@miller-tech/uap 1.9.1 → 1.10.1

package/config/chat_template.jinja

@@ -0,0 +1,76 @@
+{#- Qwen3.5 Chat Template - Optimized for 35B A3B tool calling -#}
+{#- Compatible with llama.cpp / vLLM / OpenAI-compatible servers -#}
+{#- Uses official Qwen3 tool call format for maximum reliability -#}
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content | trim + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+
+{%- for message in messages %}
+    {#- User message or non-first system message -#}
+    {%- if (message.role == 'user') or (message.role == 'system' and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content | trim + '<|im_end|>\n' }}
+
+    {#- Assistant message -#}
+    {%- elif message.role == 'assistant' %}
+        {{- '<|im_start|>assistant' }}
+        {%- if message.content is defined and message.content %}
+            {{- '\n' + message.content | trim }}
+        {%- endif %}
+        {%- if message.tool_calls is defined and message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tc = tool_call.function %}
+                {%- else %}
+                    {%- set tc = tool_call %}
+                {%- endif %}
+                {%- if (loop.first and message.content is defined and message.content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tc.name }}
+                {{- '", "arguments": ' }}
+                {%- if tc.arguments is string %}
+                    {{- tc.arguments }}
+                {%- else %}
+                    {{- tc.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+
+    {#- Tool response (grouped under user role per Qwen3 spec) -#}
+    {%- elif message.role == 'tool' %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != 'tool') %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content | trim }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != 'tool') %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+
+    {%- endif %}
+{%- endfor %}
+
+{#- Generation prompt for assistant turn -#}
+{%- if add_generation_prompt is defined and add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- endif %}
+{%- endif %}

package/config/lora-finetune.yaml

@@ -0,0 +1,82 @@
+# LoRA Fine-Tune Configuration for Qwen3.5 35B A3B Tool Calling
+# Compatible with: axolotl, LLaMA-Factory, unsloth
+#
+# This config trains a LoRA adapter to improve tool call format adherence.
+# The adapter is small (~50MB) and can be merged into the base model or
+# loaded at runtime with llama.cpp --lora flag.
+#
+# Usage:
+#   1. Generate training data:
+#      python3 tools/agents/scripts/generate_lora_training_data.py -n 500
+#   2. Fine-tune with axolotl:
+#      accelerate launch -m axolotl.cli.train config/lora-finetune.yaml
+#   3. Or with unsloth (faster, less VRAM):
+#      unsloth train --config config/lora-finetune.yaml
+
+# Base model
+base_model: Qwen/Qwen3.5-35B-A3B
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+trust_remote_code: true
+
+# LoRA configuration
+adapter: lora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_modules:
+  - q_proj
+  - k_proj
+  - v_proj
+  - o_proj
+  - gate_proj
+  - up_proj
+  - down_proj
+lora_target_linear: true
+
+# Dataset
+datasets:
+  - path: tool_call_training_data.jsonl
+    type: chat_template
+    chat_template: chatml
+    field_messages: messages
+
+# Training parameters
+num_epochs: 3
+micro_batch_size: 1
+gradient_accumulation_steps: 8
+learning_rate: 2.0e-4
+lr_scheduler: cosine
+warmup_ratio: 0.1
+optimizer: adamw_torch
+weight_decay: 0.01
+max_grad_norm: 1.0
+
+# Sequence length
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+# Memory optimization
+bf16: true
+tf32: true
+gradient_checkpointing: true
+flash_attention: true
+
+# Output
+output_dir: ./output/qwen35-tool-call-lora
+save_strategy: epoch
+save_total_limit: 2
+logging_steps: 10
+
+# Evaluation
+val_set_size: 0.05
+eval_steps: 50
+
+# Wandb (optional)
+# wandb_project: qwen35-tool-call-lora
+# wandb_run_id:
+
+# Special tokens
+special_tokens:
+  pad_token: '<|endoftext|>'

package/config/model-profiles/claude-haiku-3.5.json

@@ -0,0 +1,62 @@
+{
+  "_profile": "claude-haiku-3.5",
+  "_description": "Anthropic Claude 3.5 Haiku (claude-3-5-haiku-20241022). Fastest and most cost-effective Claude model. 200K context with strong tool calling for high-throughput tasks. Best for rapid iteration, simple tool use, classification, and cost-sensitive workloads where latency matters.",
+
+  "model": "claude-3-5-haiku-20241022",
+  "provider": "anthropic",
+  "api_base_url": "https://api.anthropic.com",
+  "max_tokens": 8192,
+  "temperature": 1.0,
+  "top_p": 1.0,
+  "timeout_ms": 60000,
+  "context_window": 200000,
+
+  "optimize_for_tool_calls": true,
+  "enable_thinking": false,
+
+  "tool_calling": {
+    "parallel_tool_calls": true,
+    "tool_choice": "auto",
+    "max_tool_calls_per_turn": 128,
+    "_comment": "Haiku 3.5 supports parallel tool use with very low latency. Great for high-volume tool calling."
+  },
+
+  "extended_thinking": {
+    "enabled": false,
+    "_comment": "Extended thinking not available on Haiku 3.5. Use Sonnet 4.6 or Opus 4.6 for thinking mode."
+  },
+
+  "dynamic_temperature": {
+    "enabled": false,
+    "_comment": "Anthropic API handles temperature differently. Default temp=1.0 is recommended."
+  },
+
+  "tool_call_batching": {
+    "enabled": true,
+    "system_prompt_suffix": "When multiple tools are needed, call ALL of them in a single response. Do not call one tool and wait - emit all tool calls together.",
+    "_comment": "Haiku 3.5 supports parallel tool calls with sub-second latency."
+  },
+
+  "pricing": {
+    "input_per_1m": 0.8,
+    "output_per_1m": 4.0,
+    "cache_write_per_1m": 1.0,
+    "cache_read_per_1m": 0.08,
+    "currency": "USD",
+    "_comment": "As of 2025. Extremely cost-effective. ~19x cheaper than Opus 4.6 on input."
+  },
+
+  "rate_limits": {
+    "requests_per_minute": 4000,
+    "tokens_per_minute": 400000,
+    "_comment": "Highest throughput of all Claude models. Designed for high-volume use cases."
+  },
+
+  "running_config": {
+    "description": "Anthropic hosted API. Requires ANTHROPIC_API_KEY environment variable.",
+    "env_vars": {
+      "ANTHROPIC_API_KEY": "<your-api-key>",
+      "UAP_MODEL_PROFILE": "claude-haiku-3.5"
+    }
+  }
+}

package/config/model-profiles/claude-opus-4.6.json

@@ -0,0 +1,63 @@
+{
+  "_profile": "claude-opus-4.6",
+  "_description": "Anthropic Claude Opus 4.6 (claude-opus-4-6-20250616). Most capable Claude model with extended thinking, parallel tool use, and 200K context. Best for complex multi-step reasoning, large codebase navigation, and agentic workflows where accuracy matters more than speed.",
+
+  "model": "claude-opus-4-6-20250616",
+  "provider": "anthropic",
+  "api_base_url": "https://api.anthropic.com",
+  "max_tokens": 32768,
+  "temperature": 1.0,
+  "top_p": 1.0,
+  "timeout_ms": 300000,
+  "context_window": 200000,
+
+  "optimize_for_tool_calls": true,
+  "enable_thinking": true,
+
+  "tool_calling": {
+    "parallel_tool_calls": true,
+    "tool_choice": "auto",
+    "max_tool_calls_per_turn": 128,
+    "_comment": "Opus 4.6 has native parallel tool use. tool_choice='auto' is recommended; 'required' forces tool use every turn."
+  },
+
+  "extended_thinking": {
+    "enabled": true,
+    "budget_tokens": 10000,
+    "_comment": "Extended thinking gives Opus 4.6 a scratchpad for complex reasoning. Budget controls max thinking tokens per turn. Set to 0 to disable."
+  },
+
+  "dynamic_temperature": {
+    "enabled": false,
+    "_comment": "Anthropic API handles temperature differently. Default temp=1.0 is recommended by Anthropic for most use cases."
+  },
+
+  "tool_call_batching": {
+    "enabled": true,
+    "system_prompt_suffix": "When multiple tools are needed, call ALL of them in a single response. Do not call one tool and wait - emit all tool calls together.",
+    "_comment": "Opus 4.6 natively supports parallel tool calls. This prompt reinforces the behavior."
+  },
+
+  "pricing": {
+    "input_per_1m": 15.0,
+    "output_per_1m": 75.0,
+    "cache_write_per_1m": 18.75,
+    "cache_read_per_1m": 1.5,
+    "currency": "USD",
+    "_comment": "As of 2025. Prompt caching gives 90% discount on cache hits."
+  },
+
+  "rate_limits": {
+    "requests_per_minute": 1000,
+    "tokens_per_minute": 80000,
+    "_comment": "Default tier limits. Higher tiers available via Anthropic."
+  },
+
+  "running_config": {
+    "description": "Anthropic hosted API. Requires ANTHROPIC_API_KEY environment variable.",
+    "env_vars": {
+      "ANTHROPIC_API_KEY": "<your-api-key>",
+      "UAP_MODEL_PROFILE": "claude-opus-4.6"
+    }
+  }
+}

package/config/model-profiles/claude-sonnet-4.6.json

@@ -0,0 +1,63 @@
+{
+  "_profile": "claude-sonnet-4.6",
+  "_description": "Anthropic Claude Sonnet 4.6 (claude-sonnet-4-6-20250514). Best balance of speed, cost, and capability. Excellent tool calling, 200K context, and strong coding performance. Ideal for most agentic tasks where you want fast iteration without sacrificing quality.",
+
+  "model": "claude-sonnet-4-6-20250514",
+  "provider": "anthropic",
+  "api_base_url": "https://api.anthropic.com",
+  "max_tokens": 16384,
+  "temperature": 1.0,
+  "top_p": 1.0,
+  "timeout_ms": 120000,
+  "context_window": 200000,
+
+  "optimize_for_tool_calls": true,
+  "enable_thinking": true,
+
+  "tool_calling": {
+    "parallel_tool_calls": true,
+    "tool_choice": "auto",
+    "max_tool_calls_per_turn": 128,
+    "_comment": "Sonnet 4.6 has native parallel tool use. Excellent tool call reliability out of the box."
+  },
+
+  "extended_thinking": {
+    "enabled": true,
+    "budget_tokens": 5000,
+    "_comment": "Extended thinking available on Sonnet 4.6. Lower budget than Opus since Sonnet is faster and cheaper."
+  },
+
+  "dynamic_temperature": {
+    "enabled": false,
+    "_comment": "Anthropic API handles temperature differently. Default temp=1.0 is recommended."
+  },
+
+  "tool_call_batching": {
+    "enabled": true,
+    "system_prompt_suffix": "When multiple tools are needed, call ALL of them in a single response. Do not call one tool and wait - emit all tool calls together.",
+    "_comment": "Sonnet 4.6 natively supports parallel tool calls."
+  },
+
+  "pricing": {
+    "input_per_1m": 3.0,
+    "output_per_1m": 15.0,
+    "cache_write_per_1m": 3.75,
+    "cache_read_per_1m": 0.3,
+    "currency": "USD",
+    "_comment": "As of 2025. 5x cheaper than Opus 4.6 with strong performance."
+  },
+
+  "rate_limits": {
+    "requests_per_minute": 2000,
+    "tokens_per_minute": 160000,
+    "_comment": "Default tier limits. Higher throughput than Opus."
+  },
+
+  "running_config": {
+    "description": "Anthropic hosted API. Requires ANTHROPIC_API_KEY environment variable.",
+    "env_vars": {
+      "ANTHROPIC_API_KEY": "<your-api-key>",
+      "UAP_MODEL_PROFILE": "claude-sonnet-4.6"
+    }
+  }
+}

package/config/model-profiles/gemini-2.5-flash.json

@@ -0,0 +1,62 @@
+{
+  "_profile": "gemini-2.5-flash",
+  "_description": "Google Gemini 2.5 Flash (gemini-2.5-flash-preview-05-20). Fast and cost-effective with 1M context and thinking mode. Best for high-throughput agentic tasks where speed and cost matter. Strong tool calling with very low latency.",
+
+  "model": "gemini-2.5-flash-preview-05-20",
+  "provider": "google",
+  "api_base_url": "https://generativelanguage.googleapis.com/v1beta",
+  "max_tokens": 65536,
+  "temperature": 1.0,
+  "top_p": 0.95,
+  "timeout_ms": 120000,
+  "context_window": 1048576,
+
+  "optimize_for_tool_calls": true,
+  "enable_thinking": true,
+
+  "tool_calling": {
+    "parallel_tool_calls": true,
+    "tool_choice": "auto",
+    "_comment": "Gemini 2.5 Flash supports function calling with parallel tool use. Very fast response times."
+  },
+
+  "thinking": {
+    "enabled": true,
+    "thinking_budget": 4096,
+    "_comment": "Flash has thinking mode but with lower budget for speed. Increase for harder tasks."
+  },
+
+  "dynamic_temperature": {
+    "enabled": false,
+    "_comment": "Google recommends temperature=1.0 for most use cases."
+  },
+
+  "tool_call_batching": {
+    "enabled": true,
+    "system_prompt_suffix": "When multiple tools are needed, call ALL of them in a single response.",
+    "_comment": "Flash supports parallel function calls with very low latency."
+  },
+
+  "pricing": {
+    "input_per_1m_under_200k": 0.15,
+    "input_per_1m_over_200k": 0.3,
+    "output_per_1m": 0.6,
+    "thinking_per_1m": 0.7,
+    "currency": "USD",
+    "_comment": "As of 2025. Extremely cost-effective. Free tier available."
+  },
+
+  "rate_limits": {
+    "requests_per_minute": 2000,
+    "tokens_per_minute": 4000000,
+    "_comment": "Very high throughput. Designed for high-volume use cases."
+  },
+
+  "running_config": {
+    "description": "Google AI Studio or Vertex AI. Requires GOOGLE_API_KEY or GOOGLE_APPLICATION_CREDENTIALS.",
+    "env_vars": {
+      "GOOGLE_API_KEY": "<your-api-key>",
+      "UAP_MODEL_PROFILE": "gemini-2.5-flash"
+    }
+  }
+}

package/config/model-profiles/gemini-2.5-pro.json

@@ -0,0 +1,62 @@
+{
+  "_profile": "gemini-2.5-pro",
+  "_description": "Google Gemini 2.5 Pro (gemini-2.5-pro-preview-06-05). Top-tier reasoning model with 1M token context, native tool use, and thinking mode. Excels at code generation, long-document analysis, and complex multi-step tasks. Competitive with Opus 4 at lower cost.",
+
+  "model": "gemini-2.5-pro-preview-06-05",
+  "provider": "google",
+  "api_base_url": "https://generativelanguage.googleapis.com/v1beta",
+  "max_tokens": 65536,
+  "temperature": 1.0,
+  "top_p": 0.95,
+  "timeout_ms": 300000,
+  "context_window": 1048576,
+
+  "optimize_for_tool_calls": true,
+  "enable_thinking": true,
+
+  "tool_calling": {
+    "parallel_tool_calls": true,
+    "tool_choice": "auto",
+    "_comment": "Gemini 2.5 Pro supports function calling with automatic tool selection. Parallel calls supported."
+  },
+
+  "thinking": {
+    "enabled": true,
+    "thinking_budget": 8192,
+    "_comment": "Gemini 2.5 Pro has a thinking mode that improves reasoning quality. Budget controls max thinking tokens."
+  },
+
+  "dynamic_temperature": {
+    "enabled": false,
+    "_comment": "Google recommends temperature=1.0 for most use cases with Gemini 2.5."
+  },
+
+  "tool_call_batching": {
+    "enabled": true,
+    "system_prompt_suffix": "When multiple tools are needed, call ALL of them in a single response.",
+    "_comment": "Gemini 2.5 Pro supports parallel function calls."
+  },
+
+  "pricing": {
+    "input_per_1m_under_200k": 1.25,
+    "input_per_1m_over_200k": 2.5,
+    "output_per_1m": 10.0,
+    "thinking_per_1m": 3.5,
+    "currency": "USD",
+    "_comment": "As of 2025. Tiered input pricing based on context length. Thinking tokens billed separately."
+  },
+
+  "rate_limits": {
+    "requests_per_minute": 1000,
+    "tokens_per_minute": 4000000,
+    "_comment": "Very high token throughput. Free tier available with lower limits."
+  },
+
+  "running_config": {
+    "description": "Google AI Studio or Vertex AI. Requires GOOGLE_API_KEY or GOOGLE_APPLICATION_CREDENTIALS.",
+    "env_vars": {
+      "GOOGLE_API_KEY": "<your-api-key>",
+      "UAP_MODEL_PROFILE": "gemini-2.5-pro"
+    }
+  }
+}

package/config/model-profiles/generic.json

@@ -0,0 +1,54 @@
+{
+  "_profile": "generic",
+  "_description": "Generic defaults for any OpenAI-compatible model. Works out of the box with llama.cpp, vLLM, Ollama, OpenAI, or any server that speaks the /v1/chat/completions API. Conservative settings that prioritize reliability over speed.",
+
+  "model": "default",
+  "max_tokens": 4096,
+  "temperature": 0.6,
+  "top_p": 0.9,
+  "timeout_ms": 120000,
+  "context_window": 32768,
+
+  "optimize_for_tool_calls": true,
+  "enable_thinking": false,
+
+  "dynamic_temperature": {
+    "enabled": true,
+    "decay": 0.5,
+    "floor": 0.2,
+    "_comment": "On retry, temp = max(floor, base_temp * decay^attempt). Reduces randomness for structured output."
+  },
+
+  "tool_call_batching": {
+    "enabled": true,
+    "system_prompt_suffix": "When multiple tools are needed, call ALL of them in a single response. Do not call one tool and wait - emit all tool calls together.",
+    "_comment": "Reduces round-trips by encouraging multi-tool emission in one turn."
+  },
+
+  "server_optimization": {
+    "flash_attention": false,
+    "kv_cache_type_k": "f16",
+    "kv_cache_type_v": "f16",
+    "prompt_cache": {
+      "enabled": false
+    },
+    "speculative_decoding": {
+      "enabled": false,
+      "_comment": "Enable and configure a draft model for 2-3x speedup. Requires a small model from the same family."
+    },
+    "_comment": "Generic defaults with no hardware-specific optimizations. Override per your GPU/CPU setup."
+  },
+
+  "running_config": {
+    "description": "Generic OpenAI-compatible server. Point TARGET_URL to your inference server.",
+    "llm_server": {
+      "description": "Any OpenAI-compatible server on port 8080",
+      "command": "llama-server --model /path/to/model.gguf --port 8080 --host 0.0.0.0 --ctx-size 32768",
+      "vram_estimate": "Depends on model size and quantization"
+    },
+    "env_vars": {
+      "TARGET_URL": "http://127.0.0.1:8080",
+      "UAP_MODEL_PROFILE": "generic"
+    }
+  }
+}

package/config/model-profiles/glm-5.json

@@ -0,0 +1,61 @@
+{
+  "_profile": "glm-5",
+  "_description": "Zhipu GLM-5 (glm-5-plus). Chinese-English bilingual reasoning model with deep thinking, tool calling, and 128K context. Strong at code generation, math, and structured output. Competitive with frontier models at significantly lower cost.",
+
+  "model": "glm-5-plus",
+  "provider": "zhipu",
+  "api_base_url": "https://open.bigmodel.cn/api/paas/v4",
+  "max_tokens": 16384,
+  "temperature": 0.7,
+  "top_p": 0.9,
+  "timeout_ms": 180000,
+  "context_window": 128000,
+
+  "optimize_for_tool_calls": true,
+  "enable_thinking": true,
+
+  "tool_calling": {
+    "parallel_tool_calls": true,
+    "tool_choice": "auto",
+    "_comment": "GLM-5 supports OpenAI-compatible function calling with parallel tool use."
+  },
+
+  "thinking": {
+    "enabled": true,
+    "_comment": "GLM-5 has deep thinking mode for complex reasoning tasks. Enabled by default."
+  },
+
+  "dynamic_temperature": {
+    "enabled": true,
+    "decay": 0.5,
+    "floor": 0.2,
+    "_comment": "Lower temperature on retries for more deterministic tool call output."
+  },
+
+  "tool_call_batching": {
+    "enabled": true,
+    "system_prompt_suffix": "When multiple tools are needed, call ALL of them in a single response. Do not call one tool and wait - emit all tool calls together.",
+    "_comment": "GLM-5 supports parallel function calls."
+  },
+
+  "pricing": {
+    "input_per_1m_cny": 5.0,
+    "output_per_1m_cny": 5.0,
+    "currency": "CNY",
+    "_comment": "As of 2025. Priced in CNY. Very competitive pricing for frontier-level performance."
+  },
+
+  "rate_limits": {
+    "requests_per_minute": 300,
+    "tokens_per_minute": 300000,
+    "_comment": "Default tier limits. Higher tiers available via Zhipu."
+  },
+
+  "running_config": {
+    "description": "Zhipu AI hosted API. Requires ZHIPU_API_KEY environment variable. OpenAI-compatible endpoint.",
+    "env_vars": {
+      "ZHIPU_API_KEY": "<your-api-key>",
+      "UAP_MODEL_PROFILE": "glm-5"
+    }
+  }
+}

package/config/model-profiles/gpt-4.1.json

@@ -0,0 +1,64 @@
+{
+  "_profile": "gpt-4.1",
+  "_description": "OpenAI GPT-4.1 (gpt-4.1-2025-04-14). Purpose-built for coding and instruction following. 1M token context window, superior at long-document understanding, and optimized for agentic tool use. Best OpenAI model for software engineering tasks.",
+
+  "model": "gpt-4.1-2025-04-14",
+  "provider": "openai",
+  "api_base_url": "https://api.openai.com/v1",
+  "max_tokens": 32768,
+  "temperature": 0.7,
+  "top_p": 1.0,
+  "timeout_ms": 180000,
+  "context_window": 1047576,
+
+  "optimize_for_tool_calls": true,
+  "enable_thinking": false,
+
+  "tool_calling": {
+    "parallel_tool_calls": true,
+    "tool_choice": "auto",
+    "strict_mode": true,
+    "_comment": "GPT-4.1 excels at agentic tool use with very high reliability. Parallel function calling supported."
+  },
+
+  "structured_output": {
+    "json_mode": true,
+    "strict_json_schema": true,
+    "_comment": "Supports structured outputs with guaranteed schema adherence."
+  },
+
+  "dynamic_temperature": {
+    "enabled": true,
+    "decay": 0.5,
+    "floor": 0.2,
+    "_comment": "Lower temperature on retries for more deterministic output."
+  },
+
+  "tool_call_batching": {
+    "enabled": true,
+    "system_prompt_suffix": "When multiple tools are needed, call ALL of them in a single response using parallel function calling.",
+    "_comment": "GPT-4.1 natively supports parallel function calls with high reliability."
+  },
+
+  "pricing": {
+    "input_per_1m": 2.0,
+    "output_per_1m": 8.0,
+    "cached_input_per_1m": 0.5,
+    "currency": "USD",
+    "_comment": "As of 2025. Cheaper than GPT-4o with better coding performance."
+  },
+
+  "rate_limits": {
+    "requests_per_minute": 5000,
+    "tokens_per_minute": 800000,
+    "_comment": "Tier 3+ limits."
+  },
+
+  "running_config": {
+    "description": "OpenAI hosted API. Requires OPENAI_API_KEY environment variable.",
+    "env_vars": {
+      "OPENAI_API_KEY": "<your-api-key>",
+      "UAP_MODEL_PROFILE": "gpt-4.1"
+    }
+  }
+}