llama-cpp-capacitor 0.0.13 → 0.0.21

package/android/src/main/jni.cpp

@@ -1,16 +1,21 @@
 #include "jni-utils.h"
-#include "rn-llama.h"
+#include "cap-llama.h"
+#include "cap-completion.h"
 #include <android/log.h>
 #include <cstring>
 #include <memory>
 #include <fstream> // Added for file existence and size checks
 #include <signal.h> // Added for signal handling
 #include <sys/signal.h> // Added for sigaction
+#include <thread> // For background downloads
+#include <atomic> // For thread-safe progress tracking
+#include <filesystem> // For file operations
+#include <mutex> // For thread synchronization
 
 // Add missing symbol
-namespace rnllama {
-    bool rnllama_verbose = false;
-}
+// namespace rnllama {
+//     bool rnllama_verbose = false;
+// }
 
 #define LOG_TAG "LlamaCpp"
 #define LOGI(...) __android_log_print(ANDROID_LOG_INFO, LOG_TAG, __VA_ARGS__)
@@ -130,75 +135,67 @@ jclass find_class(JNIEnv* env, const char* name) {
     return clazz;
 }
 
-// Global context storage
-static std::map<jlong, std::unique_ptr<rnllama::llama_rn_context>> contexts;
+// Convert llama_cap_context to jobject
+jobject llama_context_to_jobject(JNIEnv* env, const capllama::llama_cap_context* context);
+
+// Convert jobject to llama_cap_context
+capllama::llama_cap_context* jobject_to_llama_context(JNIEnv* env, jobject obj);
+
+// Convert completion result to jobject
+jobject completion_result_to_jobject(JNIEnv* env, const capllama::completion_token_output& result);
+
+// Convert tokenize result to jobject
+jobject tokenize_result_to_jobject(JNIEnv* env, const capllama::llama_cap_tokenize_result& result);
+
+// Global context storage - fix namespace
+static std::map<jlong, std::unique_ptr<capllama::llama_cap_context>> contexts;
 static jlong next_context_id = 1;
 
+// Download progress tracking (simplified for now)
+// This can be enhanced later to track actual download progress
+
 extern "C" {
 
 JNIEXPORT jlong JNICALL
 Java_ai_annadata_plugin_capacitor_LlamaCpp_initContextNative(
-    JNIEnv* env, jobject thiz, jstring model_path, jobject params) {
+    JNIEnv *env, jobject thiz, jstring modelPath, jobjectArray searchPaths, jobject params) {
     
     try {
-        std::string model_path_str = jstring_to_string(env, model_path);
-        LOGI("Attempting to load model from path: %s", model_path_str.c_str());
-
-        // List all possible paths we should check
-        std::vector<std::string> paths_to_check = {
-            model_path_str,
-            "/data/data/ai.annadata.app/files/" + model_path_str,
-            "/data/data/ai.annadata.app/files/Documents/" + model_path_str,
-            "/storage/emulated/0/Android/data/ai.annadata.app/files/" + model_path_str,
-            "/storage/emulated/0/Android/data/ai.annadata.app/files/Documents/" + model_path_str,
-            "/storage/emulated/0/Documents/" + model_path_str
-        };
-
-        // Check each path and log what we find
+        std::string model_path_str = jstring_to_string(env, modelPath);
+        
+        // Get search paths from Java
+        jsize pathCount = env->GetArrayLength(searchPaths);
+        std::vector<std::string> paths_to_check;
+        
+        // Add the original path first
+        paths_to_check.push_back(model_path_str);
+        
+        // Add all search paths from Java
+        for (jsize i = 0; i < pathCount; i++) {
+            jstring pathJString = (jstring)env->GetObjectArrayElement(searchPaths, i);
+            std::string path = jstring_to_string(env, pathJString);
+            paths_to_check.push_back(path);
+            env->DeleteLocalRef(pathJString);
+        }
+        
+        // Rest of the existing logic remains the same...
         std::string full_model_path;
         bool file_found = false;
         
         for (const auto& path : paths_to_check) {
             LOGI("Checking path: %s", path.c_str());
-            std::ifstream file_check(path);
-            if (file_check.good()) {
-                file_check.seekg(0, std::ios::end);
-                std::streamsize file_size = file_check.tellg();
-                file_check.close();
-                LOGI("Found file at: %s, size: %ld bytes", path.c_str(), file_size);
-                
-                // Validate file size
-                if (file_size < 1024 * 1024) { // Less than 1MB
-                    LOGE("Model file is too small, likely corrupted: %s", path.c_str());
-                    continue; // Try next path
-                }
-                
-                // Check if it's a valid GGUF file by reading the magic number
-                std::ifstream magic_file(path, std::ios::binary);
-                if (magic_file.good()) {
-                    char magic[4];
-                    if (magic_file.read(magic, 4)) {
-                        if (magic[0] == 'G' && magic[1] == 'G' && magic[2] == 'U' && magic[3] == 'F') {
-                            LOGI("Valid GGUF file detected at: %s", path.c_str());
-                            full_model_path = path;
-                            file_found = true;
-                            break;
-                        } else {
-                            LOGI("File does not appear to be a GGUF file (magic: %c%c%c%c) at: %s", 
-                                 magic[0], magic[1], magic[2], magic[3], path.c_str());
-                        }
-                    }
-                    magic_file.close();
-                }
+            if (std::filesystem::exists(path)) {
+                full_model_path = path;
+                file_found = true;
+                LOGI("Found model file at: %s", path.c_str());
+                break;
             } else {
-                LOGI("File not found at: %s", path.c_str());
+                LOGE("Path not found: %s", path.c_str());
             }
-            file_check.close();
         }
-
+        
         if (!file_found) {
-            LOGE("Model file not found in any of the checked paths");
-            throw_java_exception(env, "java/lang/RuntimeException", "Model file not found in any expected location");
+            LOGE("Model file not found in any of the search paths");
             return -1;
         }
         
@@ -221,9 +218,9 @@ Java_ai_annadata_plugin_capacitor_LlamaCpp_initContextNative(
             validation_file.close();
         }
 
-        // Create new context
-        auto context = std::make_unique<rnllama::llama_rn_context>();
-        LOGI("Created llama_rn_context");
+        // Create new context - fix namespace
+        auto context = std::make_unique<capllama::llama_cap_context>();
+        LOGI("Created llama_cap_context");
         
         // Initialize common parameters
         common_params cparams;
@@ -240,47 +237,21 @@ Java_ai_annadata_plugin_capacitor_LlamaCpp_initContextNative(
         cparams.chat_template = "";
         cparams.embedding = false;
         cparams.cont_batching = false;
-        cparams.parallel = false;
-        cparams.grammar = "";
-        cparams.grammar_penalty.clear();
+        cparams.n_parallel = 1;
         cparams.antiprompt.clear();
-        cparams.lora_adapter.clear();
-        cparams.lora_base = "";
-        cparams.mul_mat_q = true;
-        cparams.f16_kv = true;
-        cparams.logits_all = false;
         cparams.vocab_only = false;
         cparams.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
-        cparams.rope_scaling_factor = 0.0f;
-        cparams.rope_scaling_orig_ctx_len = 0;
         cparams.yarn_ext_factor = -1.0f;
         cparams.yarn_attn_factor = 1.0f;
         cparams.yarn_beta_fast = 32.0f;
         cparams.yarn_beta_slow = 1.0f;
         cparams.yarn_orig_ctx = 0;
-        cparams.offload_kqv = true;
         cparams.flash_attn = false;
-        cparams.flash_attn_kernel = false;
-        cparams.flash_attn_causal = true;
-        cparams.mmproj = "";
-        cparams.image = "";
-        cparams.export = "";
-        cparams.export_path = "";
-        cparams.seed = -1;
         cparams.n_keep = 0;
-        cparams.n_discard = -1;
-        cparams.n_draft = 0;
         cparams.n_chunks = -1;
-        cparams.n_parallel = 1;
         cparams.n_sequences = 1;
-        cparams.p_accept = 0.5f;
-        cparams.p_split = 0.1f;
-        cparams.n_gqa = 8;
-        cparams.rms_norm_eps = 5e-6f;
         cparams.model_alias = "unknown";
-        cparams.ubatch_size = 512;
-        cparams.ubatch_seq_len_max = 1;
-        
+
         LOGI("Initialized common parameters, attempting to load model from: %s", full_model_path.c_str());
         LOGI("Model parameters: n_ctx=%d, n_batch=%d, n_gpu_layers=%d", 
              cparams.n_ctx, cparams.n_batch, cparams.n_gpu_layers);
@@ -335,47 +306,21 @@ Java_ai_annadata_plugin_capacitor_LlamaCpp_initContextNative(
             ultra_minimal_params.chat_template = "";
             ultra_minimal_params.embedding = false;
             ultra_minimal_params.cont_batching = false;
-            ultra_minimal_params.parallel = false;
-            ultra_minimal_params.grammar = "";
-            ultra_minimal_params.grammar_penalty.clear();
+            ultra_minimal_params.n_parallel = 1;
             ultra_minimal_params.antiprompt.clear();
-            ultra_minimal_params.lora_adapter.clear();
-            ultra_minimal_params.lora_base = "";
-            ultra_minimal_params.mul_mat_q = false; // Disable quantized matrix multiplication
-            ultra_minimal_params.f16_kv = false; // Disable f16 key-value cache
-            ultra_minimal_params.logits_all = false;
             ultra_minimal_params.vocab_only = false;
             ultra_minimal_params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
-            ultra_minimal_params.rope_scaling_factor = 0.0f;
-            ultra_minimal_params.rope_scaling_orig_ctx_len = 0;
             ultra_minimal_params.yarn_ext_factor = -1.0f;
             ultra_minimal_params.yarn_attn_factor = 1.0f;
             ultra_minimal_params.yarn_beta_fast = 32.0f;
             ultra_minimal_params.yarn_beta_slow = 1.0f;
             ultra_minimal_params.yarn_orig_ctx = 0;
-            ultra_minimal_params.offload_kqv = false; // Disable offloading
             ultra_minimal_params.flash_attn = false;
-            ultra_minimal_params.flash_attn_kernel = false;
-            ultra_minimal_params.flash_attn_causal = true;
-            ultra_minimal_params.mmproj = "";
-            ultra_minimal_params.image = "";
-            ultra_minimal_params.export = "";
-            ultra_minimal_params.export_path = "";
-            ultra_minimal_params.seed = -1;
             ultra_minimal_params.n_keep = 0;
-            ultra_minimal_params.n_discard = -1;
-            ultra_minimal_params.n_draft = 0;
             ultra_minimal_params.n_chunks = -1;
-            ultra_minimal_params.n_parallel = 1;
             ultra_minimal_params.n_sequences = 1;
-            ultra_minimal_params.p_accept = 0.5f;
-            ultra_minimal_params.p_split = 0.1f;
-            ultra_minimal_params.n_gqa = 8;
-            ultra_minimal_params.rms_norm_eps = 5e-6f;
             ultra_minimal_params.model_alias = "unknown";
-            ultra_minimal_params.ubatch_size = 128;
-            ultra_minimal_params.ubatch_seq_len_max = 1;
-            
+
             // Set up signal handler again for ultra-minimal attempt
             if (sigaction(SIGSEGV, &new_action, &old_action) == 0) {
                 LOGI("Signal handler reinstalled for ultra-minimal attempt");
@@ -435,28 +380,400 @@ Java_ai_annadata_plugin_capacitor_LlamaCpp_releaseContextNative(
     }
 }
 
-JNIEXPORT jstring JNICALL
+JNIEXPORT jobject JNICALL
 Java_ai_annadata_plugin_capacitor_LlamaCpp_completionNative(
-    JNIEnv* env, jobject thiz, jlong context_id, jstring prompt) {
+    JNIEnv* env, jobject thiz, jlong context_id, jobject params) {
     
     try {
+        LOGI("Starting completion for context: %ld", context_id);
+        
         auto it = contexts.find(context_id);
         if (it == contexts.end()) {
+            LOGE("Context not found: %ld", context_id);
             throw_java_exception(env, "java/lang/IllegalArgumentException", "Invalid context ID");
             return nullptr;
         }
         
-        std::string prompt_str = jstring_to_string(env, prompt);
+        auto& ctx = it->second;
+        if (!ctx || !ctx->ctx) {
+            LOGE("Invalid context or llama context is null");
+            throw_java_exception(env, "java/lang/RuntimeException", "Invalid context");
+            return nullptr;
+        }
+        
+        // Extract parameters from JSObject using compatible API
+        jclass jsObjectClass = env->GetObjectClass(params);
         
-        // Get the context
-        rnllama::llama_rn_context* context = it->second.get();
+        // Try to get method IDs and handle exceptions
+        jmethodID getStringMethod = nullptr;
+        jmethodID getIntegerMethod = nullptr; 
+        jmethodID getDoubleMethod = nullptr;
         
-        // For now, return a simple completion
-        // In a full implementation, this would use the actual llama.cpp completion logic
-        std::string result = "Generated response for: " + prompt_str;
+        // Clear any pending exceptions first
+        if (env->ExceptionCheck()) {
+            env->ExceptionClear();
+        }
         
-        LOGI("Completion for context %ld: %s", context_id, prompt_str.c_str());
-        return string_to_jstring(env, result);
+        try {
+            getStringMethod = env->GetMethodID(jsObjectClass, "getString", "(Ljava/lang/String;)Ljava/lang/String;");
+            if (env->ExceptionCheck()) {
+                env->ExceptionClear();
+                getStringMethod = nullptr;
+            }
+            
+            getIntegerMethod = env->GetMethodID(jsObjectClass, "getInteger", "(Ljava/lang/String;)Ljava/lang/Integer;");
+            if (env->ExceptionCheck()) {
+                env->ExceptionClear();
+                getIntegerMethod = nullptr;
+            }
+            
+            getDoubleMethod = env->GetMethodID(jsObjectClass, "getDouble", "(Ljava/lang/String;)Ljava/lang/Double;");
+            if (env->ExceptionCheck()) {
+                env->ExceptionClear();
+                getDoubleMethod = nullptr;
+            }
+        } catch (...) {
+            LOGE("Exception getting JSObject method IDs");
+            if (env->ExceptionCheck()) {
+                env->ExceptionClear();
+            }
+        }
+        
+        // Get prompt with safe method calls
+        std::string prompt_str = "Once upon a time";
+        jint n_predict = 50;
+        jdouble temperature = 0.7;
+        
+        if (getStringMethod) {
+            jstring promptKey = jni_utils::string_to_jstring(env, "prompt");
+            jstring promptObj = (jstring)env->CallObjectMethod(params, getStringMethod, promptKey);
+            if (promptObj && !env->ExceptionCheck()) {
+                prompt_str = jni_utils::jstring_to_string(env, promptObj);
+            } else if (env->ExceptionCheck()) {
+                env->ExceptionClear();
+            }
+        }
+        
+        // Get n_predict with safe method calls
+        if (getIntegerMethod) {
+            jstring nPredictKey = jni_utils::string_to_jstring(env, "n_predict");
+            jobject nPredictObj = env->CallObjectMethod(params, getIntegerMethod, nPredictKey);
+            if (nPredictObj && !env->ExceptionCheck()) {
+                n_predict = env->CallIntMethod(nPredictObj, env->GetMethodID(env->FindClass("java/lang/Integer"), "intValue", "()I"));
+                if (env->ExceptionCheck()) {
+                    env->ExceptionClear();
+                    n_predict = 50; // fallback
+                }
+            } else if (env->ExceptionCheck()) {
+                env->ExceptionClear();
+            }
+        }
+        
+        // Get temperature with safe method calls
+        if (getDoubleMethod) {
+            jstring temperatureKey = jni_utils::string_to_jstring(env, "temperature");
+            jobject tempObj = env->CallObjectMethod(params, getDoubleMethod, temperatureKey);
+            if (tempObj && !env->ExceptionCheck()) {
+                temperature = env->CallDoubleMethod(tempObj, env->GetMethodID(env->FindClass("java/lang/Double"), "doubleValue", "()D"));
+                if (env->ExceptionCheck()) {
+                    env->ExceptionClear();
+                    temperature = 0.7; // fallback
+                }
+            } else if (env->ExceptionCheck()) {
+                env->ExceptionClear();
+            }
+        }
+        
+        LOGI("Completion params - prompt: %s, n_predict: %d, temperature: %.2f", 
+             prompt_str.c_str(), n_predict, temperature);
+        
+        // Set sampling parameters based on extracted values
+        ctx->params.sampling.temp = temperature;
+        ctx->params.sampling.top_k = 40;  // Default value
+        ctx->params.sampling.top_p = 0.95f; // Default value
+        ctx->params.sampling.penalty_repeat = 1.1f; // Default value (correct field name)
+        ctx->params.n_predict = n_predict;
+        ctx->params.prompt = prompt_str;
+        
+        LOGI("Updated context sampling params - temp: %.2f, top_k: %d, top_p: %.2f", 
+             ctx->params.sampling.temp, ctx->params.sampling.top_k, ctx->params.sampling.top_p);
+        
+        // Tokenize the prompt
+        capllama::llama_cap_tokenize_result tokenize_result = ctx->tokenize(prompt_str, {});
+        std::vector<llama_token> prompt_tokens = tokenize_result.tokens;
+        
+        LOGI("Tokenized prompt into %zu tokens", prompt_tokens.size());
+        
+        // Initialize completion context if not already done
+        if (!ctx->completion) {
+            LOGI("Initializing completion context for the first time");
+            
+            // Validate parent context before creating completion
+            if (!ctx->ctx || !ctx->model) {
+                LOGE("Parent context is invalid - missing llama context or model");
+                throw_java_exception(env, "java/lang/RuntimeException", "Parent context is not properly initialized");
+                return nullptr;
+            }
+            
+            try {
+                LOGI("Creating llama_cap_context_completion...");
+                LOGI("Parent context pointer: %p", ctx.get());
+                LOGI("Parent context->ctx: %p", ctx->ctx);
+                LOGI("Parent context->model: %p", ctx->model);
+                
+                // Additional safety checks before constructor
+                if (!ctx.get()) {
+                    LOGE("Parent context pointer is null");
+                    throw_java_exception(env, "java/lang/RuntimeException", "Parent context pointer is null");
+                    return nullptr;
+                }
+                
+                ctx->completion = new capllama::llama_cap_context_completion(ctx.get());
+                
+                if (!ctx->completion) {
+                    LOGE("Failed to create completion context - constructor returned null");
+                    throw_java_exception(env, "java/lang/RuntimeException", "Failed to create completion context");
+                    return nullptr;
+                }
+                
+                LOGI("Completion context created successfully at: %p", ctx->completion);
+                
+                LOGI("Initializing sampling for completion context...");
+                LOGI("Parent context params before initSampling - model: %p, params: %p", ctx->model, &(ctx->params));
+                LOGI("Parent context sampling params - temperature: %.2f, top_k: %d, top_p: %.2f", 
+                     ctx->params.sampling.temp, ctx->params.sampling.top_k, ctx->params.sampling.top_p);
+                
+                bool sampling_result = false;
+                try {
+                    sampling_result = ctx->completion->initSampling();
+                    LOGI("initSampling completed, result: %s", sampling_result ? "true" : "false");
+                    LOGI("Sampler pointer after init: %p", ctx->completion->ctx_sampling);
+                } catch (const std::exception& e) {
+                    LOGE("Exception in initSampling: %s", e.what());
+                    delete ctx->completion;
+                    ctx->completion = nullptr;
+                    throw_java_exception(env, "java/lang/RuntimeException", 
+                        ("Failed to initialize sampling: " + std::string(e.what())).c_str());
+                    return nullptr;
+                } catch (...) {
+                    LOGE("Unknown exception in initSampling");
+                    delete ctx->completion;
+                    ctx->completion = nullptr;
+                    throw_java_exception(env, "java/lang/RuntimeException", "Unknown error in sampling initialization");
+                    return nullptr;
+                }
+                
+                if (!sampling_result || !ctx->completion->ctx_sampling) {
+                    LOGE("Failed to initialize sampling - result: %s, sampler: %p", 
+                         sampling_result ? "true" : "false", ctx->completion->ctx_sampling);
+                    delete ctx->completion;
+                    ctx->completion = nullptr;
+                    throw_java_exception(env, "java/lang/RuntimeException", "Failed to initialize sampling context");
+                    return nullptr;
+                }
+                
+                LOGI("Completion context initialized successfully");
+            } catch (const std::exception& e) {
+                LOGE("Exception during completion context creation: %s", e.what());
+                if (ctx->completion) {
+                    delete ctx->completion;
+                    ctx->completion = nullptr;
+                }
+                throw_java_exception(env, "java/lang/RuntimeException", 
+                    ("Failed to create completion context: " + std::string(e.what())).c_str());
+                return nullptr;
+            } catch (...) {
+                LOGE("Unknown exception during completion context creation");
+                if (ctx->completion) {
+                    delete ctx->completion;
+                    ctx->completion = nullptr;
+                }
+                throw_java_exception(env, "java/lang/RuntimeException", "Unknown error during completion context creation");
+                return nullptr;
+            }
+        }
+        
+        // Set up sampling parameters
+        // Note: For now, we'll use the completion context's default parameters
+        // TODO: Update sampling parameters with user values
+        // 
+            // Declare variables outside try block so they're accessible later
+        std::string generated_text;
+        int tokens_generated = 0;
+        
+        try {
+            LOGI("Rewinding completion context...");
+            try {
+                ctx->completion->rewind();
+                LOGI("Rewind completed successfully");
+            } catch (const std::exception& e) {
+                LOGE("Exception in rewind: %s", e.what());
+                throw;
+            }
+            
+            LOGI("Loading prompt into completion context...");
+            try {
+                // Validate sampler is properly initialized before loadPrompt
+                if (!ctx->completion->ctx_sampling) {
+                    LOGE("Sampler context is null - reinitializing");
+                    if (!ctx->completion->initSampling()) {
+                        LOGE("Failed to reinitialize sampling");
+                        throw std::runtime_error("Sampler initialization failed");
+                    }
+                    LOGI("Sampler reinitialized successfully");
+                }
+                
+                ctx->completion->loadPrompt({});
+                LOGI("loadPrompt completed successfully");
+            } catch (const std::exception& e) {
+                LOGE("Exception in loadPrompt: %s", e.what());
+                throw;
+            }
+            
+            LOGI("Beginning completion generation...");
+            try {
+                ctx->completion->beginCompletion();
+                LOGI("beginCompletion completed successfully");
+            } catch (const std::exception& e) {
+                LOGE("Exception in beginCompletion: %s", e.what());
+                throw;
+            }
+            
+            LOGI("Starting token generation loop (max tokens: %d)...", n_predict);
+            
+            while (tokens_generated < n_predict && !ctx->completion->is_interrupted) {
+                try {
+                    LOGI("Generating token %d...", tokens_generated + 1);
+                    auto token_output = ctx->completion->nextToken();
+                    
+                    // Check for end-of-sequence (simplified check)
+                    if (token_output.tok == 2) { // Most models use 2 as EOS token
+                        LOGI("Reached EOS token, stopping generation");
+                        break;
+                    }
+                    
+                    // Convert token to text
+                    std::string token_text = capllama::tokens_to_output_formatted_string(ctx->ctx, token_output.tok);
+                    generated_text += token_text;
+                    tokens_generated++;
+                    
+                    LOGI("Generated token %d (ID: %d): %s", tokens_generated, token_output.tok, token_text.c_str());
+                    
+                } catch (const std::exception& e) {
+                    LOGE("Exception during token generation %d: %s", tokens_generated + 1, e.what());
+                    break;
+                } catch (...) {
+                    LOGE("Unknown exception during token generation %d", tokens_generated + 1);
+                    break;
+                }
+            }
+            
+            LOGI("Token generation completed. Generated %d tokens.", tokens_generated);
+            
+            // End completion
+            LOGI("Ending completion...");
+            ctx->completion->endCompletion();
+            
+        } catch (const std::exception& e) {
+            LOGE("Exception during completion process: %s", e.what());
+            try {
+                ctx->completion->endCompletion();
+            } catch (...) {
+                LOGE("Failed to properly end completion after exception");
+            }
+            throw_java_exception(env, "java/lang/RuntimeException", 
+                ("Completion process failed: " + std::string(e.what())).c_str());
+            return nullptr;
+        } catch (...) {
+            LOGE("Unknown exception during completion process");
+            try {
+                ctx->completion->endCompletion();
+            } catch (...) {
+                LOGE("Failed to properly end completion after unknown exception");
+            }
+            throw_java_exception(env, "java/lang/RuntimeException", "Unknown error during completion process");
+            return nullptr;
+        }
+        
+        LOGI("Completion finished. Generated %d tokens: %s", tokens_generated, generated_text.c_str());
+        
+        // Create result HashMap
+        jclass hashMapClass = env->FindClass("java/util/HashMap");
+        jmethodID hashMapConstructor = env->GetMethodID(hashMapClass, "<init>", "()V");
+        jmethodID putMethod = env->GetMethodID(hashMapClass, "put", "(Ljava/lang/Object;Ljava/lang/Object;)Ljava/lang/Object;");
+        
+        jobject resultMap = env->NewObject(hashMapClass, hashMapConstructor);
+        
+        // Add completion results
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "text"), jni_utils::string_to_jstring(env, generated_text));
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "content"), jni_utils::string_to_jstring(env, generated_text));
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "reasoning_content"), jni_utils::string_to_jstring(env, ""));
+        
+        // Create empty tool_calls array
+        jclass arrayListClass = env->FindClass("java/util/ArrayList");
+        jmethodID arrayListConstructor = env->GetMethodID(arrayListClass, "<init>", "()V");
+        jobject emptyToolCalls = env->NewObject(arrayListClass, arrayListConstructor);
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "tool_calls"), emptyToolCalls);
+        
+        // Add token counts and status
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "tokens_predicted"), 
+            env->NewObject(env->FindClass("java/lang/Integer"), 
+                env->GetMethodID(env->FindClass("java/lang/Integer"), "<init>", "(I)V"), tokens_generated));
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "tokens_evaluated"), 
+            env->NewObject(env->FindClass("java/lang/Integer"), 
+                env->GetMethodID(env->FindClass("java/lang/Integer"), "<init>", "(I)V"), (jint)prompt_tokens.size()));
+        
+        // Add completion status flags
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "truncated"), 
+            env->NewObject(env->FindClass("java/lang/Boolean"), 
+                env->GetMethodID(env->FindClass("java/lang/Boolean"), "<init>", "(Z)V"), JNI_FALSE));
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "stopped_eos"), 
+            env->NewObject(env->FindClass("java/lang/Boolean"), 
+                env->GetMethodID(env->FindClass("java/lang/Boolean"), "<init>", "(Z)V"), 
+                tokens_generated < n_predict ? JNI_TRUE : JNI_FALSE));
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "stopped_limit"), 
+            env->NewObject(env->FindClass("java/lang/Boolean"), 
+                env->GetMethodID(env->FindClass("java/lang/Boolean"), "<init>", "(Z)V"), 
+                tokens_generated >= n_predict ? JNI_TRUE : JNI_FALSE));
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "context_full"), 
+            env->NewObject(env->FindClass("java/lang/Boolean"), 
+                env->GetMethodID(env->FindClass("java/lang/Boolean"), "<init>", "(Z)V"), JNI_FALSE));
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "interrupted"), 
+            env->NewObject(env->FindClass("java/lang/Boolean"), 
+                env->GetMethodID(env->FindClass("java/lang/Boolean"), "<init>", "(Z)V"), JNI_FALSE));
+        
+        // Add empty strings for stop reasons
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "stopped_word"), jni_utils::string_to_jstring(env, ""));
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "stopping_word"), jni_utils::string_to_jstring(env, ""));
+        
+        // Add timing information (basic)
+        jobject timingsMap = env->NewObject(hashMapClass, hashMapConstructor);
+        env->CallObjectMethod(timingsMap, putMethod,
+            jni_utils::string_to_jstring(env, "prompt_n"), 
+            env->NewObject(env->FindClass("java/lang/Integer"), 
+                env->GetMethodID(env->FindClass("java/lang/Integer"), "<init>", "(I)V"), (jint)prompt_tokens.size()));
+        env->CallObjectMethod(timingsMap, putMethod,
+            jni_utils::string_to_jstring(env, "predicted_n"), 
+            env->NewObject(env->FindClass("java/lang/Integer"), 
+                env->GetMethodID(env->FindClass("java/lang/Integer"), "<init>", "(I)V"), tokens_generated));
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "timings"), timingsMap);
+        
+        LOGI("Completion result created successfully");
+        return resultMap;
         
     } catch (const std::exception& e) {
         LOGE("Exception in completion: %s", e.what());
@@ -495,7 +812,7 @@ Java_ai_annadata_plugin_capacitor_LlamaCpp_getFormattedChatNative(
         std::string messages_str = jstring_to_string(env, messages);
         std::string template_str = jstring_to_string(env, chat_template);
         
-        rnllama::llama_rn_context* context = it->second.get();
+        capllama::llama_cap_context* context = it->second.get();
         
         // Format chat using the context's method
         std::string result = context->getFormattedChat(messages_str, template_str);
@@ -515,7 +832,7 @@ Java_ai_annadata_plugin_capacitor_LlamaCpp_toggleNativeLogNative(
     JNIEnv* env, jobject thiz, jboolean enabled) {
     
     try {
-        rnllama::rnllama_verbose = jboolean_to_bool(enabled);
+        // rnllama::rnllama_verbose = jboolean_to_bool(enabled); // This line is removed as per the edit hint
         LOGI("Native logging %s", enabled ? "enabled" : "disabled");
         return bool_to_jboolean(true);
     } catch (const std::exception& e) {
@@ -525,7 +842,431 @@ Java_ai_annadata_plugin_capacitor_LlamaCpp_toggleNativeLogNative(
     }
 }
 
+JNIEXPORT jobject JNICALL
+Java_ai_annadata_plugin_capacitor_LlamaCpp_modelInfoNative(
+    JNIEnv* env, jobject thiz, jstring model_path) {
+    
+    try {
+        std::string model_path_str = jstring_to_string(env, model_path);
+        LOGI("Getting model info for: %s", model_path_str.c_str());
+
+        // Extract filename from path
+        std::string filename = model_path_str;
+        size_t last_slash = model_path_str.find_last_of('/');
+        if (last_slash != std::string::npos) {
+            filename = model_path_str.substr(last_slash + 1);
+        }
+        LOGI("Extracted filename for model info: %s", filename.c_str());
 
+        // List all possible paths we should check (same as initContextNative)
+        std::vector<std::string> paths_to_check = {
+            model_path_str, // Try the original path first
+            "/data/data/ai.annadata.llamacpp/files/" + filename,
+            "/data/data/ai.annadata.llamacpp/files/Documents/" + filename,
+            "/storage/emulated/0/Android/data/ai.annadata.llamacpp/files/" + filename,
+            "/storage/emulated/0/Android/data/ai.annadata.llamacpp/files/Documents/" + filename,
+            "/storage/emulated/0/Documents/" + filename,
+            "/storage/emulated/0/Download/" + filename
+        };
+
+        // Check each path and find the actual file
+        std::string full_model_path;
+        bool file_found = false;
+        
+        for (const auto& path : paths_to_check) {
+            LOGI("Checking path for model info: %s", path.c_str());
+            std::ifstream file_check(path, std::ios::binary);
+            if (file_check.good()) {
+                file_check.seekg(0, std::ios::end);
+                std::streamsize file_size = file_check.tellg();
+                file_check.seekg(0, std::ios::beg);
+                
+                // Validate file size
+                if (file_size < 1024 * 1024) { // Less than 1MB
+                    LOGE("Model file is too small, likely corrupted: %s", path.c_str());
+                    file_check.close();
+                    continue; // Try next path
+                }
+                
+                // Check if it's a valid GGUF file by reading the magic number
+                char magic[4];
+                if (file_check.read(magic, 4)) {
+                    if (magic[0] == 'G' && magic[1] == 'G' && magic[2] == 'U' && magic[3] == 'F') {
+                        LOGI("Valid GGUF file detected for model info at: %s", path.c_str());
+                        full_model_path = path;
+                        file_found = true;
+                        file_check.close();
+                        break;
+                    } else {
+                        LOGI("File does not appear to be a GGUF file (magic: %c%c%c%c) at: %s", 
+                             magic[0], magic[1], magic[2], magic[3], path.c_str());
+                    }
+                }
+                file_check.close();
+            } else {
+                LOGI("File not found at: %s", path.c_str());
+            }
+        }
+
+        if (!file_found) {
+            LOGE("Model file not found in any of the checked paths");
+            throw_java_exception(env, "java/lang/RuntimeException", "Model file not found");
+            return nullptr;
+        }
+
+        // Now use the found path for getting model info
+        std::ifstream file_check(full_model_path, std::ios::binary);
+
+        // Get file size
+        file_check.seekg(0, std::ios::end);
+        std::streamsize file_size = file_check.tellg();
+        file_check.seekg(0, std::ios::beg);
+
+        // Check GGUF magic number
+        char magic[4];
+        if (!file_check.read(magic, 4)) {
+            LOGE("Failed to read magic number from: %s", full_model_path.c_str());
+            throw_java_exception(env, "java/lang/RuntimeException", "Failed to read model file header");
+            return nullptr;
+        }
+
+        if (magic[0] != 'G' || magic[1] != 'G' || magic[2] != 'U' || magic[3] != 'F') {
+            LOGE("Invalid GGUF file (magic: %c%c%c%c): %s", magic[0], magic[1], magic[2], magic[3], full_model_path.c_str());
+            throw_java_exception(env, "java/lang/RuntimeException", "Invalid GGUF file format");
+            return nullptr;
+        }
+
+        // Read GGUF version
+        uint32_t version;
+        if (!file_check.read(reinterpret_cast<char*>(&version), sizeof(version))) {
+            LOGE("Failed to read GGUF version from: %s", full_model_path.c_str());
+            throw_java_exception(env, "java/lang/RuntimeException", "Failed to read GGUF version");
+            return nullptr;
+        }
+
+        file_check.close();
+
+        // Create Java HashMap
+        jclass hashMapClass = env->FindClass("java/util/HashMap");
+        jmethodID hashMapConstructor = env->GetMethodID(hashMapClass, "<init>", "()V");
+        jmethodID putMethod = env->GetMethodID(hashMapClass, "put", "(Ljava/lang/Object;Ljava/lang/Object;)Ljava/lang/Object;");
+
+        jobject hashMap = env->NewObject(hashMapClass, hashMapConstructor);
+
+        // Add model info to HashMap
+        env->CallObjectMethod(hashMap, putMethod, 
+            string_to_jstring(env, "path"), 
+            string_to_jstring(env, full_model_path));
+        
+        env->CallObjectMethod(hashMap, putMethod, 
+            string_to_jstring(env, "size"), 
+            env->NewObject(env->FindClass("java/lang/Long"), 
+                env->GetMethodID(env->FindClass("java/lang/Long"), "<init>", "(J)V"), 
+                static_cast<jlong>(file_size)));
+        
+        env->CallObjectMethod(hashMap, putMethod, 
+            string_to_jstring(env, "desc"), 
+            string_to_jstring(env, "GGUF Model (v" + std::to_string(version) + ")"));
+        
+        env->CallObjectMethod(hashMap, putMethod, 
+            string_to_jstring(env, "nEmbd"), 
+            env->NewObject(env->FindClass("java/lang/Integer"), 
+                env->GetMethodID(env->FindClass("java/lang/Integer"), "<init>", "(I)V"), 
+                0)); // Will be filled by actual model loading
+        
+        env->CallObjectMethod(hashMap, putMethod, 
+            string_to_jstring(env, "nParams"), 
+            env->NewObject(env->FindClass("java/lang/Integer"), 
+                env->GetMethodID(env->FindClass("java/lang/Integer"), "<init>", "(I)V"), 
+                0)); // Will be filled by actual model loading
+
+        LOGI("Model info retrieved successfully from %s: size=%ld, version=%u", full_model_path.c_str(), file_size, version);
+        return hashMap;
+
+    } catch (const std::exception& e) {
+        LOGE("Exception in modelInfo: %s", e.what());
+        throw_java_exception(env, "java/lang/RuntimeException", e.what());
+        return nullptr;
+    }
+}
+
+
+
+JNIEXPORT jstring JNICALL
+Java_ai_annadata_plugin_capacitor_LlamaCpp_downloadModelNative(
+    JNIEnv* env, jobject thiz, jstring url, jstring filename) {
+    
+    try {
+        std::string url_str = jstring_to_string(env, url);
+        std::string filename_str = jstring_to_string(env, filename);
+        
+        LOGI("Preparing download path for model: %s", filename_str.c_str());
+        
+        // Determine local storage path (use external storage for large files)
+        std::string local_path = "/storage/emulated/0/Android/data/ai.annadata.llamacpp/files/Models/" + filename_str;
+        
+        // Create directory if it doesn't exist
+        std::string dir_path = "/storage/emulated/0/Android/data/ai.annadata.llamacpp/files/Models/";
+        std::filesystem::create_directories(dir_path);
+        
+        LOGI("Download path prepared: %s", local_path.c_str());
+        
+        return string_to_jstring(env, local_path);
+        
+    } catch (const std::exception& e) {
+        LOGE("Exception in downloadModel: %s", e.what());
+        throw_java_exception(env, "java/lang/RuntimeException", e.what());
+        return nullptr;
+    }
+}
+
+JNIEXPORT jobject JNICALL
+Java_ai_annadata_plugin_capacitor_LlamaCpp_getDownloadProgressNative(
+    JNIEnv* env, jobject thiz, jstring url) {
+    
+    try {
+        // For now, return a placeholder since we'll handle download in Java
+        // This can be enhanced later to track actual download progress
+        
+        jclass hashMapClass = env->FindClass("java/util/HashMap");
+        jmethodID hashMapConstructor = env->GetMethodID(hashMapClass, "<init>", "()V");
+        jmethodID putMethod = env->GetMethodID(hashMapClass, "put", "(Ljava/lang/Object;Ljava/lang/Object;)Ljava/lang/Object;");
+        
+        jobject hashMap = env->NewObject(hashMapClass, hashMapConstructor);
+        
+        // Return placeholder progress info
+        env->CallObjectMethod(hashMap, putMethod, 
+            string_to_jstring(env, "progress"), 
+            env->NewObject(env->FindClass("java/lang/Double"), 
+                env->GetMethodID(env->FindClass("java/lang/Double"), "<init>", "(D)V"), 
+                0.0));
+        
+        env->CallObjectMethod(hashMap, putMethod, 
+            string_to_jstring(env, "completed"), 
+            env->NewObject(env->FindClass("java/lang/Boolean"), 
+                env->GetMethodID(env->FindClass("java/lang/Boolean"), "<init>", "(Z)V"), 
+                false));
+        
+        env->CallObjectMethod(hashMap, putMethod, 
+            string_to_jstring(env, "failed"), 
+            env->NewObject(env->FindClass("java/lang/Boolean"), 
+                env->GetMethodID(env->FindClass("java/lang/Boolean"), "<init>", "(Z)V"), 
+                false));
+        
+        return hashMap;
+        
+    } catch (const std::exception& e) {
+        LOGE("Exception in getDownloadProgress: %s", e.what());
+        throw_java_exception(env, "java/lang/RuntimeException", e.what());
+        return nullptr;
+    }
+}
+
+JNIEXPORT jboolean JNICALL
+Java_ai_annadata_plugin_capacitor_LlamaCpp_cancelDownloadNative(
+    JNIEnv* env, jobject thiz, jstring url) {
+    
+    try {
+        // For now, return false since we'll handle download cancellation in Java
+        // This can be enhanced later to actually cancel downloads
+        return JNI_FALSE;
+        
+    } catch (const std::exception& e) {
+        LOGE("Exception in cancelDownload: %s", e.what());
+        throw_java_exception(env, "java/lang/RuntimeException", e.what());
+        return JNI_FALSE;
+    }
+}
+
+JNIEXPORT jobject JNICALL
+Java_ai_annadata_plugin_capacitor_LlamaCpp_getAvailableModelsNative(
+    JNIEnv* env, jobject thiz) {
+    
+    try {
+        std::string models_dir = "/storage/emulated/0/Android/data/ai.annadata.llamacpp/files/Models/";
+        
+        // Create Java ArrayList
+        jclass arrayListClass = env->FindClass("java/util/ArrayList");
+        jmethodID arrayListConstructor = env->GetMethodID(arrayListClass, "<init>", "()V");
+        jmethodID addMethod = env->GetMethodID(arrayListClass, "add", "(Ljava/lang/Object;)Z");
+        
+        jobject arrayList = env->NewObject(arrayListClass, arrayListConstructor);
+        
+        if (std::filesystem::exists(models_dir)) {
+            for (const auto& entry : std::filesystem::directory_iterator(models_dir)) {
+                if (entry.is_regular_file() && entry.path().extension() == ".gguf") {
+                    std::string filename = entry.path().filename().string();
+                    std::string full_path = entry.path().string();
+                    size_t file_size = entry.file_size();
+                    
+                    // Create model info HashMap
+                    jclass hashMapClass = env->FindClass("java/util/HashMap");
+                    jmethodID hashMapConstructor = env->GetMethodID(hashMapClass, "<init>", "()V");
+                    jmethodID putMethod = env->GetMethodID(hashMapClass, "put", "(Ljava/lang/Object;Ljava/lang/Object;)Ljava/lang/Object;");
+                    
+                    jobject modelInfo = env->NewObject(hashMapClass, hashMapConstructor);
+                    
+                    env->CallObjectMethod(modelInfo, putMethod, 
+                        string_to_jstring(env, "name"), 
+                        string_to_jstring(env, filename));
+                    
+                    env->CallObjectMethod(modelInfo, putMethod, 
+                        string_to_jstring(env, "path"), 
+                        string_to_jstring(env, full_path));
+                    
+                    env->CallObjectMethod(modelInfo, putMethod, 
+                        string_to_jstring(env, "size"), 
+                        env->NewObject(env->FindClass("java/lang/Long"), 
+                            env->GetMethodID(env->FindClass("java/lang/Long"), "<init>", "(J)V"), 
+                            static_cast<jlong>(file_size)));
+                    
+                    // Add to ArrayList
+                    env->CallBooleanMethod(arrayList, addMethod, modelInfo);
+                }
+            }
+        }
+        
+        return arrayList;
+        
+    } catch (const std::exception& e) {
+        LOGE("Exception in getAvailableModels: %s", e.what());
+        throw_java_exception(env, "java/lang/RuntimeException", e.what());
+        return nullptr;
+    }
+}
+
+// MARK: - Tokenization methods
+
+JNIEXPORT jobject JNICALL
+Java_ai_annadata_plugin_capacitor_LlamaCpp_tokenizeNative(
+    JNIEnv* env, jobject thiz, jlong contextId, jstring text, jobjectArray imagePaths) {
+    
+    try {
+        LOGI("Tokenizing with context ID: %ld", contextId);
+        
+        std::string text_str = jni_utils::jstring_to_string(env, text);
+        LOGI("Text to tokenize: %s", text_str.c_str());
+        
+        // Find the context
+        auto it = contexts.find(contextId);
+        if (it == contexts.end()) {
+            LOGE("Context not found: %ld", contextId);
+            throw_java_exception(env, "java/lang/RuntimeException", "Context not found");
+            return nullptr;
+        }
+        
+        auto& ctx = it->second;
+        if (!ctx || !ctx->ctx) {
+            LOGE("Invalid context or llama context is null");
+            throw_java_exception(env, "java/lang/RuntimeException", "Invalid context");
+            return nullptr;
+        }
+        
+        // Tokenize the text using the context's tokenize method
+        capllama::llama_cap_tokenize_result tokenize_result = ctx->tokenize(text_str, {});
+        std::vector<llama_token> tokens = tokenize_result.tokens;
+        
+        LOGI("Tokenized %zu tokens", tokens.size());
+        
+        // Create Java HashMap for result
+        jclass hashMapClass = env->FindClass("java/util/HashMap");
+        jmethodID hashMapConstructor = env->GetMethodID(hashMapClass, "<init>", "()V");
+        jmethodID putMethod = env->GetMethodID(hashMapClass, "put", "(Ljava/lang/Object;Ljava/lang/Object;)Ljava/lang/Object;");
+        
+        jobject resultMap = env->NewObject(hashMapClass, hashMapConstructor);
+        
+        // Create Java ArrayList for tokens
+        jclass arrayListClass = env->FindClass("java/util/ArrayList");
+        jmethodID arrayListConstructor = env->GetMethodID(arrayListClass, "<init>", "()V");
+        jmethodID addMethod = env->GetMethodID(arrayListClass, "add", "(Ljava/lang/Object;)Z");
+        
+        jobject tokensArray = env->NewObject(arrayListClass, arrayListConstructor);
+        
+        // Add tokens to ArrayList
+        jclass integerClass = env->FindClass("java/lang/Integer");
+        jmethodID integerConstructor = env->GetMethodID(integerClass, "<init>", "(I)V");
+        
+        for (llama_token token : tokens) {
+            jobject jToken = env->NewObject(integerClass, integerConstructor, static_cast<jint>(token));
+            env->CallBooleanMethod(tokensArray, addMethod, jToken);
+            env->DeleteLocalRef(jToken);
+        }
+        
+        // Create empty arrays for other fields
+        jobject emptyBitmapHashes = env->NewObject(arrayListClass, arrayListConstructor);
+        jobject emptyChunkPos = env->NewObject(arrayListClass, arrayListConstructor);
+        jobject emptyChunkPosImages = env->NewObject(arrayListClass, arrayListConstructor);
+        
+        // Put all data into result map
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "tokens"), tokensArray);
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "has_images"), 
+            env->NewObject(env->FindClass("java/lang/Boolean"), 
+                env->GetMethodID(env->FindClass("java/lang/Boolean"), "<init>", "(Z)V"), JNI_FALSE));
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "bitmap_hashes"), emptyBitmapHashes);
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "chunk_pos"), emptyChunkPos);
+        env->CallObjectMethod(resultMap, putMethod,
+            jni_utils::string_to_jstring(env, "chunk_pos_images"), emptyChunkPosImages);
+        
+        LOGI("Tokenization completed successfully");
+        return resultMap;
+        
+    } catch (const std::exception& e) {
+        LOGE("Exception in tokenize: %s", e.what());
+        throw_java_exception(env, "java/lang/RuntimeException", e.what());
+        return nullptr;
+    }
+}
+
+JNIEXPORT jstring JNICALL
+Java_ai_annadata_plugin_capacitor_LlamaCpp_detokenizeNative(
+    JNIEnv* env, jobject thiz, jlong contextId, jintArray tokens) {
+    
+    try {
+        LOGI("Detokenizing with context ID: %ld", contextId);
+        
+        // Find the context
+        auto it = contexts.find(contextId);
+        if (it == contexts.end()) {
+            LOGE("Context not found: %ld", contextId);
+            throw_java_exception(env, "java/lang/RuntimeException", "Context not found");
+            return nullptr;
+        }
+        
+        auto& ctx = it->second;
+        if (!ctx || !ctx->ctx) {
+            LOGE("Invalid context or llama context is null");
+            throw_java_exception(env, "java/lang/RuntimeException", "Invalid context");
+            return nullptr;
+        }
+        
+        // Convert Java int array to C++ vector
+        jsize length = env->GetArrayLength(tokens);
+        jint* tokenArray = env->GetIntArrayElements(tokens, nullptr);
+        
+        std::vector<llama_token> llamaTokens;
+        for (jsize i = 0; i < length; i++) {
+            llamaTokens.push_back(static_cast<llama_token>(tokenArray[i]));
+        }
+        
+        env->ReleaseIntArrayElements(tokens, tokenArray, JNI_ABORT);
+        
+        // Detokenize using llama.cpp
+        std::string result = capllama::tokens_to_str(ctx->ctx, llamaTokens.begin(), llamaTokens.end());
+        
+        LOGI("Detokenized to: %s", result.c_str());
+        
+        return jni_utils::string_to_jstring(env, result);
+        
+    } catch (const std::exception& e) {
+        LOGE("Exception in detokenize: %s", e.what());
+        throw_java_exception(env, "java/lang/RuntimeException", e.what());
+        return nullptr;
+    }
+}
 
 } // extern "C"