react-native-ai-core 0.1.0 → 0.2.0

package/android/src/main/java/com/aicore/AiCoreModule.kt

@@ -1,5 +1,6 @@
 package com.aicore
 
+import android.content.Intent
 import android.os.Build
 import com.facebook.react.bridge.Arguments
 import com.facebook.react.bridge.Promise
@@ -11,6 +12,7 @@ import com.google.mediapipe.tasks.genai.llminference.LlmInferenceSession
 import com.google.mediapipe.tasks.genai.llminference.ProgressListener
 import com.google.mlkit.genai.common.DownloadStatus
 import com.google.mlkit.genai.common.FeatureStatus
+import com.google.mlkit.genai.common.GenAiException
 import com.google.mlkit.genai.prompt.Generation
 import com.google.mlkit.genai.prompt.GenerativeModel
 import com.google.mlkit.genai.prompt.TextPart
@@ -19,9 +21,11 @@ import kotlinx.coroutines.CoroutineScope
 import kotlinx.coroutines.Dispatchers
 import kotlinx.coroutines.SupervisorJob
 import kotlinx.coroutines.cancel
+import kotlinx.coroutines.delay
 import kotlinx.coroutines.flow.catch
 import kotlinx.coroutines.launch
 import java.io.File
+import java.util.concurrent.CountDownLatch
 import java.util.concurrent.ExecutorService
 import java.util.concurrent.Executors
 
@@ -34,7 +38,6 @@ class AiCoreModule(reactContext: ReactApplicationContext) :
   private val executor: ExecutorService = Executors.newSingleThreadExecutor()
   private val coroutineScope = CoroutineScope(Dispatchers.IO + SupervisorJob())
 
-  // Historial de conversación: lista de pares (mensaje usuario, respuesta asistente)
   private val conversationHistory = mutableListOf<Pair<String, String>>()
 
   companion object {
@@ -43,34 +46,38 @@ class AiCoreModule(reactContext: ReactApplicationContext) :
     const val EVENT_STREAM_COMPLETE = "AICore_streamComplete"
     const val EVENT_STREAM_ERROR    = "AICore_streamError"
     private const val DEFAULT_TEMPERATURE      = 0.7f
-    private const val DEFAULT_MAX_TOKENS        = 2048  // ventana MediaPipe (entrada+salida)
-    private const val DEFAULT_MAX_OUTPUT_TOKENS = 256   // salida ML Kit/AICore (límite API: 1–256)
+    private const val DEFAULT_MAX_TOKENS        = 4096  // MediaPipe context window (input+output)
+    private const val REQUESTED_MAX_OUTPUT_TOKENS = 256
+    private const val FALLBACK_MAX_OUTPUT_TOKENS = 256
     private const val DEFAULT_TOP_K             = 40
-    // Límite de caracteres del historial para no superar ~3000 tokens de entrada
+    private const val PROMPT_CHAR_BUDGET        = 4000
     private const val HISTORY_MAX_CHARS         = 9000
+    private const val MAX_CONTINUATIONS         = 12
+    private const val MAX_STREAM_IDLE_RETRIES   = 3
+    private const val QUOTA_ERROR_CODE          = 9  // AICore NPU quota exceeded error code
+    private const val CONTINUATION_DELAY_MS     = 1200L
+    private const val QUOTA_RETRY_DELAY_MS      = 1800L
+    private const val MAX_QUOTA_RETRIES         = 2
+    private const val MAX_NON_STREAM_QUOTA_RETRIES = 6
+    private const val BACKGROUND_RETRY_DELAY_MS  = 2000L
+    private const val MAX_BACKGROUND_RETRIES     = 1
+    private const val CONTINUATION_PROMPT       = "Continue exactly from the last generated character. Do not repeat, restart, summarize, explain, or add headings. Output only the direct continuation."
+    private const val END_MARKER                = "-E-"
+    private const val END_MARKER_INSTRUCTION    = "[INTERNAL] Append $END_MARKER only once at the true end of the final answer. Never use it before the end."
     private val STANDARD_MODEL_PATHS = listOf(
       "/data/local/tmp/gemini-nano.bin",
       "/sdcard/Download/gemini-nano.bin"
     )
   }
 
-  // ── Historial ──────────────────────────────────────────────────────────────
-
   @Synchronized
   private fun buildContextualPrompt(userPrompt: String): String {
-    if (conversationHistory.isEmpty()) return userPrompt
-    val sb = StringBuilder()
-    for ((u, a) in conversationHistory) {
-      sb.append("User: ").append(u).append("\nAssistant: ").append(a).append("\n")
-    }
-    sb.append("User: ").append(userPrompt).append("\nAssistant:")
-    return sb.toString()
+    return buildPromptWithBudget(userPrompt, null, END_MARKER_INSTRUCTION)
   }
 
   @Synchronized
   private fun saveToHistory(userPrompt: String, assistantResponse: String) {
     conversationHistory.add(Pair(userPrompt, assistantResponse))
-    // Eliminar turnos más antiguos si el historial supera el límite de caracteres
     var total = conversationHistory.sumOf { it.first.length + it.second.length }
     while (total > HISTORY_MAX_CHARS && conversationHistory.size > 1) {
       val removed = conversationHistory.removeAt(0)
@@ -83,12 +90,250 @@ class AiCoreModule(reactContext: ReactApplicationContext) :
     conversationHistory.clear()
   }
 
+  private fun trimFromStart(text: String, maxChars: Int): String {
+    if (text.length <= maxChars) return text
+    return text.takeLast(maxChars)
+  }
+
+  private fun trimFromEnd(text: String, maxChars: Int): String {
+    if (text.length <= maxChars) return text
+    return text.take(maxChars)
+  }
+
+  @Synchronized
+  private fun buildPromptWithBudget(
+    userPrompt: String,
+    assistantPrefix: String?,
+    hiddenUserPrompt: String?
+  ): String {
+    val hiddenInstruction = hiddenUserPrompt?.let { "\n$it\nAssistant:" } ?: ""
+    val assistantBase = "\nAssistant:"
+    val normalizedUserPrompt = trimFromEnd(userPrompt, PROMPT_CHAR_BUDGET)
+    val historySnapshot = conversationHistory.toMutableList()
+
+    while (true) {
+      val sb = StringBuilder()
+      for ((u, a) in historySnapshot) {
+        sb.append("User: ").append(u).append("\nAssistant: ").append(a).append("\n")
+      }
+      sb.append("User: ").append(normalizedUserPrompt).append(assistantBase)
+      if (assistantPrefix != null) {
+        sb.append(' ').append(assistantPrefix)
+      }
+      sb.append(hiddenInstruction)
+
+      val candidate = sb.toString()
+      if (candidate.length < PROMPT_CHAR_BUDGET || historySnapshot.isEmpty()) {
+        if (candidate.length <= PROMPT_CHAR_BUDGET) return candidate
+
+        val fixedPrefix = "User: $normalizedUserPrompt$assistantBase"
+        val availableForAssistant = (PROMPT_CHAR_BUDGET - fixedPrefix.length - hiddenInstruction.length - 1)
+          .coerceAtLeast(0)
+        val trimmedAssistantPrefix = assistantPrefix?.let { trimFromStart(it, availableForAssistant) }
+
+        return buildString {
+          append(fixedPrefix)
+          if (trimmedAssistantPrefix != null && trimmedAssistantPrefix.isNotEmpty()) {
+            append(' ').append(trimmedAssistantPrefix)
+          }
+          append(hiddenInstruction)
+        }
+      }
+
+      historySnapshot.removeAt(0)
+    }
+  }
+
+  private fun shouldContinueResponse(text: String): Boolean {
+    if (text.isBlank()) return false
+    val trimmed = text.trimEnd()
+    return !(trimmed.endsWith('.') || trimmed.endsWith('!') || trimmed.endsWith('?'))
+  }
+
+  private fun buildContinuationPrompt(originalUserPrompt: String, partialResponse: String): String {
+    return buildPromptWithBudget(
+      originalUserPrompt,
+      partialResponse,
+      "$CONTINUATION_PROMPT\n$END_MARKER_INSTRUCTION"
+    )
+  }
+
+  private fun containsEndMarker(text: String): Boolean {
+    return text.contains(END_MARKER)
+  }
+
+  private fun stripEndMarker(text: String): String {
+    return text.replace(END_MARKER, "")
+  }
+
+  private fun sanitizeVisibleText(text: String): String {
+    var cleaned = text
+    cleaned = cleaned.replace(Regex("(?i)\\[\\s*internal\\s*\\][^\\n\\r]*"), "")
+    cleaned = cleaned.replace(Regex("(?i)append\\s+\\Q$END_MARKER\\E[^\\n\\r]*"), "")
+    cleaned = cleaned.replace(Regex("(?i)never use it before the end\\.?"), "")
+    return cleaned
+  }
+
+  private fun adjustChunkBoundary(existing: String, incoming: String): String {
+    if (incoming.isEmpty() || existing.isEmpty()) return incoming
+
+    var chunk = incoming
+    val last = existing.last()
+    val first = chunk.first()
+
+    if (last.isLetterOrDigit() && first.isLetterOrDigit()) {
+      chunk = " $chunk"
+    }
+
+    if (existing.last() == ' ' && chunk.firstOrNull() == ' ') {
+      chunk = chunk.trimStart(' ')
+    }
+
+    return chunk
+  }
+
+  private fun emitStreamToken(token: String, done: Boolean) {
+    sendEvent(EVENT_STREAM_TOKEN, Arguments.createMap().apply {
+      putString("token", token)
+      putBoolean("done", done)
+    })
+  }
+
+  private fun isOutOfRangeError(error: Throwable): Boolean {
+    return error.message?.contains("out of range", ignoreCase = true) == true
+  }
+
+  private fun isQuotaError(error: Throwable): Boolean {
+    return error is GenAiException && error.errorCode == QUOTA_ERROR_CODE
+  }
+
+  private fun isBackgroundError(error: Throwable): Boolean {
+    val msg = error.message?.lowercase() ?: ""
+    return msg.contains("background usage is blocked") ||
+           msg.contains("use the api when your app is in the foreground")
+  }
+
+  private suspend fun generateMlKitChunk(model: GenerativeModel, prompt: String): String {
+    var quotaRetries = 0
+    var backgroundRetries = 0
+    while (true) {
+      try {
+        val request = generateContentRequest(TextPart(prompt)) {
+          maxOutputTokens = REQUESTED_MAX_OUTPUT_TOKENS
+        }
+        return model.generateContent(request).candidates.firstOrNull()?.text ?: ""
+      } catch (error: Exception) {
+        if (isQuotaError(error) && quotaRetries < MAX_QUOTA_RETRIES) {
+          quotaRetries++
+          delay(QUOTA_RETRY_DELAY_MS)
+          continue
+        }
+        if (isBackgroundError(error) && backgroundRetries < MAX_BACKGROUND_RETRIES) {
+          backgroundRetries++
+          delay(BACKGROUND_RETRY_DELAY_MS)
+          continue
+        }
+        if (!isOutOfRangeError(error)) throw error
+        while (true) {
+          try {
+            val fallbackRequest = generateContentRequest(TextPart(prompt)) {
+              maxOutputTokens = FALLBACK_MAX_OUTPUT_TOKENS
+            }
+            return model.generateContent(fallbackRequest).candidates.firstOrNull()?.text ?: ""
+          } catch (fallbackError: Exception) {
+            if (isQuotaError(fallbackError) && quotaRetries < MAX_QUOTA_RETRIES) {
+              quotaRetries++
+              delay(QUOTA_RETRY_DELAY_MS)
+              continue
+            }
+            throw fallbackError
+          }
+        }
+      }
+    }
+  }
+
+  private suspend fun streamMlKitChunk(
+    model: GenerativeModel,
+    prompt: String,
+    onToken: (String) -> Unit
+  ): Boolean {
+    suspend fun collectWithLimit(limit: Int): Boolean {
+      var quotaRetries = 0
+      var backgroundRetries = 0
+      while (true) {
+        var quotaHit = false
+        var backgroundHit = false
+        val request = generateContentRequest(TextPart(prompt)) {
+          maxOutputTokens = limit
+        }
+        try {
+          model.generateContentStream(request)
+            .catch { error ->
+              if (isQuotaError(error)) {
+                quotaHit = true
+              } else if (isBackgroundError(error)) {
+                backgroundHit = true
+              } else {
+                throw error
+              }
+            }
+            .collect { chunk ->
+              val token = chunk.candidates.firstOrNull()?.text ?: ""
+              onToken(token)
+            }
+        } catch (error: Exception) {
+          if (isQuotaError(error)) {
+            quotaHit = true
+          } else if (isBackgroundError(error)) {
+            backgroundHit = true
+          } else {
+            throw error
+          }
+        }
+        if (!quotaHit && !backgroundHit) return false
+        if (quotaHit) {
+          if (quotaRetries >= MAX_QUOTA_RETRIES) return true
+          quotaRetries++
+          delay(QUOTA_RETRY_DELAY_MS)
+        } else {
+          if (backgroundRetries >= MAX_BACKGROUND_RETRIES) return false
+          backgroundRetries++
+          delay(BACKGROUND_RETRY_DELAY_MS)
+        }
+      }
+    }
+
+    return try {
+      collectWithLimit(REQUESTED_MAX_OUTPUT_TOKENS)
+    } catch (error: Exception) {
+      if (!isOutOfRangeError(error)) throw error
+      collectWithLimit(FALLBACK_MAX_OUTPUT_TOKENS)
+    }
+  }
+
   private fun sendEvent(name: String, params: WritableMap?) {
     reactApplicationContext
       .getJSModule(DeviceEventManagerModule.RCTDeviceEventEmitter::class.java)
       .emit(name, params)
   }
 
+  private fun startInferenceService() {
+    val intent = Intent(reactApplicationContext, InferenceService::class.java)
+    if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.O) {
+      reactApplicationContext.startForegroundService(intent)
+    } else {
+      reactApplicationContext.startService(intent)
+    }
+  }
+
+  private fun stopInferenceService() {
+    val intent = Intent(reactApplicationContext, InferenceService::class.java).apply {
+      action = InferenceService.ACTION_STOP
+    }
+    reactApplicationContext.startService(intent)
+  }
+
   private fun createErrorMap(code: String, message: String): WritableMap =
     Arguments.createMap().apply {
       putString("code", code)
@@ -96,7 +341,7 @@ class AiCoreModule(reactContext: ReactApplicationContext) :
     }
 
   private fun createMediaPipeSession(): LlmInferenceSession {
-    val inference = llmInference ?: throw IllegalStateException("LLM no inicializado.")
+    val inference = llmInference ?: throw IllegalStateException("LLM not initialized.")
     val opts = LlmInferenceSession.LlmInferenceSessionOptions.builder()
       .setTemperature(DEFAULT_TEMPERATURE)
       .setTopK(DEFAULT_TOP_K)
@@ -104,6 +349,20 @@ class AiCoreModule(reactContext: ReactApplicationContext) :
     return LlmInferenceSession.createFromOptions(inference, opts)
   }
 
+  private fun buildPrompt(userPrompt: String, useConversationHistory: Boolean): String {
+    return if (useConversationHistory) buildContextualPrompt(userPrompt) else userPrompt
+  }
+
+  private fun maybeSaveToHistory(
+    userPrompt: String,
+    assistantResponse: String,
+    useConversationHistory: Boolean
+  ) {
+    if (useConversationHistory) {
+      saveToHistory(userPrompt, assistantResponse)
+    }
+  }
+
   override fun initialize(modelPath: String, promise: Promise) {
     mlkitModel = null
     llmInference?.close()
@@ -131,9 +390,9 @@ class AiCoreModule(reactContext: ReactApplicationContext) :
                 }
               }
             }
-            FeatureStatus.DOWNLOADING -> promise.reject("ALREADY_DOWNLOADING", "Ya se está descargando.")
-            FeatureStatus.UNAVAILABLE -> promise.reject("AICORE_UNAVAILABLE", "Gemini Nano no disponible en este dispositivo.")
-            else -> promise.reject("AICORE_UNKNOWN", "Estado desconocido.")
+            FeatureStatus.DOWNLOADING -> promise.reject("ALREADY_DOWNLOADING", "Model download already in progress.")
+            FeatureStatus.UNAVAILABLE -> promise.reject("AICORE_UNAVAILABLE", "Gemini Nano is not available on this device.")
+            else -> promise.reject("AICORE_UNKNOWN", "Unknown AICore status.")
           }
         } catch (e: Exception) {
           promise.reject("AICORE_ERROR", e.message, e)
@@ -143,7 +402,7 @@ class AiCoreModule(reactContext: ReactApplicationContext) :
       executor.execute {
         try {
           if (!File(modelPath).exists()) {
-            promise.reject("MODEL_NOT_FOUND", "Modelo no encontrado en: $modelPath")
+            promise.reject("MODEL_NOT_FOUND", "Model file not found at: $modelPath")
             return@execute
           }
           val options = LlmInference.LlmInferenceOptions.builder()
@@ -165,105 +424,287 @@ class AiCoreModule(reactContext: ReactApplicationContext) :
   }
 
   override fun generateResponse(prompt: String, promise: Promise) {
+    generateResponseInternal(prompt, true, promise)
+  }
+
+  override fun generateResponseStateless(prompt: String, promise: Promise) {
+    generateResponseInternal(prompt, false, promise)
+  }
+
+  private fun generateResponseInternal(
+    prompt: String,
+    useConversationHistory: Boolean,
+    promise: Promise
+  ) {
     val mlkit = mlkitModel
     val mediapipe = llmInference
-    val contextualPrompt = buildContextualPrompt(prompt)
+    startInferenceService()
     when {
       mlkit != null -> coroutineScope.launch {
         try {
-          val request = generateContentRequest(TextPart(contextualPrompt)) {
-            maxOutputTokens = DEFAULT_MAX_OUTPUT_TOKENS
+          val rawTotal = StringBuilder()
+          val visibleTotal = StringBuilder()
+          var currentPrompt = buildPrompt(prompt, useConversationHistory)
+          var continuations = 0
+          var continuationJoinPending = false
+          var quotaRetries = 0
+          while (true) {
+            val part = try {
+              generateMlKitChunk(mlkit, currentPrompt)
+            } catch (e: GenAiException) {
+              if (e.errorCode == QUOTA_ERROR_CODE) {
+                if (quotaRetries < MAX_NON_STREAM_QUOTA_RETRIES) {
+                  quotaRetries++
+                  delay(QUOTA_RETRY_DELAY_MS)
+                  continue
+                }
+                if (rawTotal.isNotEmpty()) break
+              }
+              throw e
+            }
+            quotaRetries = 0
+            rawTotal.append(part)
+            if (containsEndMarker(rawTotal.toString())) break
+
+            val cleanPart = sanitizeVisibleText(stripEndMarker(part))
+            val partForUi = if (continuationJoinPending) {
+              continuationJoinPending = false
+              adjustChunkBoundary(visibleTotal.toString(), cleanPart)
+            } else {
+              cleanPart
+            }
+            visibleTotal.append(partForUi)
+            val visible = visibleTotal.toString()
+            if (
+              useConversationHistory &&
+              shouldContinueResponse(visible) &&
+              continuations < MAX_CONTINUATIONS
+            ) {
+              currentPrompt = buildContinuationPrompt(prompt, visible)
+              continuations++
+              continuationJoinPending = true
+              delay(CONTINUATION_DELAY_MS)
+            } else break
+          }
+          val full = if (visibleTotal.isNotEmpty()) {
+            sanitizeVisibleText(visibleTotal.toString())
+          } else {
+            sanitizeVisibleText(stripEndMarker(rawTotal.toString()))
           }
-          val response = mlkit.generateContent(request).candidates.firstOrNull()?.text ?: ""
-          saveToHistory(prompt, response)
-          promise.resolve(response)
+          maybeSaveToHistory(prompt, full, useConversationHistory)
+          stopInferenceService()
+          promise.resolve(full)
         } catch (e: Exception) {
+          stopInferenceService()
           promise.reject("GENERATION_ERROR", e.message, e)
         }
       }
       mediapipe != null -> executor.execute {
         var session: LlmInferenceSession? = null
         try {
-          session = createMediaPipeSession()
-          session.addQueryChunk(contextualPrompt)
-          val response = session.generateResponse()
-          saveToHistory(prompt, response)
-          promise.resolve(response)
+          val rawTotal = StringBuilder()
+          val visibleTotal = StringBuilder()
+          var currentPrompt = buildPrompt(prompt, useConversationHistory)
+          var continuations = 0
+          var continuationJoinPending = false
+          while (true) {
+            session = createMediaPipeSession()
+            session.addQueryChunk(currentPrompt)
+            val part = session.generateResponse()
+            session.close()
+            session = null
+            rawTotal.append(part)
+            if (containsEndMarker(rawTotal.toString())) break
+
+            val cleanPart = sanitizeVisibleText(stripEndMarker(part))
+            val partForUi = if (continuationJoinPending) {
+              continuationJoinPending = false
+              adjustChunkBoundary(visibleTotal.toString(), cleanPart)
+            } else {
+              cleanPart
+            }
+            visibleTotal.append(partForUi)
+            val visible = visibleTotal.toString()
+            if (
+              useConversationHistory &&
+              shouldContinueResponse(visible) &&
+              continuations < MAX_CONTINUATIONS
+            ) {
+              currentPrompt = buildContinuationPrompt(prompt, visible)
+              continuations++
+              continuationJoinPending = true
+            } else break
+          }
+          val full = if (visibleTotal.isNotEmpty()) {
+            sanitizeVisibleText(visibleTotal.toString())
+          } else {
+            sanitizeVisibleText(stripEndMarker(rawTotal.toString()))
+          }
+          maybeSaveToHistory(prompt, full, useConversationHistory)
+          stopInferenceService()
+          promise.resolve(full)
         } catch (e: Exception) {
+          stopInferenceService()
           promise.reject("GENERATION_ERROR", e.message, e)
         } finally {
           session?.close()
         }
       }
-      else -> promise.reject("NOT_INITIALIZED", "LLM no inicializado.")
+      else -> promise.reject("NOT_INITIALIZED", "LLM not initialized.")
     }
   }
 
   override fun generateResponseStream(prompt: String) {
     val mlkit = mlkitModel
     val mediapipe = llmInference
-    val contextualPrompt = buildContextualPrompt(prompt)
+    startInferenceService()
     when {
       mlkit != null -> coroutineScope.launch {
+        val total = StringBuilder()
+        val rawTotal = StringBuilder()
+        var currentPrompt = buildContextualPrompt(prompt)
+        var continuations = 0
+        var continuationJoinPending = false
+        var firstDeltaInPass = false
+        var idleRetries = 0
         var streamError = false
-        val fullResponse = StringBuilder()
+        var markerReached = false
         try {
-          val request = generateContentRequest(TextPart(contextualPrompt)) {
-            maxOutputTokens = DEFAULT_MAX_OUTPUT_TOKENS
-          }
-          mlkit.generateContentStream(request)
-            .catch { e ->
+          while (true) {
+            val beforeLength = total.length
+            firstDeltaInPass = true
+            var quotaHit = false
+            try {
+              quotaHit = streamMlKitChunk(mlkit, currentPrompt) { token ->
+                rawTotal.append(token)
+                if (containsEndMarker(rawTotal.toString())) {
+                  markerReached = true
+                }
+
+                val visibleNow = sanitizeVisibleText(stripEndMarker(rawTotal.toString()))
+                if (visibleNow.length > total.length) {
+                  val delta = visibleNow.substring(total.length)
+                  val adjustedDelta = if (continuationJoinPending && firstDeltaInPass) {
+                    continuationJoinPending = false
+                    firstDeltaInPass = false
+                    adjustChunkBoundary(total.toString(), delta)
+                  } else {
+                    firstDeltaInPass = false
+                    delta
+                  }
+                  if (adjustedDelta.isNotEmpty()) {
+                    total.append(adjustedDelta)
+                    emitStreamToken(adjustedDelta, false)
+                  }
+                }
+              }
+            } catch (e: GenAiException) {
+              if (e.errorCode == QUOTA_ERROR_CODE && total.isNotEmpty()) {
+                quotaHit = true
+              } else {
+                throw e
+              }
+            } catch (e: Exception) {
               streamError = true
+              stopInferenceService()
               sendEvent(EVENT_STREAM_ERROR, createErrorMap("STREAM_ERROR", e.message ?: "Error"))
+              break
             }
-            .collect { chunk ->
-              val token = chunk.candidates.firstOrNull()?.text ?: ""
-              fullResponse.append(token)
-              sendEvent(EVENT_STREAM_TOKEN, Arguments.createMap().apply {
-                putString("token", token)
-                putBoolean("done", false)
-              })
+            if (streamError) return@launch
+            if (markerReached) break
+
+            val appendedNewText = total.length > beforeLength
+            if (!appendedNewText && shouldContinueResponse(total.toString()) && idleRetries < MAX_STREAM_IDLE_RETRIES) {
+              idleRetries++
+              currentPrompt = buildContinuationPrompt(prompt, total.toString())
+              continuationJoinPending = true
+              delay(if (quotaHit) QUOTA_RETRY_DELAY_MS else CONTINUATION_DELAY_MS)
+              continue
             }
-          if (!streamError) {
-            saveToHistory(prompt, fullResponse.toString())
-            sendEvent(EVENT_STREAM_TOKEN, Arguments.createMap().apply {
-              putString("token", "")
-              putBoolean("done", true)
-            })
-            sendEvent(EVENT_STREAM_COMPLETE, Arguments.createMap())
+
+            idleRetries = 0
+            if (shouldContinueResponse(total.toString()) && continuations < MAX_CONTINUATIONS) {
+              currentPrompt = buildContinuationPrompt(prompt, total.toString())
+              continuations++
+              continuationJoinPending = true
+              delay(if (quotaHit) QUOTA_RETRY_DELAY_MS else CONTINUATION_DELAY_MS)
+            } else break
           }
+          saveToHistory(prompt, sanitizeVisibleText(total.toString()))
+          emitStreamToken("", true)
+          stopInferenceService()
+          sendEvent(EVENT_STREAM_COMPLETE, Arguments.createMap())
         } catch (e: Exception) {
           if (!streamError) {
+            stopInferenceService()
             sendEvent(EVENT_STREAM_ERROR, createErrorMap("STREAM_ERROR", e.message ?: "Error"))
           }
         }
       }
       mediapipe != null -> executor.execute {
-        val fullResponse = StringBuilder()
+        val total = StringBuilder()
+        val rawTotal = StringBuilder()
+        var currentPrompt = buildContextualPrompt(prompt)
+        var continuations = 0
+        var continuationJoinPending = false
+        var firstDeltaInPass = false
         var session: LlmInferenceSession? = null
+        var markerReached = false
         try {
-          session = createMediaPipeSession()
-          val capturedSession = session
-          session.addQueryChunk(contextualPrompt)
-          session.generateResponseAsync(ProgressListener<String> { partial, done ->
-            val token = partial ?: ""
-            fullResponse.append(token)
-            sendEvent(EVENT_STREAM_TOKEN, Arguments.createMap().apply {
-              putString("token", token)
-              putBoolean("done", done)
+          while (true) {
+            val latch = CountDownLatch(1)
+            firstDeltaInPass = true
+            session = createMediaPipeSession()
+            val capturedSession = session
+            session.addQueryChunk(currentPrompt)
+            session.generateResponseAsync(ProgressListener<String> { partial, done ->
+              val token = partial ?: ""
+              rawTotal.append(token)
+              if (containsEndMarker(rawTotal.toString())) {
+                markerReached = true
+              }
+
+              val visibleNow = sanitizeVisibleText(stripEndMarker(rawTotal.toString()))
+              if (visibleNow.length > total.length) {
+                val delta = visibleNow.substring(total.length)
+                val adjustedDelta = if (continuationJoinPending && firstDeltaInPass) {
+                  continuationJoinPending = false
+                  firstDeltaInPass = false
+                  adjustChunkBoundary(total.toString(), delta)
+                } else {
+                  firstDeltaInPass = false
+                  delta
+                }
+                if (adjustedDelta.isNotEmpty()) {
+                  total.append(adjustedDelta)
+                  emitStreamToken(adjustedDelta, false)
+                }
+              }
+              if (done) {
+                capturedSession.close()
+                latch.countDown()
+              }
             })
-            if (done) {
-              saveToHistory(prompt, fullResponse.toString())
-              sendEvent(EVENT_STREAM_COMPLETE, Arguments.createMap())
-              capturedSession.close()
-            }
-          })
+            latch.await()
+            session = null
+            if (markerReached) break
+            if (shouldContinueResponse(total.toString()) && continuations < MAX_CONTINUATIONS) {
+              currentPrompt = buildContinuationPrompt(prompt, total.toString())
+              continuations++
+              continuationJoinPending = true
+            } else break
+          }
+          saveToHistory(prompt, sanitizeVisibleText(total.toString()))
+          emitStreamToken("", true)
+          stopInferenceService()
+          sendEvent(EVENT_STREAM_COMPLETE, Arguments.createMap())
         } catch (e: Exception) {
           session?.close()
+          stopInferenceService()
           sendEvent(EVENT_STREAM_ERROR, createErrorMap("STREAM_ERROR", e.message ?: "Error"))
         }
       }
-      else -> sendEvent(EVENT_STREAM_ERROR, createErrorMap("NOT_INITIALIZED", "LLM no inicializado."))
+      else -> sendEvent(EVENT_STREAM_ERROR, createErrorMap("NOT_INITIALIZED", "LLM not initialized."))
     }
   }
 
@@ -317,6 +758,7 @@ class AiCoreModule(reactContext: ReactApplicationContext) :
   override fun invalidate() {
     super.invalidate()
     try {
+      stopInferenceService()
       llmInference?.close()
       llmInference = null
       mlkitModel = null