,EAAE,MAAM,CAAC,WAAW,IAAI,IAAI,CAAC,MAAM,CAAC,WAAW;oBAC1D,gBAAgB,EAAE,KAAK;iBACxB;aACF,CAAC,CAAC;YAEH,IAAI,KAAK,EAAE,MAAM,KAAK,IAAI,MAAM,EAAE,CAAC;gBACjC,IAAI,KAAK,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC;oBACrB,MAAM;wBACJ,IAAI,EAAE,qBAAqB;wBAC3B,KAAK,EAAE;4BACL,IAAI,EAAE,YAAY;4BAClB,IAAI,EAAE,KAAK,CAAC,KAAK,CAAC,IAAI;yBACvB;qBACF,CAAC;gBACJ,CAAC;YACH,CAAC;YAED,MAAM;gBACJ,IAAI,EAAE,cAAc;aACrB,CAAC;QAEJ,CAAC;QAAC,OAAO,KAAK,EAAE,CAAC;YACf,MAAM,aAAa,GAAkB;gBACnC,IAAI,EAAE,qBAAqB;gBAC3B,OAAO,EAAE,qBAAqB,KAAK,EAAE;gBACrC,QAAQ,EAAE,WAAW;gBACrB,SAAS,EAAE,IAAI;aAChB,CAAC;YACF,MAAM,aAAa,CAAC;QACtB,CAAC;IACH,CAAC;IAED;;OAEG;IACH,oBAAoB,CAAC,QAAkB;QACrC,MAAM,SAAS,GAAG,CAAC,MAAM,EAAE,QAAQ,CAAC,CAAC;QACrC,OAAO,QAAQ,CAAC,KAAK,CAAC,CAAC,CAAC,EAAE,CAAC,SAAS,CAAC,QAAQ,CAAC,CAAC,CAAC,CAAC,CAAC;IACpD,CAAC;IAED;;OAEG;IACH,YAAY;QACV,OAAO;YACL,OAAO,EAAE,IAAI,CAAC,MAAM,CAAC,OAAO;YAC5B,IAAI,EAAE,IAAI,CAAC,MAAM,CAAC,YAAY,CAAC,CAAC,CAAC,YAAY,CAAC,CAAC,CAAC,KAAK;YACrD,sBAAsB,EAAE,KAAK,EAAE,+CAA+C;YAC9E,SAAS,EAAE,IAAI,CAAC,SAAS;YACzB,MAAM,EAAE,IAAI,CAAC,MAAM,CAAC,iBAAiB,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,SAAS;SAC1D,CAAC;IACJ,CAAC;IAED;;OAEG;IACH,OAAO,CAAC,YAAqB;QAC3B,IAAI,CAAC,MAAM,CAAC,YAAY,GAAG,YAAY,CAAC;IAC1C,CAAC;CACF","sourcesContent":["/**\n * ONNX Runtime Provider for Phi-4 Model\n *\n * Hybrid implementation with fallback to HuggingFace Inference API\n * when local ONNX model is not available\n */\n\nimport { HfInference } from '@huggingface/inference';\nimport type {\n  LLMProvider,\n  ChatParams,\n  ChatResponse,\n  StreamChunk,\n  ProviderError,\n  Message,\n  ContentBlock\n} from '../types.js';\n\nexport interface ONNXPhi4Config {\n  modelId?: string;\n  useLocalONNX?: boolean;\n  huggingfaceApiKey?: string;\n  maxTokens?: number;\n  temperature?: number;\n}\n\nexport class ONNXPhi4Provider implements LLMProvider {\n  name = 'onnx-phi4';\n  type = 'custom' as const;\n  supportsStreaming = true;\n  supportsTools = false;\n  supportsMCP = false;\n\n  private config: Required<ONNXPhi4Config>;\n  private hf: HfInference;\n  private modelPath = './models/phi-4/cpu_and_mobile/cpu-int4-rtn-block-32-acc-level-4/model.onnx';\n\n  constructor(config: ONNXPhi4Config = {}) {\n    this.config = {\n      modelId: config.modelId || 'microsoft/Phi-3-mini-4k-instruct',\n      useLocalONNX: config.useLocalONNX ?? false, // Default to API until local model downloaded\n      huggingfaceApiKey: config.huggingfaceApiKey || process.env.HUGGINGFACE_API_KEY || '',\n      maxTokens: config.maxTokens || 512,\n      temperature: config.temperature || 0.7\n    };\n\n    this.hf = new HfInference(this.config.huggingfaceApiKey);\n  }\n\n  /**\n   * Format messages for Phi-4 chat template\n   */\n  private formatMessages(messages: Message[]): string {\n    let prompt = '';\n\n    for (const msg of messages) {\n      const content = typeof msg.content === 'string'\n        ? msg.content\n        : msg.content.map(c => c.type === 'text' ? c.text : '').join('');\n\n      if (msg.role === 'system') {\n        prompt += `<|system|>\\n${content}<|end|>\\n`;\n      } else if (msg.role === 'user') {\n        prompt += `<|user|>\\n${content}<|end|>\\n`;\n      } else if (msg.role === 'assistant') {\n        prompt += `<|assistant|>\\n${content}<|end|>\\n`;\n      }\n    }\n\n    prompt += '<|assistant|>\\n';\n    return prompt;\n  }\n\n  /**\n   * Chat completion via HuggingFace Inference API\n   */\n  private async chatViaAPI(params: ChatParams): Promise<ChatResponse> {\n    const startTime = Date.now();\n    const prompt = this.formatMessages(params.messages);\n\n    try {\n      const result = await this.hf.textGeneration({\n        model: this.config.modelId,\n        inputs: prompt,\n        parameters: {\n          max_new_tokens: params.maxTokens || this.config.maxTokens,\n          temperature: params.temperature || this.config.temperature,\n          return_full_text: false,\n          do_sample: true,\n          top_p: 0.9\n        }\n      });\n\n      const latency = Date.now() - startTime;\n      const generatedText = result.generated_text;\n\n      // Clean up the response\n      const assistantResponse = generatedText\n        .split('<|end|>')[0]\n        .trim();\n\n      // Estimate token counts\n      const inputTokens = Math.ceil(prompt.length / 4);\n      const outputTokens = Math.ceil(assistantResponse.length / 4);\n\n      const content: ContentBlock[] = [{\n        type: 'text',\n        text: assistantResponse\n      }];\n\n      return {\n        id: `onnx-phi4-${Date.now()}`,\n        model: this.config.modelId,\n        content,\n        stopReason: 'end_turn',\n        usage: {\n          inputTokens,\n          outputTokens\n        },\n        metadata: {\n          provider: 'onnx-phi4',\n          model: this.config.modelId,\n          latency,\n          cost: outputTokens * 0.000002, // Rough estimate\n          mode: 'api'\n        }\n      };\n\n    } catch (error) {\n      const providerError: ProviderError = {\n        name: 'ONNXPhi4APIError',\n        message: `HuggingFace API inference failed: ${error}`,\n        provider: 'onnx-phi4',\n        retryable: true\n      };\n      throw providerError;\n    }\n  }\n\n  /**\n   * Chat completion via local ONNX (not yet implemented)\n   */\n  private async chatViaONNX(params: ChatParams): Promise<ChatResponse> {\n    throw new Error('Local ONNX inference not yet implemented. Download model.onnx.data first.');\n  }\n\n  /**\n   * Chat completion (uses API or local ONNX based on config)\n   */\n  async chat(params: ChatParams): Promise<ChatResponse> {\n    if (this.config.useLocalONNX) {\n      return this.chatViaONNX(params);\n    } else {\n      return this.chatViaAPI(params);\n    }\n  }\n\n  /**\n   * Streaming generation via HuggingFace API\n   */\n  async *stream(params: ChatParams): AsyncGenerator<StreamChunk> {\n    const prompt = this.formatMessages(params.messages);\n\n    try {\n      const stream = this.hf.textGenerationStream({\n        model: this.config.modelId,\n        inputs: prompt,\n        parameters: {\n          max_new_tokens: params.maxTokens || this.config.maxTokens,\n          temperature: params.temperature || this.config.temperature,\n          return_full_text: false\n        }\n      });\n\n      for await (const chunk of stream) {\n        if (chunk.token.text) {\n          yield {\n            type: 'content_block_delta',\n            delta: {\n              type: 'text_delta',\n              text: chunk.token.text\n            }\n          };\n        }\n      }\n\n      yield {\n        type: 'message_stop'\n      };\n\n    } catch (error) {\n      const providerError: ProviderError = {\n        name: 'ONNXPhi4StreamError',\n        message: `Streaming failed: ${error}`,\n        provider: 'onnx-phi4',\n        retryable: true\n      };\n      throw providerError;\n    }\n  }\n\n  /**\n   * Validate capabilities\n   */\n  validateCapabilities(features: string[]): boolean {\n    const supported = ['chat', 'stream'];\n    return features.every(f => supported.includes(f));\n  }\n\n  /**\n   * Get model info\n   */\n  getModelInfo() {\n    return {\n      modelId: this.config.modelId,\n      mode: this.config.useLocalONNX ? 'local-onnx' : 'api',\n      supportsLocalInference: false, // Will be true when model.onnx.data downloaded\n      modelPath: this.modelPath,\n      apiKey: this.config.huggingfaceApiKey ? '***' : undefined\n    };\n  }\n\n  /**\n   * Switch between API and local ONNX\n   */\n  setMode(useLocalONNX: boolean) {\n    this.config.useLocalONNX = useLocalONNX;\n  }\n}\n"]}