diff --git a/ai_docs/anthropic_openai_compat.md b/ai_docs/anthropic_openai_compat.md index 8119044..1fbbc07 100644 --- a/ai_docs/anthropic_openai_compat.md +++ b/ai_docs/anthropic_openai_compat.md @@ -22,6 +22,8 @@ print(response.choices[0].message.content) ``` ## Available Models +- claude-opus-4-20250514 +- claude-sonnet-4-20250514 - claude-3-opus-20240229 - claude-3-sonnet-20240229 - claude-3-haiku-20240307 @@ -34,6 +36,8 @@ print(response.choices[0].message.content) 3. Audio input not supported 4. Prompt caching not supported via OpenAI SDK (use native Anthropic SDK) 5. Rate limits follow Anthropic's standard limits +6. Extended context windows (200K tokens) for long-horizon tasks +7. Enhanced reasoning capabilities for complex problem-solving ## Supported Features - ✅ Chat completions @@ -69,4 +73,19 @@ response = client.chat.completions.create( ## Important Notes - This compatibility layer is for testing and comparison, not production - For full features (PDFs, citations, prompt caching), use native Anthropic SDK -- Most unsupported fields are silently ignored rather than producing errors \ No newline at end of file +- Most unsupported fields are silently ignored rather than producing errors +- Claude 4 models (Opus and Sonnet) support extended context windows and enhanced reasoning capabilities +- For the latest model availability and features, refer to Anthropic's official API documentation + +## Claude 4 Latest Capabilities (August 2025) + +### Enhanced Reasoning +- **Long-Horizon Tasks**: Can execute complex, multi-step reasoning over extended periods +- **Coding Proficiency**: Optimized for software development and code analysis +- **Data Analysis**: Advanced capabilities for analyzing large datasets and complex information + +### Use Cases +- **Agentic Workflows**: Configurable reasoning efforts for different complexity levels +- **Extended Context**: 200K token context windows for comprehensive document analysis +- **Complex Problem-Solving**: Multi-step reasoning for engineering and research tasks +- **Code Generation**: Advanced coding assistance with understanding of large codebases \ No newline at end of file diff --git a/ai_docs/gpt_oss_ollama_guide.md b/ai_docs/gpt_oss_ollama_guide.md index e86a994..f74f358 100644 --- a/ai_docs/gpt_oss_ollama_guide.md +++ b/ai_docs/gpt_oss_ollama_guide.md @@ -4,17 +4,43 @@ Source: [OpenAI Cookbook](https://cookbook.openai.com/articles/gpt-oss/run-local ## Model Options -### gpt-oss-20b -- Smaller model -- Best with ≥16GB VRAM or unified memory -- Perfect for higher-end consumer GPUs or Apple Silicon Macs - -### gpt-oss-120b -- Full-sized model -- Best with ≥60GB VRAM or unified memory -- Ideal for multi-GPU or workstation setups - -**Note:** Models ship MXFP4 quantized out of the box. CPU offloading is possible but slower. +### gpt-oss:20b +- **Release Date**: August 5, 2025 +- **License**: Apache 2.0 +- **Smaller model** optimized for consumer hardware +- **Best with ≥16GB VRAM or unified memory** +- **Perfect for higher-end consumer GPUs or Apple Silicon Macs** +- **Context Length**: 131K tokens +- **Hardware Requirements**: 40GB VRAM (FP16), ~10GB VRAM (4-bit quantized) + +### gpt-oss:120b +- **Release Date**: August 5, 2025 +- **License**: Apache 2.0 +- **Full-sized model** for high-performance tasks +- **Best with ≥80GB VRAM or unified memory** +- **Ideal for multi-GPU or workstation setups** +- **Context Length**: 32K tokens +- **Hardware Requirements**: 240GB VRAM (FP16), ~60GB VRAM (4-bit quantized) + +**Note:** Models ship MXFP4 quantized out of the box. CPU offloading is possible but slower. Both models are now available on Amazon Bedrock and AWS SageMaker for cloud deployment. + +## Latest Capabilities (August 2025) + +### Enhanced Reasoning +- **Configurable Reasoning Efforts**: Adjust reasoning depth for different task complexities +- **Agentic Workflows**: Optimized for autonomous AI agents and multi-step reasoning +- **Extended Context**: 131K tokens for gpt-oss:20b, 32K for gpt-oss:120b + +### Use Cases +- **On-Device AI**: Privacy-focused applications with local processing +- **Cost-Efficient Deployment**: Free open-weight models for production use +- **Extended Document Analysis**: Process large documents and codebases +- **Multi-Step Problem Solving**: Complex reasoning tasks requiring sustained attention + +### Cloud Integration +- **Amazon Bedrock**: Native integration for enterprise deployment +- **AWS SageMaker**: Pre-configured endpoints and scaling +- **Hybrid Deployment**: Combine local Ollama with cloud services ## Quick Setup diff --git a/ai_docs/new_openai_gpt_models.md b/ai_docs/new_openai_gpt_models.md index 8d33cc8..bf7f12b 100644 --- a/ai_docs/new_openai_gpt_models.md +++ b/ai_docs/new_openai_gpt_models.md @@ -12,6 +12,7 @@ OpenAI has introduced two groundbreaking model families in August 2025: - **Announcement Date**: August 7, 2025 - **Rollout**: Gradual API access to Plus subscribers - **Full Availability**: Rolling out across platforms +- **Latest Update**: August 11, 2025 - General availability expanding ### Model IDs for API Access @@ -25,6 +26,8 @@ OpenAI has introduced two groundbreaking model families in August 2025: | Chat-optimized | `gpt-5-chat-latest` | Non-reasoning, fast streaming | | **Aliases** | | | | Latest stable | `gpt-5-2025-08-06` | Date-versioned stable release | +| **New Models** | | | +| Deep Research | `gpt-deep-research` | AI agent for extensive web browsing and analysis | ### Model Variants @@ -139,8 +142,9 @@ chat_response = client.chat.completions.create( ### Overview - **Announcement Date**: August 5, 2025 -- **Type**: Open-weight models under MIT license -- **Availability**: Weights available on Hugging Face +- **Type**: Open-weight models under Apache 2.0 license +- **Availability**: Weights available on Hugging Face and AWS +- **Latest Update**: August 11, 2025 - Now available on Amazon Bedrock and SageMaker ### Model Variants @@ -252,6 +256,7 @@ trainer.train() #### Cloud Deployment - **AWS SageMaker**: Pre-configured endpoints for gpt-oss models +- **Amazon Bedrock**: Native integration for gpt-oss-120b and gpt-oss-20b - **Azure AI Foundry**: Native GPT-5 integration via Azure OpenAI Service - **Google Cloud Vertex AI**: Model garden integration for open models - **Replicate**: One-click deployment for gpt-oss variants @@ -262,6 +267,20 @@ trainer.train() - **Core ML**: Apple Silicon optimization for on-device inference - **Mobile SDKs**: Direct integration via OpenAI mobile libraries +### New Models + +#### gpt-deep-research +- **Model ID**: `gpt-deep-research` +- **Release Date**: February 2, 2025 +- **Type**: AI agent leveraging OpenAI's o3 model +- **Capabilities**: + - Extensive web browsing and data analysis + - Multi-step research tasks + - Comprehensive report generation + - Data synthesis and analysis +- **Availability**: Initially ChatGPT Pro subscribers, lightweight version for all users since April 2025 +- **Use Cases**: In-depth research, data analysis, comprehensive reporting + ## Comparison Matrix | Feature | GPT-4o | GPT-5 | GPT-5-mini | GPT-5-nano | gpt-oss-120b | gpt-oss-20b | diff --git a/ai_docs/ollama_openai_compat.md b/ai_docs/ollama_openai_compat.md index bc7e293..a6a1ba9 100644 --- a/ai_docs/ollama_openai_compat.md +++ b/ai_docs/ollama_openai_compat.md @@ -30,10 +30,29 @@ print(response.choices[0].message.content) ## Available Models Any model you have pulled with Ollama can be used. Popular options: + +### Latest State-of-the-Art Models (August 2025) +- **gpt-oss:20b**: OpenAI's open-weight model optimized for 16GB GPUs, supports 131K context length +- **gpt-oss:120b**: OpenAI's larger model requiring 80GB GPU for enhanced capabilities +- **mistral:7b**: Mistral AI's efficient 7B parameter model with strong performance +- **mixtral:8x7b**: Mistral AI's mixture of experts model for robust multi-task performance +- **mistral:large**: Mistral AI's latest large model for advanced reasoning +- **magistral:small**: Mistral AI's first AI reasoning model with chain-of-thought capabilities (open-source) +- **magistral:medium**: Mistral AI's commercial reasoning model for complex problem-solving +- **gemma:2b**: Google's lightweight 2B parameter model for general text generation +- **gemma:7b**: Google's 7B parameter model with enhanced capabilities +- **qwen2:0.5b**: Alibaba's ultra-lightweight model for resource-constrained applications +- **qwen2:7b**: Alibaba's 7B parameter model with improved performance +- **llama3.2:3b**: Meta's latest 3B parameter model optimized for efficiency +- **llama3.2:8b**: Meta's 8B parameter model with enhanced reasoning +- **llama3.2:70b**: Meta's 70B parameter model for complex tasks +- **llama3.3:70b**: Meta's latest 70B model emphasizing cost-efficiency and performance +- **deepseek:r1**: DeepSeek's open-source model with strong coding and mathematics capabilities +- **grok:3**: xAI's real-time conversational AI with up-to-date information access + +### Established Popular Models - llama2 -- mistral - codellama -- mixtral - deepseek-coder - phi - neural-chat @@ -42,6 +61,25 @@ Any model you have pulled with Ollama can be used. Popular options: To pull a model: `ollama pull ` +**Note**: The gpt-oss models are OpenAI's latest open-weight releases, offering enterprise-grade performance on consumer hardware. Use the exact model names with colons (e.g., `gpt-oss:20b`) when pulling models. + +## Latest Model Capabilities (August 2025) + +### Reasoning and Problem-Solving +- **Mistral Magistral Series**: First AI reasoning models with chain-of-thought capabilities for complex problem-solving +- **GPT-OSS Models**: Enhanced reasoning with configurable effort levels for agentic workflows +- **Llama 3.3**: Optimized for mathematical computations and instruction following + +### Specialized Use Cases +- **DeepSeek R1**: Strong performance in coding assistance and mathematical problem-solving +- **Grok 3**: Real-time conversational AI with access to current information +- **Claude 4 Series**: Extended context windows and enhanced reasoning for long-horizon tasks + +### Hardware Optimization +- **gpt-oss:20b**: Optimized for 16GB+ VRAM with 131K context support +- **gpt-oss:120b**: High-performance model requiring 80GB+ VRAM +- **Magistral Small**: Open-source reasoning model for accessible deployment + ## API Endpoint - Base URL: `http://localhost:11434/v1` diff --git a/ai_docs/openai_agent_sdk_core.md b/ai_docs/openai_agent_sdk_core.md index 1edf6d6..a42285d 100644 --- a/ai_docs/openai_agent_sdk_core.md +++ b/ai_docs/openai_agent_sdk_core.md @@ -13,6 +13,25 @@ The OpenAI Agents SDK enables building production-ready agentic AI applications - **Sessions**: Automatic conversation history management - **Streaming**: Real-time progress monitoring +### Latest Model Capabilities (August 2025) + +#### GPT-5 Series for Engineering +- **gpt-5**: PhD-level expertise for complex architectural decisions +- **gpt-5-mini**: 80% performance at 20% cost for production workflows +- **gpt-5-nano**: Ultra-fast responses for CI/CD and edge deployment +- **400K Context**: Process entire codebases in single conversations + +#### GPT-OSS for Local Development +- **gpt-oss:20b**: Consumer hardware optimization (16GB+ VRAM) +- **gpt-oss:120b**: High-performance local deployment (80GB+ VRAM) +- **Apache 2.0 License**: Free for commercial use +- **131K Context**: Extended reasoning for complex engineering tasks + +#### Claude 4 for Research +- **200K Context**: Comprehensive codebase analysis +- **Enhanced Reasoning**: Multi-step problem solving +- **Long-Horizon Tasks**: Extended engineering workflows + ## Installation and Setup ```bash diff --git a/ai_docs/openai_agents_model_configuration.md b/ai_docs/openai_agents_model_configuration.md index 170aa0a..056ec65 100644 --- a/ai_docs/openai_agents_model_configuration.md +++ b/ai_docs/openai_agents_model_configuration.md @@ -17,6 +17,20 @@ The OpenAI Agents SDK supports multiple model providers through different config - Fallback for providers without Responses API support - More widely compatible with OpenAI-compatible endpoints +### 2. Latest Model Families (August 2025) + +#### GPT-5 Series +- **gpt-5**: Flagship reasoning model with PhD-level expertise +- **gpt-5-mini**: 80% of GPT-5 capabilities at 20% cost +- **gpt-5-nano**: Ultra-lightweight for edge deployment +- **gpt-5-chat-latest**: Non-reasoning chat model for fast responses + +#### GPT-OSS Series +- **gpt-oss:20b**: Open-weight model for consumer hardware (16GB+ VRAM) +- **gpt-oss:120b**: High-performance open-weight model (80GB+ VRAM) +- **License**: Apache 2.0 - Free for commercial use +- **Availability**: Local deployment via Ollama, cloud via AWS Bedrock + ## Non-OpenAI Model Integration ### Method 1: LiteLLM Integration (Simplest) @@ -159,14 +173,17 @@ Solution: Use providers that support JSON schema outputs, or handle malformed JS ## Feature Compatibility Matrix -| Feature | OpenAI | Anthropic (via LiteLLM) | Ollama | -|---------|--------|-------------------------|---------| -| Responses API | ✅ | ❌ | ❌ | -| Chat Completions | ✅ | ✅ | ✅ | -| Structured Outputs | ✅ | ⚠️ | ❌ | -| Multimodal | ✅ | ✅ | ⚠️ | -| Function Calling | ✅ | ✅ | ❌ | -| File/Web Search | ✅ | ❌ | ❌ | +| Feature | OpenAI | Anthropic (via LiteLLM) | Ollama | GPT-OSS | +|---------|--------|-------------------------|---------|---------| +| Responses API | ✅ | ❌ | ❌ | ❌ | +| Chat Completions | ✅ | ✅ | ✅ | ✅ | +| Structured Outputs | ✅ | ⚠️ | ❌ | ⚠️ | +| Multimodal | ✅ | ✅ | ⚠️ | ❌ | +| Function Calling | ✅ | ✅ | ❌ | ⚠️ | +| File/Web Search | ✅ | ❌ | ❌ | ❌ | +| Extended Context | ✅ (400K) | ✅ (200K) | ✅ (131K) | ✅ (131K) | +| Reasoning Effort | ✅ (GPT-5) | ✅ (Claude 4) | ❌ | ⚠️ | +| Local Deployment | ❌ | ❌ | ✅ | ✅ | ## Key Implementation Notes @@ -215,4 +232,72 @@ ollama_agent = Agent( openai_client=ollama_client ) ) +``` + +## Integration with Latest Models + +### GPT-5 Reasoning Models + +```python +# Using GPT-5 with reasoning capabilities +reasoning_agent = Agent( + name="Research Agent", + model="gpt-5", # or "gpt-5-mini" for cost optimization + model_settings=ModelSettings( + temperature=0.1, + extra_args={ + "reasoning_effort": "high", # GPT-5 specific parameter + "max_tokens": 128000, # Up to 128K output tokens + } + ) +) + +# For edge deployment with minimal reasoning +edge_agent = Agent( + name="Edge Agent", + model="gpt-5-nano", + model_settings=ModelSettings( + extra_args={ + "reasoning_effort": "minimal", # Fastest, cheapest + } + ) +) +``` + +### GPT-OSS Local Models + +```python +# Local GPT-OSS deployment via Ollama +from openai import AsyncOpenAI + +ollama_client = AsyncOpenAI( + base_url="http://localhost:11434/v1", + api_key="ollama" +) + +local_agent = Agent( + name="Local Agent", + model=OpenAIChatCompletionsModel( + model="gpt-oss:20b", # Use exact Ollama model name + openai_client=ollama_client + ), + model_settings=ModelSettings( + temperature=0.2, + max_tokens=131000, # Full context window + ) +) +``` + +### Claude 4 Models + +```python +# Claude 4 with extended context +claude_agent = Agent( + name="Claude Agent", + model="litellm/anthropic/claude-opus-4-20250514", + model_settings=ModelSettings( + temperature=0.1, + max_tokens=200000, # Claude 4's 200K context + ) +) ``` \ No newline at end of file