hack/tools/scripts/llm-key-hunter/results/_cache/content/ayanzadeh93_benchmark-accessibility.json

{"repo": "ayanzadeh93/benchmark-accessibility", "files": {"vlm_factory.py": {"501a3536d9f6129392f74e262f34d3201d05775b": {"t": 1785674522.176054, "n": 7004, "h": "3d17ba055b46ec6d", "c": "#!/usr/bin/env python3\n\"\"\"\nVLM Factory - Easy switching between GPT-4o and Qwen3-VL\n\"\"\"\n\nimport logging\nfrom typing import Optional\nfrom vlm_base import BaseVLMExtractor\n\nlogging.basicConfig(level=logging.INFO)\nlogger = logging.getLogger(__name__)\n\n\nclass VLMFactory:\n    \"\"\"Factory for creating VLM extractors.\"\"\"\n    \n    @staticmethod\n    def create_extractor(\n        model_type: str = \"qwen\",\n        api_key: Optional[str] = None,\n        device: str = \"auto\"\n    ) -> BaseVLMExtractor:\n        \"\"\"\n        Create VLM extractor.\n        \n        Args:\n            model_type: \"qwen\", \"florence2\", \"gpt4o\", \"gpt5nano\"\n            api_key: API key for GPT models\n            device: \"auto\", \"cuda\", or \"cpu\"\n            \n        Returns:\n            VLM extractor instance\n            \n        Examples:\n            # Use Florence-2 (FASTEST, free, local GPU)\n            extractor = VLMFactory.create_extractor(\"florence2\", device=\"cuda\")\n            \n            # Use Qwen (free, local GPU)\n            extractor = VLMFactory.create_extractor(\"qwen\", device=\"cuda\")\n            \n            # Use GPT-4o (API)\n            extractor = VLMFactory.create_extractor(\"gpt4o\", api_key=\"sk-...\")\n        \"\"\"\n        model_type = model_type.lower().strip()\n        \n        if model_type in [\"florence2\", \"florence\", \"florence-2\"]:\n            logger.info(\"Creating Florence-2 extractor (FASTEST, GPU-accelerated)\")\n            from vlm_florence2 import Florence2Extractor\n            return Florence2Extractor(\n                device=device,\n                model_size=\"base\"  # base is faster than large\n            )\n        \n        elif model_type in [\"qwen\", \"qwen3\", \"qwen3vl\", \"qwen-vl\"]:\n            logger.info(\"Creating Qwen3-VL extractor (GPU-accelerated)\")\n            from vlm_qwen import Qwen3VLExtractor\n            return Qwen3VLExtractor(\n                device=device\n            )\n        \n        elif model_type in [\"llava\", \"llava15\", \"llava-1.5\", \"llava1.5\"]:\n            logger.info(\"Creating LLaVA-1.5-7B extractor (GPU-accelerated)\")\n            from vlm_llava import LLaVAExtractor\n            return LLaVAExtractor(\n                device=device\n            )\n        \n        elif model_type in [\"gpt4o\", \"gpt-4o\", \"openai\"]:\n            logger.info(\"Creating GPT-4o extractor (API)\")\n            from vlm_gpt4o import GPT4oExtractor\n            return GPT4oExtractor(api_key=api_key, model_name=\"gpt-4o\")\n        \n        elif model_type in [\"gpt5nano\", \"gpt-5-nano\", \"gpt5-nano\", \"gptnano\"]:\n            logger.info(\"Creating GPT-5 Nano extractor (API)\")\n            from vlm_gpt4o import GPT4oExtractor\n            return GPT4oExtractor(api_key=api_key, model_name=\"gpt-5-nano-2025-08-07\")\n        \n        elif model_type in [\"gpt5mini\", \"gpt-5-mini\", \"gpt5-mini\", \"gptmini\"]:\n            logger.info(\"Creating GPT-5 Mini extractor (API, CHEAPEST)\")\n            from vlm_gpt4o import GPT4oExtractor\n            return GPT4oExtractor(api_key=api_key, model_name=\"gpt-5-mini-2025-08-07\")\n\n        elif model_type in [\n            \"openrouter_trinity\",\n            \"openrouter_llama32_11b_vision\",\n            \"openrouter_llama4_maverick\",\n            \"openrouter_molmo_8b\",\n            \"openrouter_ministral_3b\",\n            \"openrouter_gpt_oss_safeguard_20b\",\n            \"openrouter_qwen3_vl_235b\",\n            \"openrouter_qwen3_vl_8b\",\n            \"openrouter_qwen_vl_plus\",\n            \"qwen8b\",  # Alias for openrouter_qwen3_vl_8b\n            \"openrouter\",\n        ]:\n            from vlm_openrouter import OpenRouterExtractor\n\n            # Map friendly names to OpenRouter model ids\n            model_map = {\n                \"openrouter_trinity\": \"arcee-ai/trinity-large-preview:free\",\n                \"openrouter_llama32_11b_vision\": \"meta-llama/llama-3.2-11b-vision-instruct\",\n                \"openrouter_llama4_maverick\": \"meta-llama/llama-4-maverick\",\n                \"openrouter_molmo_8b\": \"allenai/molmo-2-8b:free\",\n                \"openrouter_ministral_3b\": \"mistralai/ministral-3b-2512\",\n                \"openrouter_gpt_oss_safeguard_20b\": \"openai/gpt-oss-safeguard-20b\",\n                \"openrouter_qwen3_vl_235b\": \"qwen/qwen3-vl-235b-a22b-instruct\",\n                \"openrouter_qwen3_vl_8b\": \"qwen/qwen3-vl-8b-instruct\",\n                \"qwen8b\": \"qwen/qwen3-vl-8b-instruct\",\n                \"openrouter_qwen_vl_plus\": \"qwen/qwen-vl-plus\",\n            }\n            # qwen8b -> use openrouter_qwen3_vl_8b model id\n            model_id = model_map.get(model_type, model_type)\n            logger.info(f\"Creating OpenRouter extractor (model: {model_id})\")\n            return OpenRouterExtractor(api_key=api_key, model_id=model_id)\n        \n        else:\n            raise ValueError(f\"Unknown model type: {model_type}. Use 'florence2', 'qwen', 'llava', 'gpt4o', 'gpt5nano', or 'gpt5mini'\")\n    \n    @staticmethod\n    def list_available_models():\n        \"\"\"List available VLM models.\"\"\"\n        print(\"\\n\" + \"=\"*60)\n        print(\"Available VLM Models\")\n        print(\"=\"*60)\n        \n        print(\"\\n1. Florence-2 (florence2) ⭐ FASTEST\")\n        print(\"   - Local GPU (FREE)\")\n        print(\"   - Speed: 1-2 seconds per image on RTX GPU\")\n        print(\"   - Size: 230M params (tiny!)\")\n        print(\"   - Optimized for: Object detection, grounding, captioning\")\n        \n        print(\"\\n2. LLaVA-1.5-7B (llava)\")\n        print(\"   - Local GPU (FREE)\")\n        print(\"   - Speed: 3-5 seconds per image on RTX GPU\")\n        print(\"   - Size: 7B params\")\n        print(\"   - Optimized for: VQA, detailed image understanding\")\n        print(\"   - Usage: VLMFactory.create_extractor('florence2', device='cuda')\")\n        \n        print(\"\\n2. Qwen3-VL-2B (qwen)\")\n        print(\"   - Local GPU (FREE)\")\n        print(\"   - Speed: 2-3 seconds per image\")\n        print(\"   - Size: 2B params\")\n        print(\"   - Usage: VLMFactory.create_extractor('qwen', device='cuda')\")\n\n        print(\"\\n3. GPT-4o (gpt4o)\")\n        print(\"   - API-based (paid)\")\n        print(\"   - Requires: OpenAI API key\")\n        print(\"   - Cost: ~$0.005 per image\")\n        print(\"   - Usage: VLMFactory.create_extractor('gpt4o', api_key='sk-...')\")\n        \n        print(\"\\n4. GPT-5 Nano (gpt5nano)\")\n        print(\"   - API-based (paid, cheaper than GPT-4o)\")\n        print(\"   - Requires: OpenAI API key\")\n        print(\"   - Cost: ~$0.00005 per image (input), ~$0.0004 per image (output)\")\n        print(\"   - Usage: VLMFactory.create_extractor('gpt5nano', api_key='sk-...')\")\n        \n        print(\"\\n\" + \"=\"*60)\n\n\n# Convenience functions\ndef create_qwen_extractor(device: str = \"auto\") -> BaseVLMExtractor:\n    \"\"\"Create Qwen3-VL extractor (shortcut).\"\"\"\n    return VLMFactory.create_extractor(\"qwen\", device=device)\n\n\ndef create_gpt4o_extractor(api_key: Optional[str] = None) -> BaseVLMExtractor:\n    \"\"\"Create GPT-4o extractor (shortcut).\"\"\"\n    return VLMFactory.create_extractor(\"gpt4o\", api_key=api_key)\n\n\nif __name__ == \"__main__\":\n    VLMFactory.list_available_models()"}}}}