feat(knowledge): support PDF and PPTX uploads for knowledge base and …

.env.example

@@ -150,6 +150,15 @@ AI_QUALITY_OPTIMIZATION_QUEUE_REPLICAS=2
 GEOFLOW_DEBUG_KNOWLEDGE_QUERY_EMBEDDING=false
 # 私网出站仅允许精确 host:port（逗号分隔）；留空时全部拒绝，不支持通配符或路径。
 GEOFLOW_OUTBOUND_PRIVATE_TARGETS=
+# 出站 HTTP 代理：仅当 GEOFLOW_OUTBOUND_PROXY 与 GEOFLOW_OUTBOUND_PROXY_HOSTS 同时配置才生效。
+# 留空时所有 GEOFlow 出站保持直连（默认行为，AI/搜索/更新器走各自原有路径）。
+# 典型用法：需要绕过目标站点 WAF（jshh.com、Cloudflare 高防）时为特定 host 启用住宅代理。
+# 示例住宅代理 URL：  http://user:pass@resi.example.com:8231
+# 示例 SOCKS5：       socks5://user:pass@resi.example.com:1080
+# host 支持精确（jshh.com）、后缀（*.jshh.com）、通配（* 或 . 三选一匹配所有公网 host）。
+# 私网/内网目标（private_targets）永远走直连，不受代理配置影响。
+GEOFLOW_OUTBOUND_PROXY=
+GEOFLOW_OUTBOUND_PROXY_HOSTS=
 GEOFLOW_OUTBOUND_JSON_MAX_BYTES=4194304
 GEOFLOW_OUTBOUND_AI_MAX_BYTES=8388608
 GEOFLOW_OUTBOUND_IMPORT_MAX_BYTES=5242880

.gitignore

@@ -30,3 +30,18 @@ __pycache__/
 Homestead.json
 Homestead.yaml
 Thumbs.db
+
+# AI tool caches (per-machine)
+.agents/
+.boost/
+.claude/
+.codex/
+.cursor/
+.gemini/
+.bak/
+.dockerignore.local
+
+_route_refs.txt
+
+.untracked-stash/
+

app/Data/Admin/AdminAiModelTestSnapshot.php

@@ -27,6 +27,7 @@ public function __construct(
         public readonly ?int $maxTokens,
         public readonly bool $gemini,
         public readonly bool $usesOpenAiResponses,
+        public readonly bool $volcengineMultimodal,
         public readonly bool $preparedAsSuperAdmin,
         public readonly ?AiUsageReservation $reservation,
         private readonly string $encryptedApiKey,

app/Http/Controllers/Admin/AiModelController.php

@@ -399,6 +399,7 @@ public function testConnection(Request $request, int $modelId): JsonResponse
             $modelName = trim($snapshot->providerModelId);
             $isGemini = $snapshot->gemini;
             $usesOpenAiResponses = $snapshot->usesOpenAiResponses;
+            $isVolcengineMultimodal = $snapshot->volcengineMultimodal;
 
             if ($endpoint === '') {
                 return $this->modelTestResponse(
@@ -515,7 +516,7 @@ public function testConnection(Request $request, int $modelId): JsonResponse
                 ? $request->withHeaders(['x-goog-api-key' => $apiKey])
                 : $request->withToken($apiKey);
 
-            $testPayload = $this->buildTestPayload($modelName, $modelType, $isGemini, $usesOpenAiResponses);
+            $testPayload = $this->buildTestPayload($modelName, $modelType, $isGemini, $usesOpenAiResponses, $isVolcengineMultimodal);
             $response = $this->safeHttp->post(
                 $request,
                 $endpoint,
@@ -555,7 +556,7 @@ function () use (&$outboundAttempted, $usageSession, $model, $testPayload): void
                 );
             }
 
-            if (! $this->isValidTestResponse($json, $modelType, $isGemini, $usesOpenAiResponses)) {
+            if (! $this->isValidTestResponse($json, $modelType, $isGemini, $usesOpenAiResponses, $isVolcengineMultimodal)) {
                 $usageSession->discarded('direct.p1', 'ai_provider_response_invalid', $json['usage'] ?? null);
                 if ($reservation instanceof AiUsageReservation) {
                     $this->recordModelTestAttempt($reservation);
@@ -961,7 +962,8 @@ private function buildTestPayload(
         string $modelName,
         string $modelType,
         bool $isGemini = false,
-        bool $usesOpenAiResponses = false
+        bool $usesOpenAiResponses = false,
+        bool $isVolcengineMultimodal = false
     ): array {
         if ($isGemini) {
             if ($modelType === 'embedding') {
@@ -1005,6 +1007,15 @@ private function buildTestPayload(
             ];
         }
 
+        if ($modelType === 'embedding' && $isVolcengineMultimodal) {
+            return [
+                'model' => $modelName,
+                'input' => [
+                    ['type' => 'text', 'text' => 'GEOFlow embedding connection test'],
+                ],
+            ];
+        }
+
         if ($modelType === 'embedding') {
             return [
                 'model' => $modelName,
@@ -1034,7 +1045,8 @@ private function isValidTestResponse(
         mixed $json,
         string $modelType,
         bool $isGemini = false,
-        bool $usesOpenAiResponses = false
+        bool $usesOpenAiResponses = false,
+        bool $isVolcengineMultimodal = false
     ): bool {
         if (! is_array($json)) {
             return false;

app/Http/Controllers/Admin/AiVisibilityAnalyticsController.php

@@ -179,13 +179,29 @@ private function keywordLibraries(): Collection
         $keywords = Keyword::query()->where('keyword', '!=', '')->orderBy('id')->limit(1000)
             ->get(['id', 'library_id', 'keyword'])->groupBy('library_id');
 
+        $recentlySampled = Schema::hasTable('ai_visibility_runs')
+            ? AiVisibilityRun::query()
+                ->whereIn('provider_type', AiVisibilityRun::SAMPLE_PROVIDERS)
+                ->whereNotNull('keyword')
+                ->where('keyword', '!=', '')
+                ->where('created_at', '>=', now()->subDays(7))
+                ->distinct()
+                ->pluck('keyword')
+                ->mapWithKeys(static fn (string $keyword): array => [$keyword => true])
+                ->all()
+            : [];
+
         return KeywordLibrary::query()->whereIn('id', $keywords->keys())->orderBy('id')
             ->get(['id', 'name'])
             ->map(static fn (KeywordLibrary $library): array => [
                 'id' => $library->id,
                 'name' => $library->name,
                 'keywords' => $keywords->get($library->id, collect())
-                    ->map(static fn (Keyword $keyword): array => ['id' => $keyword->id, 'keyword' => (string) $keyword->keyword])
+                    ->map(static fn (Keyword $keyword): array => [
+                        'id' => $keyword->id,
+                        'keyword' => (string) $keyword->keyword,
+                        'recently_sampled' => isset($recentlySampled[(string) $keyword->keyword]),
+                    ])
                     ->values()->all(),
             ])->values();
     }

app/Http/Controllers/Admin/EnterpriseKnowledgeController.php

@@ -67,7 +67,7 @@ public function store(Request $request): RedirectResponse
             'description' => ['nullable', 'string', 'max:1000'],
             'content' => ['nullable', 'string'],
             'enterprise_files' => ['nullable', 'array', 'max:10'],
-            'enterprise_files.*' => ['file', File::types(['txt', 'md', 'markdown', 'docx'])->max(8 * 1024)],
+            'enterprise_files.*' => ['file', File::types(['txt', 'md', 'markdown', 'docx', 'pdf', 'ppt', 'pptx'])->max(8 * 1024)],
         ], [
             'enterprise_files.max' => __('admin.enterprise_knowledge.error.files_limit'),
         ]);

app/Http/Controllers/Admin/KnowledgeBaseController.php

@@ -566,9 +566,9 @@ private function validateKnowledgeImportForm(Request $request): array
             'risk_level' => ['nullable', 'in:low,medium,high'],
             'review_status' => ['nullable', 'in:unreviewed,reviewed'],
             'import_action' => ['nullable', 'in:save,save_and_chunk'],
-            'knowledge_file' => ['nullable', File::types(['txt', 'md', 'docx'])->max(8 * 1024)],
+            'knowledge_file' => ['nullable', File::types(['txt', 'md', 'docx', 'pdf', 'ppt', 'pptx'])->max(8 * 1024)],
             'knowledge_files' => ['nullable', 'array', 'max:10'],
-            'knowledge_files.*' => ['file', File::types(['txt', 'md', 'docx'])->max(8 * 1024)],
+            'knowledge_files.*' => ['file', File::types(['txt', 'md', 'docx', 'pdf', 'ppt', 'pptx'])->max(8 * 1024)],
         ], [
             'knowledge_file.mimes' => __('admin.knowledge_bases.error.file_type_invalid'),
             'knowledge_file.max' => __('admin.knowledge_bases.error.file_too_large'),
@@ -1106,6 +1106,34 @@ private function parseUploadedKnowledgeFile(string $absolutePath, string $origin
             ];
         }
 
+        if ($extension === 'pdf') {
+            $content = app(\App\Services\GeoFlow\KnowledgeSourceParser::class)->extractPdfContent($absolutePath);
+            if ($content === '') {
+                throw new \RuntimeException(__('admin.knowledge_bases.error.file_type_invalid'));
+            }
+
+            return [
+                'content' => $content,
+                'file_type' => 'pdf',
+            ];
+        }
+
+        if ($extension === 'pptx') {
+            $content = app(\App\Services\GeoFlow\KnowledgeSourceParser::class)->extractPptxContent($absolutePath);
+            if ($content === '') {
+                throw new \RuntimeException(__('admin.knowledge_bases.error.file_type_invalid'));
+            }
+
+            return [
+                'content' => $content,
+                'file_type' => 'presentation',
+            ];
+        }
+
+        if ($extension === 'ppt') {
+            throw new \RuntimeException(__('admin.knowledge_bases.error.ppt_legacy_not_supported'));
+        }
+
         throw new \RuntimeException(__('admin.knowledge_bases.error.file_type_invalid'));
     }
 

app/Providers/AppServiceProvider.php

@@ -31,6 +31,7 @@
 use App\Services\GeoFlow\TaskMonitoringQueryService;
 use App\Services\Outbound\FinalOutboundSecurityPolicy;
 use App\Services\Outbound\LaravelPinnedOutboundTransport;
+use App\Services\Outbound\OutboundProxyPolicy;
 use App\Services\Outbound\SafeOutboundHttpClient;
 use App\Services\Outbound\SecureHttpFactory;
 use App\Services\Outbound\SystemHostResolver;
@@ -71,6 +72,12 @@ public function register(): void
         $this->app->bind(AgentClient::class, UnixSocketAgentClient::class);
         $this->app->bind(AiModelWriteLock::class, DatabaseAiModelWriteLock::class);
         $this->app->singleton(FinalOutboundSecurityPolicy::class);
+        $this->app->singleton(OutboundProxyPolicy::class, function (): OutboundProxyPolicy {
+            return OutboundProxyPolicy::fromConfig(
+                config('geoflow.outbound_proxy_url'),
+                (string) config('geoflow.outbound_proxy_hosts', ''),
+            );
+        });
         $this->app->bind(OutboundTransport::class, function () use ($fixedContextCapability): LaravelPinnedOutboundTransport {
             return new LaravelPinnedOutboundTransport($fixedContextCapability);
         });

app/Services/Admin/AdminAiModelTestPreparationService.php

@@ -57,6 +57,11 @@ public function prepare(Admin $authenticatedActor, int $modelId): AdminAiModelTe
                 providerModelId: (string) $lockedModel->model_id,
                 maxTokens: $lockedModel->max_tokens === null ? null : (int) $lockedModel->max_tokens,
                 gemini: OpenAiRuntimeProvider::isGeminiProviderUrl($endpoint),
+                volcengineMultimodal: $modelType === 'embedding'
+                    && OpenAiRuntimeProvider::isVolcengineMultimodalEmbedding(
+                        (string) $lockedModel->api_url,
+                        (string) $lockedModel->model_id,
+                    ),
                 usesOpenAiResponses: $modelType === 'chat'
                     && OpenAiRuntimeProvider::resolveChatDriver(
                         (string) $lockedModel->api_url,
@@ -114,6 +119,7 @@ public function prepareSystemBinding(
                 providerModelId: (string) $lockedModel->model_id,
                 maxTokens: $lockedModel->max_tokens === null ? null : (int) $lockedModel->max_tokens,
                 gemini: false,
+                volcengineMultimodal: false,
                 usesOpenAiResponses: $bindingType === 'ark',
                 preparedAsSuperAdmin: true,
                 reservation: $reservation,
@@ -342,6 +348,11 @@ private function resolveEndpoint(AiModel $model, string $modelType): string
             return rtrim($baseUrl, '/').'/responses';
         }
 
+        if ($modelType === 'embedding'
+            && OpenAiRuntimeProvider::isVolcengineMultimodalEmbedding((string) $model->api_url, (string) $model->model_id)) {
+            return rtrim($baseUrl, '/').OpenAiRuntimeProvider::volcengineMultimodalEmbeddingPath();
+        }
+
         return rtrim($baseUrl, '/').($modelType === 'embedding' ? '/embeddings' : '/chat/completions');
     }
 }

app/Services/GeoFlow/KnowledgeChunkSyncService.php

@@ -2368,6 +2368,10 @@ private function requestEmbeddingVectors(
             );
         }
 
+        if ($this->isVolcengineMultimodalEmbeddingMetadata($embeddingMetadata)) {
+            return $this->requestVolcengineMultimodalEmbeddings($inputs, $embeddingMetadata);
+        }
+
         return $this->requestOpenAiCompatibleEmbeddings($inputs, $embeddingMetadata);
     }
 
@@ -2435,6 +2439,79 @@ private function requestOpenAiCompatibleEmbeddings(
         );
     }
 
+    /**
+     * 直连火山方舟多模态 embedding endpoint。
+     *
+     * 该 endpoint（/embeddings/multimodal-embedding-v1）与通用 /embeddings 不兼容，
+     * 请求/响应采用 OpenAI 兼容结构但 input 必须是 [{"type":"text","text":...}] 形式。
+     *
+     * 请求通过统一安全出站网关校验并固定目标地址。
+     *
+     * @param  list<string>  $inputs
+     * @param  array{model_id:int,model_name:string,provider:string,api_url:string,api_key:string,driver:string}  $embeddingMetadata
+     */
+    private function requestVolcengineMultimodalEmbeddings(
+        array $inputs,
+        array $embeddingMetadata,
+    ): KnowledgeEmbeddingProviderResult {
+        $endpoint = rtrim((string) $embeddingMetadata['api_url'], '/')
+            .OpenAiRuntimeProvider::volcengineMultimodalEmbeddingPath();
+
+        $payload = [
+            'model' => (string) $embeddingMetadata['model_name'],
+            'input' => array_map(
+                static fn (string $text): array => ['type' => 'text', 'text' => $text],
+                $inputs,
+            ),
+        ];
+
+        $request = $this->http->acceptJson()
+            ->asJson()
+            ->withToken((string) $embeddingMetadata['api_key'])
+            ->connectTimeout(8)
+            ->timeout(45);
+        $response = $this->safeHttp->post($request, $endpoint, $payload, (int) config('geoflow.outbound_ai_max_bytes', 8 * 1024 * 1024));
+
+        if (! $response->successful()) {
+            $error = data_get($response->json(), 'error.message');
+            $message = is_string($error) && $this->isEmbeddingBatchSizeError($error)
+                ? 'Embedding provider rejected batch size.'
+                : 'Embedding provider request failed.';
+
+            throw new \RuntimeException(sprintf(
+                'HTTP request returned status code %d: %s',
+                $response->status(),
+                $message,
+            ), 0, $response->toException());
+        }
+
+        $data = $response->json();
+        $rows = is_array($data) ? ($data['data'] ?? []) : [];
+        if (! is_array($rows)) {
+            return new KnowledgeEmbeddingProviderResult([], is_array($data) ? ($data['usage'] ?? null) : null);
+        }
+
+        $embeddings = [];
+        foreach ($rows as $position => $row) {
+            if (! is_array($row)) {
+                continue;
+            }
+
+            $index = $position;
+            if (array_key_exists('index', $row) && is_numeric($row['index'])) {
+                $index = max(0, (int) $row['index']);
+            }
+
+            $embeddings[$index] = $row['embedding'] ?? null;
+        }
+        ksort($embeddings);
+
+        return new KnowledgeEmbeddingProviderResult(
+            $embeddings,
+            is_array($data) ? ($data['usage'] ?? null) : null,
+        );
+    }
+
     private function embeddingBatchSize(): int
     {
         return max(1, min(64, (int) config('geoflow.embedding_batch_size', 1)));
@@ -2489,6 +2566,19 @@ private function isGeminiEmbeddingMetadata(array $embeddingMetadata): bool
             || OpenAiRuntimeProvider::isGeminiProviderUrl((string) ($embeddingMetadata['api_url'] ?? ''));
     }
 
+    /**
+     * @param  array<string, mixed>  $embeddingMetadata
+     */
+    private function isVolcengineMultimodalEmbeddingMetadata(array $embeddingMetadata): bool
+    {
+        if ((string) ($embeddingMetadata['driver'] ?? '') === 'volcengine-multimodal') {
+            return true;
+        }
+
+        return OpenAiRuntimeProvider::isVolcengineProviderUrl((string) ($embeddingMetadata['api_url'] ?? ''))
+            && OpenAiRuntimeProvider::isVolcengineMultimodalEmbeddingModel((string) ($embeddingMetadata['model_name'] ?? ''));
+    }
+
     private function normalizeGeminiEmbeddingSegment(string $value): string
     {
         return trim(preg_replace('/\s+/u', ' ', $value) ?: $value);