alpha new hybridretriver line

migrations/Version20260225000100.php

@@ -0,0 +1,32 @@
+<?php
+
+declare(strict_types=1);
+
+namespace DoctrineMigrations;
+
+use Doctrine\DBAL\Schema\Schema;
+use Doctrine\Migrations\AbstractMigration;
+
+final class Version20260225000100 extends AbstractMigration
+{
+    public function getDescription(): string
+    {
+        return 'Adds retrieval_max_chunks and retrieval_vector_top_k to model_generation_config (without DB defaults)';
+    }
+
+    public function up(Schema $schema): void
+    {
+        $this->addSql('ALTER TABLE model_generation_config 
+            ADD retrieval_max_chunks INT NOT NULL,
+            ADD retrieval_vector_top_k INT NOT NULL
+        ');
+    }
+
+    public function down(Schema $schema): void
+    {
+        $this->addSql('ALTER TABLE model_generation_config 
+            DROP retrieval_max_chunks,
+            DROP retrieval_vector_top_k
+        ');
+    }
+}

migrations/Version20260225000200.php

@@ -0,0 +1,35 @@
+<?php
+
+declare(strict_types=1);
+
+namespace DoctrineMigrations;
+
+use Doctrine\DBAL\Schema\Schema;
+use Doctrine\Migrations\AbstractMigration;
+
+final class Version20260225000200 extends AbstractMigration
+{
+    public function getDescription(): string
+    {
+        return 'Fix schema drift (indexes, FK, longtext nullability)';
+    }
+
+    public function up(Schema $schema): void
+    {
+        $this->addSql('ALTER TABLE knowledge_tag CHANGE description description LONGTEXT DEFAULT NULL');
+        $this->addSql('ALTER TABLE knowledge_tag RENAME INDEX uniq_knowledge_tag_slug TO UNIQ_86B46255989D9B62');
+
+        $this->addSql('ALTER TABLE document DROP FOREIGN KEY FK_D8698A769407EE77');
+        $this->addSql('ALTER TABLE document ADD CONSTRAINT FK_D8698A769407EE77 FOREIGN KEY (current_version_id) REFERENCES document_version (id)');
+
+        $this->addSql('ALTER TABLE tag_rebuild_job CHANGE error_message error_message LONGTEXT DEFAULT NULL');
+
+        $this->addSql('ALTER TABLE document_tag RENAME INDEX idx_document_tag_document TO IDX_D0234567C33F7837');
+        $this->addSql('ALTER TABLE document_tag RENAME INDEX idx_document_tag_tag TO IDX_D0234567BAD26311');
+    }
+
+    public function down(Schema $schema): void
+    {
+        // Optional – in Drift-Fix-Migration meist nicht notwendig
+    }
+}

src/Controller/Admin/ModelGenerationConfigController.php

@@ -5,6 +5,7 @@ declare(strict_types=1);
 namespace App\Controller\Admin;
 
 use App\Entity\ModelGenerationConfig;
+use App\Knowledge\Retrieval\NdjsonHybridRetriever;
 use App\Repository\ModelGenerationConfigRepository;
 use App\Service\ModelGenerationConfigManager;
 use Doctrine\ORM\EntityManagerInterface;
@@ -21,7 +22,10 @@ class ModelGenerationConfigController extends AbstractController
     {
         $this->denyAccessUnlessGranted('ROLE_KNOWLEDGE_ADMIN');
 
-        $configs = $repository->findBy([], ['active' => 'DESC', 'modelName' => 'ASC', 'version' => 'DESC']);
+        $configs = $repository->findBy(
+            [],
+            ['active' => 'DESC', 'modelName' => 'ASC', 'version' => 'DESC']
+        );
 
         return $this->render('admin/model_config/list.html.twig', [
             'configs' => $configs,
@@ -38,20 +42,24 @@ class ModelGenerationConfigController extends AbstractController
 
         if ($request->isMethod('POST')) {
 
-            $modelName = $request->request->get('model_name');
+            $modelName = (string) $request->request->get('model_name');
-
             $version = $repository->findNextVersion($modelName);
 
+            $retrievalMaxChunks = (int) $request->request->get('retrieval_max_chunks', 25);
+            $retrievalVectorTopK = (int) $request->request->get('retrieval_vector_top_k', 25);
+
             $config = new ModelGenerationConfig(
                 modelName: $modelName,
                 version: $version,
-                stream: (bool)$request->request->get('stream'),
+                stream: (bool) $request->request->get('stream'),
-                temperature: (float)$request->request->get('temperature'),
+                temperature: (float) $request->request->get('temperature'),
-                topK: (int)$request->request->get('top_k'),
+                topK: (int) $request->request->get('top_k'),
-                topP: (float)$request->request->get('top_p'),
+                topP: (float) $request->request->get('top_p'),
-                repeatPenalty: (float)$request->request->get('repeat_penalty'),
+                repeatPenalty: (float) $request->request->get('repeat_penalty'),
-                numCtx: (int)$request->request->get('num_ctx'),
+                numCtx: (int) $request->request->get('num_ctx'),
-                active: false
+                active: false,
+                retrievalMaxChunks: $retrievalMaxChunks,
+                retrievalVectorTopK: $retrievalVectorTopK
             );
 
             $em->persist($config);
@@ -75,6 +83,32 @@ class ModelGenerationConfigController extends AbstractController
         return $this->redirectToRoute('admin_model_config_list');
     }
 
+    #[Route('/{id}/test-retrieval', name: 'admin_model_config_test_retrieval')]
+    public function testRetrieval(
+        ModelGenerationConfig $config,
+        Request $request,
+        NdjsonHybridRetriever $retriever
+    ): Response {
+        $this->denyAccessUnlessGranted('ROLE_KNOWLEDGE_ADMIN');
+
+        $results = [];
+        $prompt = '';
+
+        if ($request->isMethod('POST')) {
+            $prompt = trim((string) $request->request->get('prompt'));
+
+            if ($prompt !== '') {
+                $results = $retriever->retrieveForConfig($prompt, $config);
+            }
+        }
+
+        return $this->render('admin/model_config/test_retrieval.html.twig', [
+            'config'  => $config,
+            'results' => $results,
+            'prompt'  => $prompt,
+        ]);
+    }
+
     #[Route('/{id}/delete', name: 'admin_model_config_delete', methods: ['POST'])]
     public function delete(
         ModelGenerationConfig $config,
@@ -99,5 +133,4 @@ class ModelGenerationConfigController extends AbstractController
 
         return $this->redirectToRoute('admin_model_config_list');
     }
-
+}
-}

src/Entity/ModelGenerationConfig.php

@@ -9,9 +9,15 @@ use Symfony\Component\Uid\Uuid;
 
 #[ORM\Entity(repositoryClass: \App\Repository\ModelGenerationConfigRepository::class)]
 #[ORM\Table(name: 'model_generation_config')]
-#[ORM\Index(columns: ['model_name', 'active'], name: 'idx_model_active')]
+#[ORM\Index(name: 'idx_model_active', columns: ['model_name', 'active'])]
 class ModelGenerationConfig
 {
+    // -----------------------------
+    // Hard Guardrails
+    // -----------------------------
+    private const MAX_RETRIEVAL_CHUNKS = 200;
+    private const MAX_VECTOR_TOPK = 200;
+
     #[ORM\Id]
     #[ORM\Column(type: 'uuid', unique: true)]
     private Uuid $id;
@@ -37,6 +43,16 @@ class ModelGenerationConfig
     #[ORM\Column(name: 'num_ctx', type: 'integer')]
     private int $numCtx;
 
+    // -------------------------------------
+    // Retrieval-Parameter (NEU)
+    // -------------------------------------
+
+    #[ORM\Column(name: 'retrieval_max_chunks', type: 'integer')]
+    private int $retrievalMaxChunks;
+
+    #[ORM\Column(name: 'retrieval_vector_top_k', type: 'integer')]
+    private int $retrievalVectorTopK;
+
     #[ORM\Column(type: 'boolean')]
     private bool $active;
 
@@ -55,7 +71,9 @@ class ModelGenerationConfig
         float $topP = 0.8,
         float $repeatPenalty = 1.05,
         int $numCtx = 4096,
-        bool $active = false
+        bool $active = false,
+        int $retrievalMaxChunks = 25,
+        int $retrievalVectorTopK = 25,
     ) {
         $this->id = Uuid::v4();
         $this->modelName = $modelName;
@@ -68,8 +86,15 @@ class ModelGenerationConfig
         $this->numCtx = $numCtx;
         $this->active = $active;
         $this->createdAt = new \DateTimeImmutable();
+
+        $this->setRetrievalMaxChunks($retrievalMaxChunks);
+        $this->setRetrievalVectorTopK($retrievalVectorTopK);
     }
 
+    // -----------------------------
+    // Getter
+    // -----------------------------
+
     public function getId(): Uuid { return $this->id; }
     public function getModelName(): string { return $this->modelName; }
     public function isStream(): bool { return $this->stream; }
@@ -82,9 +107,35 @@ class ModelGenerationConfig
     public function getVersion(): int { return $this->version; }
     public function getCreatedAt(): \DateTimeImmutable { return $this->createdAt; }
 
+    public function getRetrievalMaxChunks(): int
+    {
+        return $this->retrievalMaxChunks;
+    }
+
+    public function getRetrievalVectorTopK(): int
+    {
+        return $this->retrievalVectorTopK;
+    }
+
+    // -----------------------------
+    // Controlled Mutators
+    // -----------------------------
+
     // Nur vom Manager nutzen
     public function setActive(bool $active): void
     {
         $this->active = $active;
     }
-}
+
+    public function setRetrievalMaxChunks(int $value): void
+    {
+        $value = max(1, min($value, self::MAX_RETRIEVAL_CHUNKS));
+        $this->retrievalMaxChunks = $value;
+    }
+
+    public function setRetrievalVectorTopK(int $value): void
+    {
+        $value = max(1, min($value, self::MAX_VECTOR_TOPK));
+        $this->retrievalVectorTopK = $value;
+    }
+}

src/Knowledge/Retrieval/NdjsonHybridRetriever.php

@@ -4,43 +4,65 @@ declare(strict_types=1);
 
 namespace App\Knowledge\Retrieval;
 
+use App\Entity\ModelGenerationConfig;
 use App\Knowledge\ChunkManager;
+use App\Repository\ModelGenerationConfigRepository;
 use App\Tag\TagRoutingService;
 use App\Vector\VectorSearchClient;
 
 final class NdjsonHybridRetriever implements RetrieverInterface
 {
     private const VECTOR_SCORE_THRESHOLD = 0.25;
-
-    /**
-     * Wenn Tag-Routing aktiv ist, erhöhen wir TopK,
-     * weil wir danach per document_id filtern.
-     */
     private const VECTOR_TOPK_MULTIPLIER_WHEN_ROUTED = 10;
-
-    /**
-     * Keyword-Scan: Mindest-Trefferanzahl an Terms, damit ein Chunk als Kandidat gilt.
-     */
     private const KEYWORD_MIN_HITS = 1;
 
+    private const HARD_MAX_CHUNKS = 200;
+    private const HARD_MAX_VECTORK = 200;
+
     public function __construct(
-        private readonly ChunkManager       $chunkManager,
+        private readonly ChunkManager $chunkManager,
-        private readonly NdjsonChunkLookup  $lookup,
+        private readonly NdjsonChunkLookup $lookup,
         private readonly VectorSearchClient $vectorClient,
-        private readonly TagRoutingService  $tagRouting,
+        private readonly TagRoutingService $tagRouting,
-        private readonly int                $maxChunks = 100,
+        private readonly ModelGenerationConfigRepository $configRepository,
-        private readonly int                $vectorTopK = 100,
     ) {}
 
-    public function retrieve(string $prompt, int $limit = null): array
+    /**
+     * Normalbetrieb – ausschließlich aktive Config.
+     */
+    public function retrieve(string $prompt): array
     {
-        $limit = $this->maxChunks;
+        $config = $this->configRepository->findActiveForModel();
+
+        if ($config === null) {
+            throw new \RuntimeException('No active ModelGenerationConfig found.');
+        }
+
+        return $this->retrieveInternal($prompt, $config);
+    }
+
+    /**
+     * Admin-Testbetrieb – explizite Config.
+     * Verändert KEINEN globalen Zustand.
+     */
+    public function retrieveForConfig(string $prompt, ModelGenerationConfig $config): array
+    {
+        return $this->retrieveInternal($prompt, $config);
+    }
+
+    /**
+     * Zentrale Retrieval-Logik (keine Duplikation).
+     */
+    private function retrieveInternal(string $prompt, ModelGenerationConfig $config): array
+    {
+        $limit = max(1, min($config->getRetrievalMaxChunks(), self::HARD_MAX_CHUNKS));
+
+        $vectorTopKBase = max(1, min($config->getRetrievalVectorTopK(), self::HARD_MAX_VECTORK));
 
         // ---------------------------------------------------------
-        // 0) Tag-Routing FIRST (soft gate)
+        // 1) Tag-Vector FIRST  -> candidateSet (DocIDs)
         // ---------------------------------------------------------
-        $candidateDocIds = $this->tagRouting->route($prompt);
+        $candidateDocIds = $this->tagRouting->route($prompt); // <= DAS muss intern auf Tag-Vector gehen
-
         $candidateSet = null;
 
         if (is_array($candidateDocIds) && $candidateDocIds !== []) {
@@ -48,31 +70,22 @@ final class NdjsonHybridRetriever implements RetrieverInterface
         }
 
         // ---------------------------------------------------------
-        // 1) Keyword first (simple streaming scan)
+        // 2) Vector chunks (primary)
         // ---------------------------------------------------------
-        $terms = $this->extractTerms($prompt);
+        $topK = $vectorTopKBase;
-
-        $keywordChunks = $this->keywordSearchStreaming($terms, $limit, $candidateSet);
-
-        if (\count($keywordChunks) >= $limit) {
-            return array_slice($keywordChunks, 0, $limit);
-        }
-
-        // ---------------------------------------------------------
-        // 2) Vector fallback / enrichment
-        //    - If routed: increase TopK, then filter by document_id
-        //    - Soft fallback: if filtering yields nothing -> global vector once
-        // ---------------------------------------------------------
-        $topK = $this->vectorTopK;
 
         if ($candidateSet !== null) {
-            $topK = max($this->vectorTopK * self::VECTOR_TOPK_MULTIPLIER_WHEN_ROUTED, $this->vectorTopK);
+            $topK = min(
-            $topK = min($topK, 200); // guardrail
+                max($vectorTopKBase * self::VECTOR_TOPK_MULTIPLIER_WHEN_ROUTED, $vectorTopKBase),
+                self::HARD_MAX_VECTORK
+            );
         }
 
         $hits = $this->vectorClient->search($prompt, $topK);
+
         if ($hits === []) {
-            return $keywordChunks;
+            // Tags-only System: kein Vector-Hit -> keine Chunks
+            return [];
         }
 
         $chunkIds = [];
@@ -87,14 +100,15 @@ final class NdjsonHybridRetriever implements RetrieverInterface
         }
 
         if ($chunkIds === []) {
-            return $keywordChunks;
+            return [];
         }
 
         $rows = $this->lookup->findByChunkIds($chunkIds);
-
-        // routed filtering by document_id
         $finalChunkIds = $chunkIds;
 
+        // ---------------------------------------------------------
+        // 3) Routed filtering (wenn candidateSet vorhanden)
+        // ---------------------------------------------------------
         if ($candidateSet !== null) {
             $filtered = [];
 
@@ -103,18 +117,20 @@ final class NdjsonHybridRetriever implements RetrieverInterface
                 if (!is_array($row)) {
                     continue;
                 }
+
                 $docId = $row['document_id'] ?? null;
                 if (!is_string($docId) || !isset($candidateSet[$docId])) {
                     continue;
                 }
+
                 $filtered[] = $id;
             }
 
-            // Soft fallback: if routing filtered everything away, retry global vector once
+            // Wenn Routing ALLES wegfiltert -> einmal global retry
             if ($filtered === []) {
-                $hits2 = $this->vectorClient->search($prompt, $this->vectorTopK);
+                $hits2 = $this->vectorClient->search($prompt, $vectorTopKBase);
                 if ($hits2 === []) {
-                    return $keywordChunks;
+                    return [];
                 }
 
                 $chunkIds2 = [];
@@ -129,7 +145,7 @@ final class NdjsonHybridRetriever implements RetrieverInterface
                 }
 
                 if ($chunkIds2 === []) {
-                    return $keywordChunks;
+                    return [];
                 }
 
                 $rows = $this->lookup->findByChunkIds($chunkIds2);
@@ -139,24 +155,25 @@ final class NdjsonHybridRetriever implements RetrieverInterface
             }
         }
 
-        foreach ($finalChunkIds as $id) {
-            if (!isset($rows[$id]['text']) || !is_string($rows[$id]['text'])) {
-                continue;
-            }
-            $keywordChunks[] = trim($rows[$id]['text']);
-        }
-
         // ---------------------------------------------------------
-        // 3) dedupe + limit
+        // 4) Collect texts + Dedupe + Limit
         // ---------------------------------------------------------
         $seen = [];
         $out = [];
 
-        foreach ($keywordChunks as $chunk) {
+        foreach ($finalChunkIds as $id) {
+            $text = $rows[$id]['text'] ?? null;
+            if (!is_string($text) || $text === '') {
+                continue;
+            }
+
+            $chunk = trim($text);
             $key = mb_strtolower((string)preg_replace('/\s+/u', ' ', $chunk));
+
             if (isset($seen[$key])) {
                 continue;
             }
+
             $seen[$key] = true;
             $out[] = $chunk;
 
@@ -168,16 +185,6 @@ final class NdjsonHybridRetriever implements RetrieverInterface
         return $out;
     }
 
-    /**
-     * Streaming Keyword Search über index.ndjson.
-     * Minimal, aber nützlich:
-     * - Score = Anzahl gefundener Terms
-     * - CandidateDocs (Tag-Routing) reduziert Scan massiv
-     *
-     * @param string[] $terms
-     * @param array<string,true>|null $candidateSet
-     * @return string[]
-     */
     private function keywordSearchStreaming(array $terms, int $limit, ?array $candidateSet): array
     {
         if ($terms === []) {
@@ -185,31 +192,28 @@ final class NdjsonHybridRetriever implements RetrieverInterface
         }
 
         $maxScore = \count($terms);
-
-        // top list: each item = ['score' => int, 'text' => string]
         $top = [];
 
         foreach ($this->chunkManager->streamAll() as $row) {
             $text = $row['text'] ?? null;
+
             if (!is_string($text) || $text === '') {
                 continue;
             }
 
             if ($candidateSet !== null) {
                 $docId = $row['document_id'] ?? null;
+
                 if (!is_string($docId) || !isset($candidateSet[$docId])) {
                     continue;
                 }
             }
 
             $haystack = mb_strtolower($text);
-
             $score = 0;
+
             foreach ($terms as $t) {
-                if ($t === '') {
+                if ($t !== '' && mb_stripos($haystack, $t) !== false) {
-                    continue;
-                }
-                if (mb_stripos($haystack, $t) !== false) {
                     $score++;
                 }
             }
@@ -223,14 +227,11 @@ final class NdjsonHybridRetriever implements RetrieverInterface
                 'text'  => trim($text),
             ];
 
-            // keep only best N (simple sort, N is tiny)
             usort($top, static function (array $a, array $b): int {
-                // higher score first
                 $cmp = ($b['score'] <=> $a['score']);
                 if ($cmp !== 0) {
                     return $cmp;
                 }
-                // shorter chunk first (often more precise)
                 return (mb_strlen($a['text']) <=> mb_strlen($b['text']));
             });
 
@@ -238,25 +239,14 @@ final class NdjsonHybridRetriever implements RetrieverInterface
                 $top = array_slice($top, 0, $limit);
             }
 
-            // early exit: perfect matches filled
             if (\count($top) === $limit && ($top[0]['score'] ?? 0) >= $maxScore) {
                 break;
             }
         }
 
-        $out = [];
+        return array_map(static fn($item) => (string)$item['text'], $top);
-        foreach ($top as $item) {
-            $out[] = (string)$item['text'];
-        }
-
-        return $out;
     }
 
-    /**
-     * Minimal term extraction (stabiles Verhalten, wenig Magie)
-     *
-     * @return string[]
-     */
     private function extractTerms(string $text): array
     {
         $text = mb_strtolower((string)preg_replace('/[^\p{L}\p{N}\s]/u', '', $text));
@@ -266,15 +256,14 @@ final class NdjsonHybridRetriever implements RetrieverInterface
             static fn(string $w) => mb_strlen($w) > 2
         ));
 
-        // unique, order preserved
         $seen = [];
         $out = [];
+
         foreach ($parts as $w) {
-            if (isset($seen[$w])) {
+            if (!isset($seen[$w])) {
-                continue;
+                $seen[$w] = true;
+                $out[] = $w;
             }
-            $seen[$w] = true;
-            $out[] = $w;
         }
 
         return $out;

src/Knowledge/Retrieval/RetrieverInterface.php

@@ -1,11 +1,28 @@
 <?php
 
+declare(strict_types=1);
+
 namespace App\Knowledge\Retrieval;
 
+/**
+ * Retrieval ist vollständig konfigurationsgetrieben.
+ *
+ * - retrievalMaxChunks stammt ausschließlich aus der aktiven ModelGenerationConfig.
+ * - retrievalVectorTopK stammt ausschließlich aus der aktiven ModelGenerationConfig.
+ * - Es existiert kein Runtime-Override.
+ *
+ * Ziel:
+ * Deterministisches, auditierbares Retrieval-Verhalten.
+ */
 interface RetrieverInterface
 {
     /**
-     * @return string[]  Plain text knowledge chunks
+     * Retrieves relevant knowledge chunks for a given prompt.
+     *
+     * The number of returned chunks is strictly defined by
+     * the active ModelGenerationConfig (retrievalMaxChunks).
+     *
+     * @return string[] Plain text knowledge chunks
      */
-    public function retrieve(string $prompt, int $limit = 10): array;
+    public function retrieve(string $prompt): array;
-}
+}

templates/admin/model_config/create.html.twig

@@ -22,7 +22,10 @@
 
                 <div class="row g-4">
 
-                    <!-- Modell -->
+                    <!-- ============================== -->
+                    <!-- Modell-Basis -->
+                    <!-- ============================== -->
+
                     <div class="col-md-6">
                         <label class="form-label">Modellname</label>
                         <input type="text"
@@ -31,11 +34,10 @@
                                placeholder="z. B. qwen3:latest"
                                required>
                         <div class="form-text text-secondary">
-                            Exakter Modellname wie im KI-Endpunkt konfiguriert (z. B. Ollama oder API).
+                            Exakter Modellname wie im KI-Endpunkt konfiguriert.
                         </div>
                     </div>
 
-                    <!-- Stream -->
                     <div class="col-md-6 d-flex align-items-center">
                         <div class="form-check form-switch mt-4">
                             <input class="form-check-input"
@@ -48,11 +50,22 @@
                             </label>
                         </div>
                         <div class="form-text text-secondary ms-3">
-                            Aktiviert Token-Streaming im Chat (empfohlen für bessere UX).
+                            Token-Streaming im Chat (empfohlen).
                         </div>
                     </div>
 
-                    <!-- Temperature -->
+                </div>
+
+                <hr class="border-secondary my-4">
+
+                <!-- ============================== -->
+                <!-- GENERATION -->
+                <!-- ============================== -->
+
+                <h5 class="text-info mb-3">Generation (LLM Sampling)</h5>
+
+                <div class="row g-4">
+
                     <div class="col-md-4">
                         <label class="form-label">Temperature</label>
                         <input type="number"
@@ -64,29 +77,24 @@
                                class="form-control bg-dark text-light border-secondary"
                                required>
                         <div class="form-text text-secondary">
-                            Steuert die Kreativität der Antworten.
+                            Kreativität des Modells. Für RAG 0.2–0.4 empfohlen.
-                            Niedrige Werte (0.2–0.4) erzeugen stabile, sachliche Ergebnisse – empfohlen für RAG-Systeme.
-                            Höhere Werte führen zu freieren, weniger deterministischen Antworten.
                         </div>
                     </div>
 
-                    <!-- Top K -->
                     <div class="col-md-4">
                         <label class="form-label">Top K</label>
                         <input type="number"
                                min="1"
+                               max="200"
                                name="top_k"
                                value="40"
                                class="form-control bg-dark text-light border-secondary"
                                required>
                         <div class="form-text text-secondary">
-                            Begrenzt die Anzahl der wahrscheinlichsten Token, aus denen das Modell auswählt.
+                            Begrenzt die Auswahl wahrscheinlicher Token.
-                            Niedrigere Werte = konservativer, höhere Werte = flexibler.
-                            20–50 ist für Wissenssysteme üblich.
                         </div>
                     </div>
 
-                    <!-- Top P -->
                     <div class="col-md-4">
                         <label class="form-label">Top P</label>
                         <input type="number"
@@ -98,13 +106,10 @@
                                class="form-control bg-dark text-light border-secondary"
                                required>
                         <div class="form-text text-secondary">
-                            Nucleus Sampling: Das Modell berücksichtigt nur Token,
+                            Nucleus Sampling – Balance zwischen Stabilität und Natürlichkeit.
-                            deren kumulative Wahrscheinlichkeit innerhalb dieses Werts liegt.
-                            0.8–0.95 bietet eine gute Balance zwischen Stabilität und Natürlichkeit.
                         </div>
                     </div>
 
-                    <!-- Repeat Penalty -->
                     <div class="col-md-6">
                         <label class="form-label">Repeat Penalty</label>
                         <input type="number"
@@ -116,12 +121,10 @@
                                class="form-control bg-dark text-light border-secondary"
                                required>
                         <div class="form-text text-secondary">
-                            Bestraft Wortwiederholungen. Werte leicht über 1.0 (z. B. 1.1–1.15)
+                            Verhindert Wiederholungen (1.1–1.15 empfohlen).
-                            verhindern Schleifen und redundante Antworten.
                         </div>
                     </div>
 
-                    <!-- Num Ctx -->
                     <div class="col-md-6">
                         <label class="form-label">Context Window (num_ctx)</label>
                         <input type="number"
@@ -132,9 +135,50 @@
                                class="form-control bg-dark text-light border-secondary"
                                required>
                         <div class="form-text text-secondary">
-                            Maximale Kontextlänge in Tokens (Systemprompt + Benutzerfrage + Retrieval-Chunks).
+                            Maximale Kontextlänge (System + Frage + Retrieval).
-                            Muss vom Modell unterstützt werden.
+                        </div>
-                            Höhere Werte ermöglichen größere Wissenskontexte, erhöhen jedoch Speicher- und Rechenbedarf.
+                    </div>
+
+                </div>
+
+                <hr class="border-secondary my-4">
+
+                <!-- ============================== -->
+                <!-- RETRIEVAL -->
+                <!-- ============================== -->
+
+                <h5 class="text-warning mb-3">Retrieval (Wissensabruf)</h5>
+
+                <div class="row g-4">
+
+                    <div class="col-md-6">
+                        <label class="form-label">Max Chunks</label>
+                        <input type="number"
+                               min="1"
+                               max="200"
+                               name="retrieval_max_chunks"
+                               value="25"
+                               class="form-control bg-dark text-light border-secondary"
+                               required>
+                        <div class="form-text text-secondary">
+                            Maximale Anzahl an Wissens-Chunks,
+                            die dem Modell übergeben werden.
+                            20–40 ist für die meisten Systeme optimal.
+                        </div>
+                    </div>
+
+                    <div class="col-md-6">
+                        <label class="form-label">Vector Top K</label>
+                        <input type="number"
+                               min="1"
+                               max="200"
+                               name="retrieval_vector_top_k"
+                               value="25"
+                               class="form-control bg-dark text-light border-secondary"
+                               required>
+                        <div class="form-text text-secondary">
+                            Anzahl der Vektor-Treffer vor Filterung.
+                            Höhere Werte erhöhen Recall, können aber Rauschen verstärken.
                         </div>
                     </div>
 
@@ -157,4 +201,4 @@
         Hinweis: Neue Konfigurationen werden zunächst inaktiv gespeichert und
         müssen separat aktiviert werden. Pro Modell kann nur eine Version aktiv sein.
     </div>
-{% endblock %}
+{% endblock %}

templates/admin/model_config/list.html.twig

@@ -24,11 +24,8 @@
                     <th>Modell</th>
                     <th>Version</th>
                     <th>Stream</th>
-                    <th>Temp</th>
+                    <th>Sampling</th>
-                    <th>Top K</th>
+                    <th class="text-warning">Retrieval</th>
-                    <th>Top P</th>
-                    <th>Repeat</th>
-                    <th>Ctx</th>
                     <th>Status</th>
                     <th class="text-end">Aktionen</th>
                 </tr>
@@ -52,27 +49,47 @@
                             {% endif %}
                         </td>
 
-                        <td>{{ config.temperature }}</td>
+                        {# ========================= #}
-                        <td>{{ config.topK }}</td>
+                        {# Sampling Group           #}
-                        <td>{{ config.topP }}</td>
+                        {# ========================= #}
-                        <td>{{ config.repeatPenalty }}</td>
+                        <td>
-                        <td>{{ config.numCtx }}</td>
+                            <div class="small">
+                                <div><strong>T:</strong> {{ config.temperature }}</div>
+                                <div><strong>K:</strong> {{ config.topK }}</div>
+                                <div><strong>P:</strong> {{ config.topP }}</div>
+                                <div><strong>R:</strong> {{ config.repeatPenalty }}</div>
+                                <div><strong>Ctx:</strong> {{ config.numCtx }}</div>
+                            </div>
+                        </td>
+
+                        {# ========================= #}
+                        {# Retrieval Group          #}
+                        {# ========================= #}
+                        <td class="text-warning">
+                            <div class="small">
+                                <div><strong>Chunks:</strong> {{ config.retrievalMaxChunks }}</div>
+                                <div><strong>VectorK:</strong> {{ config.retrievalVectorTopK }}</div>
+                            </div>
+                        </td>
 
                         <td>
                             {% if config.active %}
                                 <span class="badge bg-success">Aktiv</span>
                             {% else %}
                                 <span class="badge bg-dark border border-secondary">
-                                Inaktiv
+                                    Inaktiv
-                            </span>
+                                </span>
                             {% endif %}
                         </td>
 
                         <td class="text-end">
+                            <a href="{{ path('admin_model_config_test_retrieval', {id: config.id}) }}"
+                               class="btn btn-sm btn-outline-warning me-1">
+                                Test Retrieval
+                            </a>
 
                             {% if not config.active and is_granted('ROLE_SUPER_ADMIN') %}
 
-                                {# Aktivieren via POST + CSRF #}
                                 <form method="post"
                                       action="{{ path('admin_model_config_activate', {id: config.id}) }}"
                                       class="d-inline"
@@ -82,12 +99,11 @@
                                            name="_token"
                                            value="{{ csrf_token('activate_model_config_' ~ config.id) }}">
 
-                                    <button class="btn btn-sm btn-outline-success me-2">
+                                    <button class="btn btn-sm btn-outline-success me-1">
                                         Aktivieren
                                     </button>
                                 </form>
 
-                                {# Löschen via POST + CSRF #}
                                 <form method="post"
                                       action="{{ path('admin_model_config_delete', {id: config.id}) }}"
                                       class="d-inline"
@@ -105,13 +121,12 @@
                             {% else %}
                                 <span class="text-secondary small">—</span>
                             {% endif %}
-
                         </td>
                     </tr>
 
                 {% else %}
                     <tr>
-                        <td colspan="10" class="text-center text-secondary py-4">
+                        <td colspan="7" class="text-center text-secondary py-4">
                             Keine Konfiguration vorhanden.
                         </td>
                     </tr>
@@ -132,13 +147,11 @@
 
     <div class="card bg-black border-info">
         <div class="card-body p-0">
-
             <iframe
                     src="/index.html?admin_test=1"
                     class="w-100 border-0"
                     style="height:75vh;"
             ></iframe>
-
         </div>
     </div>
 
@@ -146,4 +159,5 @@
         Der Agent läuft im isolierten Admin-Test-Modus.
         Keine Persistenz. Keine produktive Session.
     </div>
-{% endblock %}
+
+{% endblock %}

templates/admin/model_config/test_retrieval.html.twig

@@ -0,0 +1,93 @@
+{% extends 'admin/base.html.twig' %}
+
+{% block title %}Retrieval-Test{% endblock %}
+
+{% block body %}
+
+    <div class="d-flex justify-content-between align-items-center mb-4">
+        <h1 class="h3">
+            Retrieval-Test – {{ config.modelName }} (v{{ config.version }})
+        </h1>
+
+        <a href="{{ path('admin_model_config_list') }}"
+           class="btn btn-sm btn-outline-secondary">
+            Zurück
+        </a>
+    </div>
+
+    {# ====================================================== #}
+    {# Konfigurationsübersicht                                #}
+    {# ====================================================== #}
+
+    <div class="card bg-black border-secondary text-light mb-4">
+        <div class="card-body small">
+
+            <div class="row">
+                <div class="col-md-6">
+                    <strong>Max Chunks:</strong>
+                    {{ config.retrievalMaxChunks }}
+                </div>
+
+                <div class="col-md-6">
+                    <strong>Vector Top K:</strong>
+                    {{ config.retrievalVectorTopK }}
+                </div>
+            </div>
+
+        </div>
+    </div>
+
+    {# ====================================================== #}
+    {# Testformular                                           #}
+    {# ====================================================== #}
+
+    <div class="card bg-black border-secondary text-light mb-4">
+        <div class="card-body">
+
+            <form method="post">
+
+                <div class="mb-3">
+                    <label class="form-label">Test-Prompt</label>
+                    <textarea name="prompt"
+                              rows="3"
+                              class="form-control bg-dark text-light border-secondary"
+                              required>{{ prompt }}</textarea>
+                </div>
+
+                <button type="submit"
+                        class="btn btn-outline-warning">
+                    Retrieval testen
+                </button>
+
+            </form>
+
+        </div>
+    </div>
+
+    {# ====================================================== #}
+    {# Ergebnisse                                             #}
+    {# ====================================================== #}
+
+    {% if results is not empty %}
+        <div class="card bg-black border-secondary text-light">
+            <div class="card-body">
+
+                <h5 class="text-warning mb-3">
+                    Gefundene Chunks ({{ results|length }})
+                </h5>
+
+                <div style="max-height: 500px; overflow-y: auto;">
+
+                    {% for chunk in results %}
+                        <div class="border border-secondary p-3 mb-3 small">
+                            {{ chunk }}
+                        </div>
+                    {% endfor %}
+
+                </div>
+
+            </div>
+        </div>
+    {% endif %}
+
+{% endblock %}