feat(admin): presets d'échantillonnage LLM (Speaker/Thinker) — app-side complet

Système de presets de sampling réglables depuis l'admin, applicables au Speaker
et au Thinker. Inspiré du tuning Kaz (temp/top-p/top-k/pénalités).

Core/config :
- SamplingParams (core) : +presencePenalty, +frequencyPenalty
- ConfigStore.ModelConfig : +topP/topK/repeatPenalty/presence/frequency/maxTokens
  +presetName ; SamplingPreset + RuntimeConfig.presets (bibliothèque) ; JSON ;
  5 presets d'usine (Équilibré/Kaz chaleureux/Factuel/Analyse/Créatif).
  Défauts maxTokens : Speaker 450, Thinker 128 (préserve la longueur prod).
- ContentProvider : colonnes speaker_*/thinker_* sampling + presets_json (R/W).
- KazeiaService : samplingFrom(ModelConfig) → SamplingParams pour Speaker+Thinker
  (au lieu du hardcodé). maxTokens honoré par le moteur ; le reste prêt.

Admin :
- KazeiaConfigClient : Sampling + Preset + parse/serialize presets_json + updateSampling/updatePresets
- ConfigRepository : updateSampling/updatePresets
- ParametersScreen (nav "Paramètres") : éditeurs sliders Speaker/Thinker + apply preset
  + bibliothèque CRUD ; bandeau honnête « seul Tokens max agit aujourd'hui ».

⚠ Le moteur natif n'honore que maxTokens à ce jour. Spec d'extension JNI sampling
pour le dev engine : docs/SAMPLING_ENGINE_SPEC.md (~10 lignes app à brancher après).

Vérifié device : colonnes sampling R/W (temp/max/preset_name), presets_json seedés
+ lus. Build app+admin verts, admin lancé sans crash.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

kazeia-android/app-admin/src/main/java/com/kazeia/admin/MainActivity.kt

@@ -93,6 +93,9 @@ private fun MainShell() {
                 composable(AdminDestination.Prompts.route) {
                     PromptsScreen()
                 }
+                composable(AdminDestination.Parameters.route) {
+                    com.kazeia.admin.ui.parameters.ParametersScreen()
+                }
                 composable(AdminDestination.Rag.route) {
                     com.kazeia.admin.ui.rag.RagScreen()
                 }

kazeia-android/app-admin/src/main/java/com/kazeia/admin/data/repository/ConfigRepository.kt

@@ -25,6 +25,15 @@ class ConfigRepository(context: Context) {
     fun updateDebug(enabled: Boolean): Boolean =
         client.updateConfig(debugEnabled = enabled)
 
+    /** Applique un jeu de sampling à un rôle ("speaker"|"thinker"), avec le nom
+     *  du preset source ("" = réglages custom). */
+    fun updateSampling(role: String, s: KazeiaConfigClient.Sampling, presetName: String): Boolean =
+        client.updateSampling(role, s, presetName)
+
+    /** Remplace toute la bibliothèque de presets (création / édition / suppression). */
+    fun updatePresets(presets: List<KazeiaConfigClient.Preset>): Boolean =
+        client.updatePresets(presets)
+
     fun updateSpeakerModel(modelId: String): Boolean =
         client.updateConfig(speakerModelId = modelId)
 

kazeia-android/app-admin/src/main/java/com/kazeia/admin/data/source/KazeiaConfigClient.kt

@@ -2,7 +2,10 @@ package com.kazeia.admin.data.source
 
 import android.content.ContentResolver
 import android.content.ContentValues
+import android.database.Cursor
 import android.net.Uri
+import org.json.JSONArray
+import org.json.JSONObject
 
 /**
  * Client du ContentProvider Kazeia pour les URI `/config` et `/models`.
@@ -17,20 +20,41 @@ class KazeiaConfigClient(private val resolver: ContentResolver) {
         val MODELS_URI: Uri = Uri.parse("content://$AUTHORITY/models")
     }
 
+    /** Jeu de paramètres d'échantillonnage d'un modèle (ou d'un preset). */
+    data class Sampling(
+        val temperature: Float = 0.7f,
+        val topP: Float = 0.85f,
+        val topK: Int = 40,
+        val repeatPenalty: Float = 1.1f,
+        val presencePenalty: Float = 0.0f,
+        val frequencyPenalty: Float = 0.0f,
+        val maxTokens: Int = 120
+    )
+
+    /** Preset nommé réutilisable, applicable au Speaker ou au Thinker. */
+    data class Preset(val name: String, val sampling: Sampling)
+
     data class Config(
         val cascadeEnabled: Boolean,
         val ttsEnabled: Boolean,
         val speakerModelId: String,
         val speakerSystemPrompt: String,
-        val speakerTemperature: Float,
+        val speakerSampling: Sampling,
+        val speakerPresetName: String,
         val thinkerModelId: String,
         val thinkerSystemPrompt: String,
-        val thinkerTemperature: Float,
+        val thinkerSampling: Sampling,
+        val thinkerPresetName: String,
+        val presets: List<Preset> = emptyList(),
         val ragEnabled: Boolean = false,
         val ragThreshold: Float = 0.82f,
         val ragTopK: Int = 3,
         val debugEnabled: Boolean = false
-    )
+    ) {
+        // Compat : anciens écrans qui lisaient juste la température.
+        val speakerTemperature: Float get() = speakerSampling.temperature
+        val thinkerTemperature: Float get() = thinkerSampling.temperature
+    }
 
     data class ModelInfo(
         val id: String,
@@ -44,6 +68,23 @@ class KazeiaConfigClient(private val resolver: ContentResolver) {
         val notes: String
     )
 
+    private fun Cursor.floatOr(col: String, d: Float) =
+        getColumnIndex(col).let { if (it >= 0) getFloat(it) else d }
+    private fun Cursor.intOr(col: String, d: Int) =
+        getColumnIndex(col).let { if (it >= 0) getInt(it) else d }
+    private fun Cursor.strOr(col: String, d: String) =
+        getColumnIndex(col).let { if (it >= 0) getString(it) ?: d else d }
+
+    private fun Cursor.sampling(p: String) = Sampling(
+        temperature = floatOr("${p}temperature", 0.7f),
+        topP = floatOr("${p}top_p", 0.85f),
+        topK = intOr("${p}top_k", 40),
+        repeatPenalty = floatOr("${p}repeat_penalty", 1.1f),
+        presencePenalty = floatOr("${p}presence_penalty", 0.0f),
+        frequencyPenalty = floatOr("${p}frequency_penalty", 0.0f),
+        maxTokens = intOr("${p}max_tokens", 120)
+    )
+
     fun queryConfig(): Config? = runCatching {
         resolver.query(CONFIG_URI, null, null, null, null)?.use { c ->
             if (!c.moveToFirst()) return@use null
@@ -52,18 +93,58 @@ class KazeiaConfigClient(private val resolver: ContentResolver) {
                 ttsEnabled = c.getInt(c.getColumnIndexOrThrow("tts_enabled")) == 1,
                 speakerModelId = c.getString(c.getColumnIndexOrThrow("speaker_model_id")),
                 speakerSystemPrompt = c.getString(c.getColumnIndexOrThrow("speaker_system_prompt")),
-                speakerTemperature = c.getFloat(c.getColumnIndexOrThrow("speaker_temperature")),
+                speakerSampling = c.sampling("speaker_"),
+                speakerPresetName = c.strOr("speaker_preset_name", ""),
                 thinkerModelId = c.getString(c.getColumnIndexOrThrow("thinker_model_id")),
                 thinkerSystemPrompt = c.getString(c.getColumnIndexOrThrow("thinker_system_prompt")),
-                thinkerTemperature = c.getFloat(c.getColumnIndexOrThrow("thinker_temperature")),
-                ragEnabled = c.getColumnIndex("rag_enabled").let { if (it >= 0) c.getInt(it) == 1 else false },
-                ragThreshold = c.getColumnIndex("rag_threshold").let { if (it >= 0) c.getFloat(it) else 0.82f },
-                ragTopK = c.getColumnIndex("rag_top_k").let { if (it >= 0) c.getInt(it) else 3 },
-                debugEnabled = c.getColumnIndex("debug_enabled").let { if (it >= 0) c.getInt(it) == 1 else false }
+                thinkerSampling = c.sampling("thinker_"),
+                thinkerPresetName = c.strOr("thinker_preset_name", ""),
+                presets = parsePresets(c.strOr("presets_json", "[]")),
+                ragEnabled = c.intOr("rag_enabled", 0) == 1,
+                ragThreshold = c.floatOr("rag_threshold", 0.82f),
+                ragTopK = c.intOr("rag_top_k", 3),
+                debugEnabled = c.intOr("debug_enabled", 0) == 1
             )
         }
     }.getOrNull()
 
+    private fun parsePresets(json: String): List<Preset> = runCatching {
+        val arr = JSONArray(json)
+        (0 until arr.length()).mapNotNull { i ->
+            arr.optJSONObject(i)?.let { o ->
+                Preset(
+                    o.optString("name", ""),
+                    Sampling(
+                        o.optDouble("temperature", 0.7).toFloat(),
+                        o.optDouble("top_p", 0.85).toFloat(),
+                        o.optInt("top_k", 40),
+                        o.optDouble("repeat_penalty", 1.1).toFloat(),
+                        o.optDouble("presence_penalty", 0.0).toFloat(),
+                        o.optDouble("frequency_penalty", 0.0).toFloat(),
+                        o.optInt("max_tokens", 120)
+                    )
+                )
+            }
+        }.filter { it.name.isNotBlank() }
+    }.getOrDefault(emptyList())
+
+    private fun presetsToJson(presets: List<Preset>): String {
+        val arr = JSONArray()
+        presets.forEach { p ->
+            arr.put(JSONObject().apply {
+                put("name", p.name)
+                put("temperature", p.sampling.temperature)
+                put("top_p", p.sampling.topP)
+                put("top_k", p.sampling.topK)
+                put("repeat_penalty", p.sampling.repeatPenalty)
+                put("presence_penalty", p.sampling.presencePenalty)
+                put("frequency_penalty", p.sampling.frequencyPenalty)
+                put("max_tokens", p.sampling.maxTokens)
+            })
+        }
+        return arr.toString()
+    }
+
     fun queryModels(): List<ModelInfo> = runCatching {
         resolver.query(MODELS_URI, null, null, null, null)?.use { c ->
             val out = mutableListOf<ModelInfo>()
@@ -91,10 +172,8 @@ class KazeiaConfigClient(private val resolver: ContentResolver) {
         ttsEnabled: Boolean? = null,
         speakerModelId: String? = null,
         speakerSystemPrompt: String? = null,
-        speakerTemperature: Float? = null,
         thinkerModelId: String? = null,
         thinkerSystemPrompt: String? = null,
-        thinkerTemperature: Float? = null,
         ragEnabled: Boolean? = null,
         ragThreshold: Float? = null,
         ragTopK: Int? = null,
@@ -105,17 +184,37 @@ class KazeiaConfigClient(private val resolver: ContentResolver) {
         ttsEnabled?.let { v.put("tts_enabled", it) }
         speakerModelId?.let { v.put("speaker_model_id", it) }
         speakerSystemPrompt?.let { v.put("speaker_system_prompt", it) }
-        speakerTemperature?.let { v.put("speaker_temperature", it) }
         thinkerModelId?.let { v.put("thinker_model_id", it) }
         thinkerSystemPrompt?.let { v.put("thinker_system_prompt", it) }
-        thinkerTemperature?.let { v.put("thinker_temperature", it) }
         ragEnabled?.let { v.put("rag_enabled", it) }
         ragThreshold?.let { v.put("rag_threshold", it) }
         ragTopK?.let { v.put("rag_top_k", it) }
         debugEnabled?.let { v.put("debug_enabled", it) }
+        return push(v)
+    }
+
+    /** Applique un jeu de sampling (+ nom de preset, "" = custom) à un rôle. */
+    fun updateSampling(role: String, s: Sampling, presetName: String): Boolean {
+        val p = "${role}_"   // "speaker" | "thinker"
+        val v = ContentValues().apply {
+            put("${p}temperature", s.temperature)
+            put("${p}top_p", s.topP)
+            put("${p}top_k", s.topK)
+            put("${p}repeat_penalty", s.repeatPenalty)
+            put("${p}presence_penalty", s.presencePenalty)
+            put("${p}frequency_penalty", s.frequencyPenalty)
+            put("${p}max_tokens", s.maxTokens)
+            put("${p}preset_name", presetName)
+        }
+        return push(v)
+    }
+
+    /** Remplace toute la bibliothèque de presets. */
+    fun updatePresets(presets: List<Preset>): Boolean =
+        push(ContentValues().apply { put("presets_json", presetsToJson(presets)) })
+
+    private fun push(v: ContentValues): Boolean {
         if (v.size() == 0) return true
-        return runCatching {
-            resolver.update(CONFIG_URI, v, null, null) > 0
-        }.getOrDefault(false)
+        return runCatching { resolver.update(CONFIG_URI, v, null, null) > 0 }.getOrDefault(false)
     }
 }

kazeia-android/app-admin/src/main/java/com/kazeia/admin/ui/navigation/AdminDestination.kt

@@ -8,6 +8,7 @@ import androidx.compose.material.icons.outlined.Build
 import androidx.compose.material.icons.outlined.GraphicEq
 import androidx.compose.material.icons.outlined.History
 import androidx.compose.material.icons.outlined.People
+import androidx.compose.material.icons.outlined.Tune
 import androidx.compose.material.icons.outlined.RecordVoiceOver
 import androidx.compose.ui.graphics.vector.ImageVector
 
@@ -23,6 +24,7 @@ enum class AdminDestination(
     Voices(   "voices",    com.kazeia.admin.R.string.nav_voices,    Icons.Outlined.RecordVoiceOver),
     Profiles( "profiles",  com.kazeia.admin.R.string.nav_profiles,  Icons.Outlined.People),
     Prompts(  "prompts",   com.kazeia.admin.R.string.nav_prompts,   Icons.AutoMirrored.Outlined.Article),
+    Parameters("parameters", com.kazeia.admin.R.string.nav_parameters, Icons.Outlined.Tune),
     Rag(      "rag",       com.kazeia.admin.R.string.nav_rag,       Icons.Outlined.Book),
     History(  "history",   com.kazeia.admin.R.string.nav_history,   Icons.Outlined.History),
     Telemetry("telemetry", com.kazeia.admin.R.string.nav_telemetry, Icons.Outlined.Analytics),

kazeia-android/app-admin/src/main/java/com/kazeia/admin/ui/parameters/ParametersScreen.kt

@@ -0,0 +1,356 @@
+package com.kazeia.admin.ui.parameters
+
+import androidx.compose.foundation.layout.Arrangement
+import androidx.compose.foundation.layout.Column
+import androidx.compose.foundation.layout.Row
+import androidx.compose.foundation.layout.Spacer
+import androidx.compose.foundation.layout.fillMaxSize
+import androidx.compose.foundation.layout.fillMaxWidth
+import androidx.compose.foundation.layout.height
+import androidx.compose.foundation.layout.padding
+import androidx.compose.foundation.rememberScrollState
+import androidx.compose.foundation.shape.RoundedCornerShape
+import androidx.compose.foundation.verticalScroll
+import androidx.compose.material3.Button
+import androidx.compose.material3.Card
+import androidx.compose.material3.CardDefaults
+import androidx.compose.material3.DropdownMenuItem
+import androidx.compose.material3.ExperimentalMaterial3Api
+import androidx.compose.material3.ExposedDropdownMenuBox
+import androidx.compose.material3.ExposedDropdownMenuDefaults
+import androidx.compose.material3.MaterialTheme
+import androidx.compose.material3.OutlinedTextField
+import androidx.compose.material3.Scaffold
+import androidx.compose.material3.SnackbarDuration
+import androidx.compose.material3.SnackbarHost
+import androidx.compose.material3.SnackbarHostState
+import androidx.compose.material3.Slider
+import androidx.compose.material3.Text
+import androidx.compose.material3.TextButton
+import androidx.compose.material3.TopAppBar
+import androidx.compose.material3.TopAppBarDefaults
+import androidx.compose.runtime.Composable
+import androidx.compose.runtime.LaunchedEffect
+import androidx.compose.runtime.getValue
+import androidx.compose.runtime.mutableStateOf
+import androidx.compose.runtime.remember
+import androidx.compose.runtime.rememberCoroutineScope
+import androidx.compose.runtime.setValue
+import androidx.compose.ui.Alignment
+import androidx.compose.ui.Modifier
+import androidx.compose.ui.platform.LocalContext
+import androidx.compose.ui.text.font.FontFamily
+import androidx.compose.ui.text.font.FontWeight
+import androidx.compose.ui.unit.dp
+import com.kazeia.admin.data.repository.ConfigRepository
+import com.kazeia.admin.data.source.KazeiaConfigClient.Preset
+import com.kazeia.admin.data.source.KazeiaConfigClient.Sampling
+import kotlinx.coroutines.launch
+import kotlin.math.roundToInt
+
+@OptIn(ExperimentalMaterial3Api::class)
+@Composable
+fun ParametersScreen() {
+    val context = LocalContext.current
+    val repo = remember { ConfigRepository(context) }
+    val scope = rememberCoroutineScope()
+    val snackbar = remember { SnackbarHostState() }
+
+    var loaded by remember { mutableStateOf(false) }
+    var reachable by remember { mutableStateOf(true) }
+    var presets by remember { mutableStateOf<List<Preset>>(emptyList()) }
+    var speaker by remember { mutableStateOf(Sampling()) }
+    var thinker by remember { mutableStateOf(Sampling()) }
+    var speakerPreset by remember { mutableStateOf("") }
+    var thinkerPreset by remember { mutableStateOf("") }
+
+    fun reload() {
+        val cfg = repo.config()
+        if (cfg == null) { reachable = false; loaded = true; return }
+        reachable = true
+        presets = cfg.presets
+        speaker = cfg.speakerSampling; speakerPreset = cfg.speakerPresetName
+        thinker = cfg.thinkerSampling; thinkerPreset = cfg.thinkerPresetName
+        loaded = true
+    }
+    LaunchedEffect(Unit) { reload() }
+
+    Scaffold(
+        topBar = {
+            TopAppBar(
+                title = { Text("Paramètres d'échantillonnage") },
+                colors = TopAppBarDefaults.topAppBarColors(
+                    containerColor = MaterialTheme.colorScheme.background
+                )
+            )
+        },
+        snackbarHost = { SnackbarHost(snackbar) }
+    ) { padding ->
+        if (!loaded) return@Scaffold
+        if (!reachable) {
+            Column(Modifier.fillMaxSize().padding(32.dp), verticalArrangement = Arrangement.Center) {
+                Text("Kazeia patient non joignable.", style = MaterialTheme.typography.titleLarge)
+                Text("Lance l'app Kazeia puis reviens ici.", color = MaterialTheme.colorScheme.onSurfaceVariant)
+            }
+            return@Scaffold
+        }
+        Column(
+            modifier = Modifier
+                .fillMaxSize()
+                .padding(padding)
+                .verticalScroll(rememberScrollState())
+                .padding(horizontal = 16.dp),
+            verticalArrangement = Arrangement.spacedBy(16.dp)
+        ) {
+            Spacer(Modifier.height(8.dp))
+            EngineNoticeCard()
+
+            RoleCard(
+                title = "Speaker", subtitle = "LLM qui répond au patient",
+                sampling = speaker, presetName = speakerPreset, presets = presets,
+                onChange = { speaker = it; speakerPreset = "" },
+                onApplyPreset = { p -> speaker = p.sampling; speakerPreset = p.name },
+                onSave = {
+                    val ok = repo.updateSampling("speaker", speaker, speakerPreset)
+                    scope.launch { snackbar.showSnackbar(if (ok) "Speaker mis à jour." else "Échec : patient injoignable.", duration = SnackbarDuration.Short) }
+                }
+            )
+
+            RoleCard(
+                title = "Thinker", subtitle = "LLM d'analyse (cascade)",
+                sampling = thinker, presetName = thinkerPreset, presets = presets,
+                onChange = { thinker = it; thinkerPreset = "" },
+                onApplyPreset = { p -> thinker = p.sampling; thinkerPreset = p.name },
+                onSave = {
+                    val ok = repo.updateSampling("thinker", thinker, thinkerPreset)
+                    scope.launch { snackbar.showSnackbar(if (ok) "Thinker mis à jour." else "Échec : patient injoignable.", duration = SnackbarDuration.Short) }
+                }
+            )
+
+            PresetLibraryCard(
+                presets = presets,
+                onSave = { updated ->
+                    val ok = repo.updatePresets(updated)
+                    if (ok) reload()
+                    scope.launch { snackbar.showSnackbar(if (ok) "Bibliothèque de presets enregistrée." else "Échec : patient injoignable.", duration = SnackbarDuration.Short) }
+                }
+            )
+            Spacer(Modifier.height(16.dp))
+        }
+    }
+}
+
+@Composable
+private fun EngineNoticeCard() {
+    Card(
+        modifier = Modifier.fillMaxWidth(),
+        colors = CardDefaults.cardColors(containerColor = MaterialTheme.colorScheme.tertiaryContainer.copy(alpha = 0.5f)),
+        shape = RoundedCornerShape(12.dp)
+    ) {
+        Column(Modifier.padding(16.dp)) {
+            Text("ℹ️ Effet sur la génération", fontWeight = FontWeight.SemiBold)
+            Spacer(Modifier.height(4.dp))
+            Text(
+                "Aujourd'hui le moteur n'applique que « Tokens max ». Température, top-p/k et " +
+                    "pénalités sont enregistrés et prendront effet dès que le moteur expose " +
+                    "l'échantillonnage (mise à jour à venir).",
+                style = MaterialTheme.typography.bodyMedium,
+                color = MaterialTheme.colorScheme.onSurfaceVariant
+            )
+        }
+    }
+}
+
+@OptIn(ExperimentalMaterial3Api::class)
+@Composable
+private fun RoleCard(
+    title: String,
+    subtitle: String,
+    sampling: Sampling,
+    presetName: String,
+    presets: List<Preset>,
+    onChange: (Sampling) -> Unit,
+    onApplyPreset: (Preset) -> Unit,
+    onSave: () -> Unit
+) {
+    Card(
+        modifier = Modifier.fillMaxWidth(),
+        colors = CardDefaults.cardColors(containerColor = MaterialTheme.colorScheme.surface),
+        elevation = CardDefaults.cardElevation(defaultElevation = 0.dp),
+        shape = RoundedCornerShape(12.dp)
+    ) {
+        Column(Modifier.padding(16.dp)) {
+            Text(title, style = MaterialTheme.typography.titleMedium, fontWeight = FontWeight.SemiBold)
+            Text(subtitle, style = MaterialTheme.typography.bodyMedium, color = MaterialTheme.colorScheme.onSurfaceVariant)
+            Spacer(Modifier.height(12.dp))
+
+            // Appliquer un preset
+            var expanded by remember { mutableStateOf(false) }
+            ExposedDropdownMenuBox(expanded = expanded, onExpandedChange = { expanded = !expanded }) {
+                OutlinedTextField(
+                    value = if (presetName.isNotBlank()) presetName else "Personnalisé",
+                    onValueChange = {}, readOnly = true,
+                    label = { Text("Preset appliqué") },
+                    trailingIcon = { ExposedDropdownMenuDefaults.TrailingIcon(expanded = expanded) },
+                    modifier = Modifier
+                        .menuAnchor(androidx.compose.material3.MenuAnchorType.PrimaryNotEditable, true)
+                        .fillMaxWidth()
+                )
+                ExposedDropdownMenu(expanded = expanded, onDismissRequest = { expanded = false }) {
+                    if (presets.isEmpty()) {
+                        DropdownMenuItem(text = { Text("Aucun preset") }, onClick = { expanded = false }, enabled = false)
+                    }
+                    presets.forEach { p ->
+                        DropdownMenuItem(
+                            text = { Text(p.name) },
+                            onClick = { expanded = false; onApplyPreset(p) }
+                        )
+                    }
+                }
+            }
+
+            Spacer(Modifier.height(8.dp))
+            SamplingEditor(sampling, onChange)
+            Spacer(Modifier.height(8.dp))
+            Row {
+                Spacer(Modifier.weight(1f))
+                Button(onClick = onSave) { Text("Appliquer au $title") }
+            }
+        }
+    }
+}
+
+@Composable
+private fun SamplingEditor(s: Sampling, onChange: (Sampling) -> Unit) {
+    Column(verticalArrangement = Arrangement.spacedBy(2.dp)) {
+        SliderRow("Température", s.temperature, 0f, 2f) { onChange(s.copy(temperature = it)) }
+        SliderRow("Top-p", s.topP, 0f, 1f) { onChange(s.copy(topP = it)) }
+        IntSliderRow("Top-k", s.topK, 0, 100) { onChange(s.copy(topK = it)) }
+        SliderRow("Pénalité répétition", s.repeatPenalty, 1f, 2f) { onChange(s.copy(repeatPenalty = it)) }
+        SliderRow("Pénalité présence", s.presencePenalty, 0f, 2f) { onChange(s.copy(presencePenalty = it)) }
+        SliderRow("Pénalité fréquence", s.frequencyPenalty, 0f, 2f) { onChange(s.copy(frequencyPenalty = it)) }
+        IntSliderRow("Tokens max ✓", s.maxTokens, 16, 512) { onChange(s.copy(maxTokens = it)) }
+    }
+}
+
+@Composable
+private fun SliderRow(label: String, value: Float, min: Float, max: Float, onChange: (Float) -> Unit) {
+    Column {
+        Row {
+            Text(label, style = MaterialTheme.typography.bodyMedium, modifier = Modifier.weight(1f))
+            Text(
+                String.format("%.2f", value),
+                style = MaterialTheme.typography.bodyMedium,
+                fontFamily = FontFamily.Monospace
+            )
+        }
+        Slider(
+            value = value.coerceIn(min, max),
+            onValueChange = { onChange((it * 20).roundToInt() / 20f) },  // pas 0.05
+            valueRange = min..max
+        )
+    }
+}
+
+@Composable
+private fun IntSliderRow(label: String, value: Int, min: Int, max: Int, onChange: (Int) -> Unit) {
+    Column {
+        Row {
+            Text(label, style = MaterialTheme.typography.bodyMedium, modifier = Modifier.weight(1f))
+            Text("$value", style = MaterialTheme.typography.bodyMedium, fontFamily = FontFamily.Monospace)
+        }
+        Slider(
+            value = value.coerceIn(min, max).toFloat(),
+            onValueChange = { onChange(it.roundToInt()) },
+            valueRange = min.toFloat()..max.toFloat()
+        )
+    }
+}
+
+@Composable
+private fun PresetLibraryCard(presets: List<Preset>, onSave: (List<Preset>) -> Unit) {
+    Card(
+        modifier = Modifier.fillMaxWidth(),
+        colors = CardDefaults.cardColors(containerColor = MaterialTheme.colorScheme.surface),
+        elevation = CardDefaults.cardElevation(defaultElevation = 0.dp),
+        shape = RoundedCornerShape(12.dp)
+    ) {
+        Column(Modifier.padding(16.dp)) {
+            Text("Bibliothèque de presets", style = MaterialTheme.typography.titleMedium, fontWeight = FontWeight.SemiBold)
+            Text("Modèles réutilisables, applicables au Speaker ou au Thinker.",
+                style = MaterialTheme.typography.bodyMedium, color = MaterialTheme.colorScheme.onSurfaceVariant)
+            Spacer(Modifier.height(12.dp))
+
+            var editing by remember { mutableStateOf<Int?>(null) }  // index en édition
+
+            presets.forEachIndexed { i, p ->
+                Row(verticalAlignment = Alignment.CenterVertically, modifier = Modifier.fillMaxWidth()) {
+                    Column(Modifier.weight(1f)) {
+                        Text(p.name, fontWeight = FontWeight.Medium)
+                        Text(
+                            "t=${"%.2f".format(p.sampling.temperature)} · top-p=${"%.2f".format(p.sampling.topP)} · " +
+                                "rep=${"%.2f".format(p.sampling.repeatPenalty)} · max=${p.sampling.maxTokens}",
+                            style = MaterialTheme.typography.bodySmall,
+                            color = MaterialTheme.colorScheme.onSurfaceVariant,
+                            fontFamily = FontFamily.Monospace
+                        )
+                    }
+                    TextButton(onClick = { editing = if (editing == i) null else i }) {
+                        Text(if (editing == i) "Fermer" else "Éditer")
+                    }
+                    TextButton(onClick = { onSave(presets.filterIndexed { j, _ -> j != i }) }) {
+                        Text("Suppr.", color = MaterialTheme.colorScheme.error)
+                    }
+                }
+                if (editing == i) {
+                    var draft by remember(i) { mutableStateOf(p.sampling) }
+                    var name by remember(i) { mutableStateOf(p.name) }
+                    OutlinedTextField(
+                        value = name, onValueChange = { name = it },
+                        label = { Text("Nom du preset") }, singleLine = true,
+                        modifier = Modifier.fillMaxWidth()
+                    )
+                    Spacer(Modifier.height(4.dp))
+                    SamplingEditor(draft) { draft = it }
+                    Row {
+                        Spacer(Modifier.weight(1f))
+                        Button(
+                            onClick = {
+                                editing = null
+                                onSave(presets.toMutableList().also { it[i] = Preset(name.ifBlank { p.name }, draft) })
+                            }
+                        ) { Text("Enregistrer") }
+                    }
+                }
+                Spacer(Modifier.height(8.dp))
+            }
+
+            // Nouveau preset
+            var creating by remember { mutableStateOf(false) }
+            if (!creating) {
+                TextButton(onClick = { creating = true }) { Text("+ Nouveau preset") }
+            } else {
+                var newName by remember { mutableStateOf("") }
+                var newSampling by remember { mutableStateOf(Sampling()) }
+                OutlinedTextField(
+                    value = newName, onValueChange = { newName = it },
+                    label = { Text("Nom du nouveau preset") }, singleLine = true,
+                    modifier = Modifier.fillMaxWidth()
+                )
+                Spacer(Modifier.height(4.dp))
+                SamplingEditor(newSampling) { newSampling = it }
+                Row {
+                    TextButton(onClick = { creating = false }) { Text("Annuler") }
+                    Spacer(Modifier.weight(1f))
+                    Button(
+                        enabled = newName.isNotBlank(),
+                        onClick = {
+                            creating = false
+                            onSave(presets + Preset(newName.trim(), newSampling))
+                        }
+                    ) { Text("Créer") }
+                }
+            }
+        }
+    }
+}

kazeia-android/app-admin/src/main/res/values/strings.xml

@@ -6,6 +6,7 @@
     <string name="nav_voices">Voix</string>
     <string name="nav_profiles">Profils patients</string>
     <string name="nav_prompts">Prompts</string>
+    <string name="nav_parameters">Paramètres</string>
     <string name="nav_rag">Base documentaire</string>
     <string name="nav_history">Historique</string>
     <string name="nav_telemetry">Telemetry</string>

kazeia-android/app/src/main/java/com/kazeia/config/ConfigStore.kt

@@ -26,7 +26,32 @@ class ConfigStore private constructor(private val file: File) {
     data class ModelConfig(
         val modelId: String,
         val systemPrompt: String,
-        val temperature: Float
+        // Paramètres d'échantillonnage. `temperature` historique conservé ; les
+        // autres ajoutés par la refonte presets 2026-06-18. ⚠ seul maxTokens est
+        // honoré par le moteur natif actuel — le reste prend effet quand le JNI
+        // expose le sampling (cf. docs/SAMPLING_ENGINE_SPEC.md). presetName = nom
+        // du preset appliqué (vide = réglages custom).
+        val temperature: Float,
+        val topP: Float = 0.85f,
+        val topK: Int = 40,
+        val repeatPenalty: Float = 1.1f,
+        val presencePenalty: Float = 0.0f,
+        val frequencyPenalty: Float = 0.0f,
+        val maxTokens: Int = 120,
+        val presetName: String = ""
+    )
+
+    /** Modèle d'échantillonnage nommé, réutilisable, applicable au Speaker ou au
+     *  Thinker. Géré (CRUD) depuis l'app admin, persisté dans la config. */
+    data class SamplingPreset(
+        val name: String,
+        val temperature: Float,
+        val topP: Float,
+        val topK: Int,
+        val repeatPenalty: Float,
+        val presencePenalty: Float,
+        val frequencyPenalty: Float,
+        val maxTokens: Int
     )
 
     data class RuntimeConfig(
@@ -69,7 +94,9 @@ class ConfigStore private constructor(private val file: File) {
          *  panneau de logs + métriques système (CPU/GPU/NPU/RAM). Quand false
          *  (défaut prod), le bouton est masqué — l'écran reste épuré pour le
          *  patient. Pilotable depuis l'app admin. */
-        val debugEnabled: Boolean = false
+        val debugEnabled: Boolean = false,
+        /** Bibliothèque de presets de sampling réutilisables (gérée depuis l'admin). */
+        val presets: List<SamplingPreset> = defaultPresets()
     )
 
     companion object {
@@ -111,17 +138,29 @@ class ConfigStore private constructor(private val file: File) {
             return store
         }
 
+        /** Presets de sampling d'usine — inspirés du tuning Kaz (binôme) + profils
+         *  thérapeutiques. Repère pour l'admin ; éditables/supprimables. */
+        fun defaultPresets(): List<SamplingPreset> = listOf(
+            SamplingPreset("Équilibré (défaut)", 0.7f, 0.85f, 40, 1.1f, 0.0f, 0.0f, 120),
+            SamplingPreset("Kaz chaleureux", 0.4f, 0.85f, 40, 1.05f, 0.2f, 0.2f, 256),
+            SamplingPreset("Factuel / froid", 0.3f, 0.70f, 40, 1.2f, 0.1f, 0.1f, 200),
+            SamplingPreset("Analyse (Thinker)", 0.2f, 0.70f, 40, 1.15f, 0.0f, 0.0f, 128),
+            SamplingPreset("Créatif", 0.9f, 0.95f, 60, 1.05f, 0.3f, 0.3f, 200)
+        )
+
         fun defaultConfig(): RuntimeConfig = RuntimeConfig(
             cascadeEnabled = false,  // mono-Speaker = contrat FROZEN actuel
             speaker = ModelConfig(
                 modelId = ModelRegistry.defaultSpeaker().id,
                 systemPrompt = DEFAULT_SPEAKER_PROMPT,
-                temperature = 0.7f
+                temperature = 0.7f,
+                maxTokens = 450   // réponse Speaker longue (valeur prod historique)
             ),
             thinker = ModelConfig(
                 modelId = ModelRegistry.defaultThinker().id,
                 systemPrompt = DEFAULT_THINKER_PROMPT,
-                temperature = 0.0f
+                temperature = 0.0f,
+                maxTokens = 128   // bloc clinique court
             ),
             ttsEnabled = true,
             sttEngine = "prod",
@@ -178,19 +217,51 @@ class ConfigStore private constructor(private val file: File) {
         root.put("rag_threshold", cfg.ragThreshold.toDouble())
         root.put("rag_top_k", cfg.ragTopK)
         root.put("debug_enabled", cfg.debugEnabled)
-        root.put("speaker", JSONObject().apply {
-            put("model_id", cfg.speaker.modelId)
-            put("system_prompt", cfg.speaker.systemPrompt)
-            put("temperature", cfg.speaker.temperature)
+        root.put("speaker", modelToJson(cfg.speaker))
+        root.put("thinker", modelToJson(cfg.thinker))
+        root.put("presets", org.json.JSONArray().apply {
+            cfg.presets.forEach { p ->
+                put(JSONObject().apply {
+                    put("name", p.name)
+                    put("temperature", p.temperature)
+                    put("top_p", p.topP)
+                    put("top_k", p.topK)
+                    put("repeat_penalty", p.repeatPenalty)
+                    put("presence_penalty", p.presencePenalty)
+                    put("frequency_penalty", p.frequencyPenalty)
+                    put("max_tokens", p.maxTokens)
                 })
-        root.put("thinker", JSONObject().apply {
-            put("model_id", cfg.thinker.modelId)
-            put("system_prompt", cfg.thinker.systemPrompt)
-            put("temperature", cfg.thinker.temperature)
+            }
         })
         return root.toString(2)
     }
 
+    private fun modelToJson(m: ModelConfig) = JSONObject().apply {
+        put("model_id", m.modelId)
+        put("system_prompt", m.systemPrompt)
+        put("temperature", m.temperature)
+        put("top_p", m.topP)
+        put("top_k", m.topK)
+        put("repeat_penalty", m.repeatPenalty)
+        put("presence_penalty", m.presencePenalty)
+        put("frequency_penalty", m.frequencyPenalty)
+        put("max_tokens", m.maxTokens)
+        put("preset_name", m.presetName)
+    }
+
+    private fun modelFromJson(js: JSONObject, d: ModelConfig) = ModelConfig(
+        modelId = js.optString("model_id", d.modelId),
+        systemPrompt = js.optString("system_prompt", d.systemPrompt),
+        temperature = js.optDouble("temperature", d.temperature.toDouble()).toFloat(),
+        topP = js.optDouble("top_p", d.topP.toDouble()).toFloat(),
+        topK = js.optInt("top_k", d.topK),
+        repeatPenalty = js.optDouble("repeat_penalty", d.repeatPenalty.toDouble()).toFloat(),
+        presencePenalty = js.optDouble("presence_penalty", d.presencePenalty.toDouble()).toFloat(),
+        frequencyPenalty = js.optDouble("frequency_penalty", d.frequencyPenalty.toDouble()).toFloat(),
+        maxTokens = js.optInt("max_tokens", d.maxTokens),
+        presetName = js.optString("preset_name", d.presetName)
+    )
+
     private fun fromJson(text: String): RuntimeConfig {
         val root = JSONObject(text)
         val defaults = defaultConfig()
@@ -204,20 +275,24 @@ class ConfigStore private constructor(private val file: File) {
             ragThreshold = root.optDouble("rag_threshold", defaults.ragThreshold.toDouble()).toFloat(),
             ragTopK = root.optInt("rag_top_k", defaults.ragTopK),
             debugEnabled = root.optBoolean("debug_enabled", defaults.debugEnabled),
-            speaker = root.optJSONObject("speaker")?.let { js ->
-                ModelConfig(
-                    modelId = js.optString("model_id", defaults.speaker.modelId),
-                    systemPrompt = js.optString("system_prompt", defaults.speaker.systemPrompt),
-                    temperature = js.optDouble("temperature", defaults.speaker.temperature.toDouble()).toFloat()
+            speaker = root.optJSONObject("speaker")?.let { modelFromJson(it, defaults.speaker) } ?: defaults.speaker,
+            thinker = root.optJSONObject("thinker")?.let { modelFromJson(it, defaults.thinker) } ?: defaults.thinker,
+            presets = root.optJSONArray("presets")?.let { arr ->
+                (0 until arr.length()).mapNotNull { i ->
+                    arr.optJSONObject(i)?.let { js ->
+                        SamplingPreset(
+                            name = js.optString("name", ""),
+                            temperature = js.optDouble("temperature", 0.7).toFloat(),
+                            topP = js.optDouble("top_p", 0.85).toFloat(),
+                            topK = js.optInt("top_k", 40),
+                            repeatPenalty = js.optDouble("repeat_penalty", 1.1).toFloat(),
+                            presencePenalty = js.optDouble("presence_penalty", 0.0).toFloat(),
+                            frequencyPenalty = js.optDouble("frequency_penalty", 0.0).toFloat(),
+                            maxTokens = js.optInt("max_tokens", 120)
                         )
-            } ?: defaults.speaker,
-            thinker = root.optJSONObject("thinker")?.let { js ->
-                ModelConfig(
-                    modelId = js.optString("model_id", defaults.thinker.modelId),
-                    systemPrompt = js.optString("system_prompt", defaults.thinker.systemPrompt),
-                    temperature = js.optDouble("temperature", defaults.thinker.temperature.toDouble()).toFloat()
-                )
-            } ?: defaults.thinker
+                    }
+                }.filter { it.name.isNotBlank() }
+            } ?: defaults.presets
         )
     }
 }

kazeia-android/app/src/main/java/com/kazeia/core/LlmEngine.kt

@@ -22,7 +22,12 @@ data class SamplingParams(
     val temperature: Float = 0.7f,
     val topP: Float = 0.85f,
     val topK: Int = 40,
-    val repetitionPenalty: Float = 1.2f
+    val repetitionPenalty: Float = 1.2f,
+    // Pénalités OpenAI-style (présence/fréquence). Réglables via les presets admin.
+    // ⚠ effectives seulement quand le moteur natif expose le sampling (cf. docs
+    //    SAMPLING_ENGINE_SPEC) ; aujourd'hui le JNI n'honore que maxNewTokens.
+    val presencePenalty: Float = 0.0f,
+    val frequencyPenalty: Float = 0.0f
 )
 
 data class GenerationResult(

kazeia-android/app/src/main/java/com/kazeia/service/KazeiaService.kt

@@ -1742,9 +1742,7 @@ Pas d'introduction, pas d'explication. Juste les 4 lignes en francais. /no_think
                 val result = tE.generateWithSystem(
                     prompt = "Message du patient : $patientMessage",
                     systemPromptOverride = cfg.thinker.systemPrompt,
-                    params = com.kazeia.core.SamplingParams(
-                        maxNewTokens = 128, temperature = cfg.thinker.temperature
-                    ),
+                    params = samplingFrom(cfg.thinker),
                     onToken = null,
                     tag = "THINKER"
                 )
@@ -1947,10 +1945,9 @@ Pas d'introduction, pas d'explication. Juste les 4 lignes en francais. /no_think
                 }
                 !stoppingCriteria.shouldStop(responseBuilder.toString())
             }
-            val sParams = SamplingParams(
-                maxNewTokens = 450,
-                temperature = conversationManager.currentTemperature()
-            )
+            // Sampling Speaker depuis la config (preset admin). maxTokens honoré par
+            // le moteur ; les autres knobs prendront effet avec l'API sampling natif.
+            val sParams = samplingFrom(cfg.speaker)
             // generateWithSystem propage le prompt système Speaker par tour (bullets
             // cascade ou hot-reload admin). UnifiedLlmAdapter le gère pour GGUF
             // (mono-tour si override) comme pour .pte.
@@ -2171,6 +2168,20 @@ Pas d'introduction, pas d'explication. Juste les 4 lignes en francais. /no_think
      *    libération immédiate.
      *  - Thinker model_id changé en cascade ON → release, sera lazy-loadé.
      */
+    /** ModelConfig (preset admin) → SamplingParams pour la génération. ⚠ seul
+     *  maxNewTokens est honoré par le moteur natif actuel ; le reste est prêt et
+     *  prendra effet dès que le JNI expose le sampling (docs/SAMPLING_ENGINE_SPEC). */
+    private fun samplingFrom(m: com.kazeia.config.ConfigStore.ModelConfig) =
+        com.kazeia.core.SamplingParams(
+            maxNewTokens = m.maxTokens,
+            temperature = m.temperature,
+            topP = m.topP,
+            topK = m.topK,
+            repetitionPenalty = m.repeatPenalty,
+            presencePenalty = m.presencePenalty,
+            frequencyPenalty = m.frequencyPenalty
+        )
+
     private fun handleConfigChange(new: com.kazeia.config.ConfigStore.RuntimeConfig) {
         val old = runtimeConfig
         runtimeConfig = new

kazeia-android/app/src/main/java/com/kazeia/telemetry/KazeiaTelemetryProvider.kt

@@ -258,9 +258,17 @@ class KazeiaTelemetryProvider : ContentProvider() {
             "stt_engine", "llm_engine", "tts_engine", "rag_enabled",
             "rag_threshold", "rag_top_k", "debug_enabled",
             "speaker_model_id", "speaker_system_prompt", "speaker_temperature",
-            "thinker_model_id", "thinker_system_prompt", "thinker_temperature"
+            "speaker_top_p", "speaker_top_k", "speaker_repeat_penalty",
+            "speaker_presence_penalty", "speaker_frequency_penalty",
+            "speaker_max_tokens", "speaker_preset_name",
+            "thinker_model_id", "thinker_system_prompt", "thinker_temperature",
+            "thinker_top_p", "thinker_top_k", "thinker_repeat_penalty",
+            "thinker_presence_penalty", "thinker_frequency_penalty",
+            "thinker_max_tokens", "thinker_preset_name",
+            "presets_json"
         )
         val cursor = MatrixCursor(cols)
+        val s = cfg.speaker; val t = cfg.thinker
         cursor.addRow(arrayOf(
             if (cfg.cascadeEnabled) 1 else 0,
             if (cfg.ttsEnabled) 1 else 0,
@@ -268,12 +276,31 @@ class KazeiaTelemetryProvider : ContentProvider() {
             if (cfg.ragEnabled) 1 else 0,
             cfg.ragThreshold, cfg.ragTopK,
             if (cfg.debugEnabled) 1 else 0,
-            cfg.speaker.modelId, cfg.speaker.systemPrompt, cfg.speaker.temperature,
-            cfg.thinker.modelId, cfg.thinker.systemPrompt, cfg.thinker.temperature
+            s.modelId, s.systemPrompt, s.temperature,
+            s.topP, s.topK, s.repeatPenalty, s.presencePenalty, s.frequencyPenalty, s.maxTokens, s.presetName,
+            t.modelId, t.systemPrompt, t.temperature,
+            t.topP, t.topK, t.repeatPenalty, t.presencePenalty, t.frequencyPenalty, t.maxTokens, t.presetName,
+            presetsToJson(cfg.presets)
         ))
         return cursor
     }
 
+    /** Sérialise la bibliothèque de presets en JSON (1 colonne `presets_json`). */
+    private fun presetsToJson(presets: List<com.kazeia.config.ConfigStore.SamplingPreset>): String {
+        val arr = org.json.JSONArray()
+        presets.forEach { p ->
+            arr.put(org.json.JSONObject().apply {
+                put("name", p.name); put("temperature", p.temperature)
+                put("top_p", p.topP); put("top_k", p.topK)
+                put("repeat_penalty", p.repeatPenalty)
+                put("presence_penalty", p.presencePenalty)
+                put("frequency_penalty", p.frequencyPenalty)
+                put("max_tokens", p.maxTokens)
+            })
+        }
+        return arr.toString()
+    }
+
     private fun modelsCursor(): Cursor {
         val cols = arrayOf(
             "id", "display_name", "pte_path", "tokenizer_path",
@@ -364,20 +391,9 @@ class KazeiaTelemetryProvider : ContentProvider() {
             ragTopK = values.getAsInteger("rag_top_k") ?: cur.ragTopK,
             debugEnabled = if (values.containsKey("debug_enabled"))
                 values.getAsBoolean("debug_enabled") else cur.debugEnabled,
-            speaker = com.kazeia.config.ConfigStore.ModelConfig(
-                modelId = values.getAsString("speaker_model_id") ?: cur.speaker.modelId,
-                systemPrompt = values.getAsString("speaker_system_prompt")
-                    ?: cur.speaker.systemPrompt,
-                temperature = values.getAsFloat("speaker_temperature")
-                    ?: cur.speaker.temperature
-            ),
-            thinker = com.kazeia.config.ConfigStore.ModelConfig(
-                modelId = values.getAsString("thinker_model_id") ?: cur.thinker.modelId,
-                systemPrompt = values.getAsString("thinker_system_prompt")
-                    ?: cur.thinker.systemPrompt,
-                temperature = values.getAsFloat("thinker_temperature")
-                    ?: cur.thinker.temperature
-            )
+            speaker = mergeModel(values, "speaker_", cur.speaker),
+            thinker = mergeModel(values, "thinker_", cur.thinker),
+            presets = values.getAsString("presets_json")?.let { presetsFromJson(it) } ?: cur.presets
         )
         store.save(newCfg)
         ctx.contentResolver.notifyChange(uri, null)
@@ -385,6 +401,42 @@ class KazeiaTelemetryProvider : ContentProvider() {
         return 1
     }
 
+    /** Fusionne les colonnes `<prefix>*` d'un ContentValues dans un ModelConfig
+     *  (merge partiel : un champ absent reste inchangé). */
+    private fun mergeModel(
+        v: ContentValues, p: String, cur: com.kazeia.config.ConfigStore.ModelConfig
+    ) = com.kazeia.config.ConfigStore.ModelConfig(
+        modelId = v.getAsString("${p}model_id") ?: cur.modelId,
+        systemPrompt = v.getAsString("${p}system_prompt") ?: cur.systemPrompt,
+        temperature = v.getAsFloat("${p}temperature") ?: cur.temperature,
+        topP = v.getAsFloat("${p}top_p") ?: cur.topP,
+        topK = v.getAsInteger("${p}top_k") ?: cur.topK,
+        repeatPenalty = v.getAsFloat("${p}repeat_penalty") ?: cur.repeatPenalty,
+        presencePenalty = v.getAsFloat("${p}presence_penalty") ?: cur.presencePenalty,
+        frequencyPenalty = v.getAsFloat("${p}frequency_penalty") ?: cur.frequencyPenalty,
+        maxTokens = v.getAsInteger("${p}max_tokens") ?: cur.maxTokens,
+        presetName = v.getAsString("${p}preset_name") ?: cur.presetName
+    )
+
+    private fun presetsFromJson(text: String): List<com.kazeia.config.ConfigStore.SamplingPreset> =
+        runCatching {
+            val arr = org.json.JSONArray(text)
+            (0 until arr.length()).mapNotNull { i ->
+                arr.optJSONObject(i)?.let { js ->
+                    com.kazeia.config.ConfigStore.SamplingPreset(
+                        name = js.optString("name", ""),
+                        temperature = js.optDouble("temperature", 0.7).toFloat(),
+                        topP = js.optDouble("top_p", 0.85).toFloat(),
+                        topK = js.optInt("top_k", 40),
+                        repeatPenalty = js.optDouble("repeat_penalty", 1.1).toFloat(),
+                        presencePenalty = js.optDouble("presence_penalty", 0.0).toFloat(),
+                        frequencyPenalty = js.optDouble("frequency_penalty", 0.0).toFloat(),
+                        maxTokens = js.optInt("max_tokens", 120)
+                    )
+                }
+            }.filter { it.name.isNotBlank() }
+        }.getOrDefault(emptyList())
+
     private fun upsertProfile(values: ContentValues): Int {
         val ctx = context ?: return 0
         val store = com.kazeia.profiles.ProfileStore.get(ctx)

kazeia-android/docs/SAMPLING_ENGINE_SPEC.md

@@ -0,0 +1,93 @@
+# Spec — API d'échantillonnage natif (kazeia-engine) pour les presets Kazeia
+
+**Date** : 2026-06-18
+**Demandeur** : app Kazeia (refonte presets admin)
+**Cible** : `kazeia-engine` (lib `libkazeia_engine`, JNI `EngineJni`)
+
+## Contexte
+
+L'app admin permet désormais de régler, par modèle (Speaker / Thinker) et via des
+**presets** nommés, les paramètres d'échantillonnage : `temperature`, `top_p`,
+`top_k`, `repeat_penalty`, `presence_penalty`, `frequency_penalty`, `max_tokens`.
+
+Côté app, ces valeurs sont **persistées et transmises** jusqu'au pont JNI
+(`UnifiedLlmAdapter` → `EngineLlmEngine`/`LlmSession`). **MAIS le JNI actuel
+n'expose que `maxTok`** :
+
+```kotlin
+external fun generate(h: Long, sys: String, usr: String, maxTok: Int): String
+external fun sessionAsk(h: Long, usr: String, maxTok: Int, cb: TokenCallback)
+external fun generateStream(h: Long, sys: String, usr: String, maxTok: Int, cb: TokenCallback)
+```
+
+Donc aujourd'hui **seul `max_tokens` agit** ; température/top-p/k/pénalités sont
+inertes (l'app l'indique honnêtement dans l'UI : « effectif après MAJ moteur »).
+
+## Demande — exposer le sampling au JNI
+
+Ajouter une **variante échantillonnée** des entrées de génération (sans casser les
+signatures existantes), qui prend une struct de sampling et la câble sur la chaîne
+de samplers llama.cpp. Le GGUF Qwen3.5-4B (Speaker par défaut) est la cible
+prioritaire.
+
+### Signatures proposées
+
+```kotlin
+// Struct portée côté natif (ou 7 args primitifs si plus simple en JNI) :
+//   temperature: Float, topP: Float, topK: Int,
+//   repeatPenalty: Float, presencePenalty: Float, frequencyPenalty: Float, maxTokens: Int
+external fun generateSampled(
+    h: Long, sys: String, usr: String,
+    maxTokens: Int, temperature: Float, topP: Float, topK: Int,
+    repeatPenalty: Float, presencePenalty: Float, frequencyPenalty: Float,
+    cb: TokenCallback
+)
+external fun sessionAskSampled(
+    h: Long, usr: String,
+    maxTokens: Int, temperature: Float, topP: Float, topK: Int,
+    repeatPenalty: Float, presencePenalty: Float, frequencyPenalty: Float,
+    cb: TokenCallback
+)
+```
+
+### Mapping llama.cpp (sampler chain)
+
+Construire la chaîne par requête (ou réutiliser un `llama_sampler` reconfiguré) :
+
+```c
+auto * chain = llama_sampler_chain_init(llama_sampler_chain_default_params());
+llama_sampler_chain_add(chain, llama_sampler_init_top_k(top_k));
+llama_sampler_chain_add(chain, llama_sampler_init_top_p(top_p, 1));
+llama_sampler_chain_add(chain, llama_sampler_init_penalties(
+    /*penalty_last_n*/ 64, repeat_penalty, frequency_penalty, presence_penalty));
+llama_sampler_chain_add(chain, llama_sampler_init_temp(temperature));
+llama_sampler_chain_add(chain, llama_sampler_init_dist(/*seed*/ LLAMA_DEFAULT_SEED));
+```
+
+- `temperature <= 0` ⇒ greedy (`llama_sampler_init_greedy`), ignorer les autres.
+- Conserver le **thinking-off** et le template ChatML actuels inchangés.
+- `presence/frequency_penalty = 0` ⇒ no-op (comportement neutre).
+
+### Comportement attendu
+
+- Valeurs de référence (preset « Kaz chaleureux ») : temp 0.4, top_p 0.85, top_k 40,
+  repeat 1.05, presence 0.2, frequency 0.2, max 256 — doit produire des sorties
+  visiblement moins répétitives / plus chaleureuses que le greedy actuel.
+- Déterminisme : exposer un `seed` optionnel plus tard si besoin (pas requis v1).
+
+### `.pte` (NPU) — hors scope v1
+
+Le runner ExecuTorch `.pte` décode **greedy**. Le sampling y est une évolution
+distincte (échantillonnage post-logits côté runner). Pour l'instant les presets
+n'affectent que le chemin GGUF ; documenter la limite suffit.
+
+## Côté app — déjà prêt
+
+- `SamplingParams` (com.kazeia.core) porte les 7 knobs.
+- `UnifiedLlmAdapter.generateWithSystem` reçoit les params ; il suffira de router
+  vers `*Sampled` au lieu de `ask(max)` / `generateStream(max)` quand le JNI existe.
+- `ConfigStore.ModelConfig` + presets persistent les valeurs ; l'admin les édite.
+
+Quand le JNI échantillonné est livré : ~10 lignes à changer côté `UnifiedLlmAdapter`
+(brancher `sessionAskSampled` / `generateSampled`) et retirer la mention
+« effectif après MAJ moteur » de l'UI admin.