import dspy import logging import os logger = logging.getLogger(__name__) # Model providers PROVIDERS = { "openai": "OpenAI", "anthropic": "Anthropic", "groq": "GROQ", "gemini": "Google Gemini" } max_tokens = int(os.getenv("MAX_TOKENS", 6000)) # Clamp temperature to valid range (0..1) for all models default_temperature = min(1.0, max(0.0, float(os.getenv("TEMPERATURE", "1.0")))) # OpenAI reasoning models (the gpt-5 family) only accept temperature=1.0 (or None). # dspy>=3.2 validates this at dspy.LM(...) construction, so never pass the env-derived # temperature to these models or the app will fail to import when TEMPERATURE != 1.0. reasoning_temperature = 1.0 # Claude Opus 5 / Sonnet 5 removed the sampling parameters: any temperature/top_p/top_k # value is rejected with a 400. dspy passes temperature=None straight through to litellm, # which then omits the field entirely, so use None (not 1.0) for these models. no_sampling_temperature = None # Lightweight LMs used for small internal tasks (planning, classification, etc.) small_lm = dspy.LM('anthropic/claude-haiku-4-5', temperature=default_temperature, max_tokens=300, api_key=os.getenv("ANTHROPIC_API_KEY"), cache=False) mid_lm = dspy.LM('anthropic/claude-haiku-4-5', temperature=default_temperature, max_tokens=1800, api_key=os.getenv("ANTHROPIC_API_KEY"), cache=False) # OpenAI models gpt_5_nano = dspy.LM( model="openai/gpt-5-nano", api_key=os.getenv("OPENAI_API_KEY"), temperature=reasoning_temperature, max_tokens=16_000, cache=False ) gpt_5_mini = dspy.LM( model="openai/gpt-5-mini", api_key=os.getenv("OPENAI_API_KEY"), temperature=reasoning_temperature, max_tokens=16_000, cache=False ) gpt_5_2_chat_latest = dspy.LM( model="openai/gpt-5.2-chat-latest", api_key=os.getenv("OPENAI_API_KEY"), temperature=default_temperature, max_tokens=max(max_tokens, 16000), cache=False ) gpt_5_4 = dspy.LM( model="openai/gpt-5.4", api_key=os.getenv("OPENAI_API_KEY"), temperature=reasoning_temperature, max_tokens=16_000, cache=False ) gpt_5_4_pro = dspy.LM( model="openai/gpt-5.4-pro", api_key=os.getenv("OPENAI_API_KEY"), temperature=reasoning_temperature, max_tokens=16_000, cache=False ) # Anthropic models claude_haiku_4_5 = dspy.LM( model="anthropic/claude-haiku-4-5", api_key=os.getenv("ANTHROPIC_API_KEY"), temperature=default_temperature, max_tokens=max_tokens, cache=False ) claude_sonnet_5 = dspy.LM( model="anthropic/claude-sonnet-5", api_key=os.getenv("ANTHROPIC_API_KEY"), temperature=no_sampling_temperature, max_tokens=max_tokens, cache=False ) claude_opus_5 = dspy.LM( model="anthropic/claude-opus-5", api_key=os.getenv("ANTHROPIC_API_KEY"), temperature=no_sampling_temperature, max_tokens=max_tokens, cache=False ) # Groq models deepseek_r1_distill_llama_70b = dspy.LM( model="groq/deepseek-r1-distill-llama-70b", api_key=os.getenv("GROQ_API_KEY"), temperature=default_temperature, max_tokens=max_tokens, cache=False ) gpt_oss_120B = dspy.LM( model="groq/gpt-oss-120B", api_key=os.getenv("GROQ_API_KEY"), temperature=default_temperature, max_tokens=max_tokens, cache=False ) gpt_oss_20B = dspy.LM( model="groq/gpt-oss-20B", api_key=os.getenv("GROQ_API_KEY"), temperature=default_temperature, max_tokens=max_tokens, cache=False ) # Gemini models gemini_3_pro = dspy.LM( model="gemini/gemini-3-pro", api_key=os.getenv("GEMINI_API_KEY"), temperature=float(os.getenv("TEMPERATURE", 1.0)), max_tokens=max_tokens, cache=False ) gemini_3_flash = dspy.LM( model="gemini/gemini-3-flash", api_key=os.getenv("GEMINI_API_KEY"), temperature=float(os.getenv("TEMPERATURE", 1.0)), max_tokens=max_tokens, cache=False ) MODEL_OBJECTS = { # OpenAI models "gpt-5-nano": gpt_5_nano, "gpt-5-mini": gpt_5_mini, "gpt-5.2-chat-latest": gpt_5_2_chat_latest, "gpt-5.4": gpt_5_4, "gpt-5.4-pro": gpt_5_4_pro, # Anthropic models "claude-haiku-4-5": claude_haiku_4_5, "claude-sonnet-5": claude_sonnet_5, "claude-opus-5": claude_opus_5, # Groq models "deepseek-r1-distill-llama-70b": deepseek_r1_distill_llama_70b, "gpt-oss-120B": gpt_oss_120B, "gpt-oss-20B": gpt_oss_20B, # Gemini models "gemini-3-pro": gemini_3_pro, "gemini-3-flash": gemini_3_flash } # Model IDs that used to be in MODEL_OBJECTS, mapped to their current replacement. # Deployments carry the old name in the MODEL_NAME env var or in a persisted session # model_config long after the registry drops it, so resolve instead of raising KeyError. LEGACY_MODEL_ALIASES = { # Anthropic "claude-sonnet-4-6": "claude-sonnet-5", "claude-sonnet-4-5-20250929": "claude-sonnet-5", "claude-sonnet-4-20250514": "claude-sonnet-5", "claude-3-7-sonnet-latest": "claude-sonnet-5", "claude-3-5-sonnet-latest": "claude-sonnet-5", "claude-opus-4-6": "claude-opus-5", "claude-opus-4-5-20251101": "claude-opus-5", "claude-opus-4-20250514": "claude-opus-5", "claude-opus-4-1": "claude-opus-5", "claude-3-opus-latest": "claude-opus-5", "claude-3-5-haiku-latest": "claude-haiku-4-5", # OpenAI "gpt-5": "gpt-5.4", "gpt-5.2": "gpt-5.4", "gpt-5.2-pro": "gpt-5.4-pro", # Gemini "gemini-2.5-pro-preview-03-25": "gemini-3-pro", } # Used when a name is neither a known model nor a known alias. FALLBACK_MODEL_NAME = "claude-sonnet-5" def resolve_model_name(model_name: str) -> str: """Map a model name to a key that exists in MODEL_OBJECTS. Returns the name unchanged when it is current, the replacement when it is a retired ID, and FALLBACK_MODEL_NAME when it is unknown. """ if model_name in MODEL_OBJECTS: return model_name resolved = LEGACY_MODEL_ALIASES.get(model_name) if resolved in MODEL_OBJECTS: logger.warning( f"Model '{model_name}' is no longer offered by this app; using '{resolved}' instead." ) return resolved logger.warning( f"Unknown model '{model_name}'; falling back to '{FALLBACK_MODEL_NAME}'." ) return FALLBACK_MODEL_NAME def get_model_object(model_name: str): """Get model object by name, resolving retired and unknown names""" return MODEL_OBJECTS[resolve_model_name(model_name)] # Get max tokens from environment max_tokens = int(os.getenv("MAX_TOKENS", 6000)) # Tiers based on cost per 1K tokens MODEL_TIERS = { "tier1": { "name": "Basic", "credits": 1, "models": [ "gpt-5-nano", "gpt-oss-20B" ] }, "tier2": { "name": "Standard", "credits": 3, "models": [ "claude-haiku-4-5", "gpt-5-mini", "gpt-5.2-chat-latest" ] }, "tier3": { "name": "Premium", "credits": 5, "models": [ "claude-sonnet-5", "deepseek-r1-distill-llama-70b", "gpt-oss-120B", "gemini-3-flash" ] }, "tier4": { "name": "Premium Plus", "credits": 20, "models": [ "gpt-5.4", "claude-opus-5", "gemini-3-pro" ] }, "tier5": { "name": "Ultimate", "credits": 50, "models": [ "gpt-5.4-pro" ] } } # Model metadata (display name, context window, etc.) MODEL_METADATA = { # OpenAI "gpt-5-nano": {"display_name": "GPT-5 Nano", "context_window": 64000}, "gpt-5-mini": {"display_name": "GPT-5 Mini", "context_window": 150000}, "gpt-5.2-chat-latest": {"display_name": "GPT-5.2 Chat", "context_window": 400000}, "gpt-5.4": {"display_name": "GPT-5.4", "context_window": 1050000}, "gpt-5.4-pro": {"display_name": "GPT-5.4 Pro", "context_window": 1050000}, # Anthropic "claude-haiku-4-5": {"display_name": "Claude Haiku 4.5", "context_window": 200000}, "claude-sonnet-5": {"display_name": "Claude Sonnet 5", "context_window": 1000000}, "claude-opus-5": {"display_name": "Claude Opus 5", "context_window": 1000000}, # GROQ "deepseek-r1-distill-llama-70b": {"display_name": "DeepSeek R1 Distill Llama 70b", "context_window": 32768}, "gpt-oss-120B": {"display_name": "OpenAI gpt oss 120B", "context_window": 128000}, "gpt-oss-20B": {"display_name": "OpenAI gpt oss 20B", "context_window": 128000}, # Gemini "gemini-3-pro": {"display_name": "Gemini 3 Pro", "context_window": 1000000}, "gemini-3-flash": {"display_name": "Gemini 3 Flash", "context_window": 1000000}, } MODEL_COSTS = { "openai": { "gpt-5-nano": {"input": 0.00005, "output": 0.0004}, "gpt-5-mini": {"input": 0.00025, "output": 0.002}, "gpt-5.2-chat-latest": {"input": 0.0005, "output": 0.002}, "gpt-5.4": {"input": 0.0025, "output": 0.015}, "gpt-5.4-pro": {"input": 0.03, "output": 0.18}, }, "anthropic": { "claude-haiku-4-5": {"input": 0.001, "output": 0.005}, "claude-sonnet-5": {"input": 0.003, "output": 0.015}, "claude-opus-5": {"input": 0.005, "output": 0.025}, }, "groq": { "deepseek-r1-distill-llama-70b": {"input": 0.00075, "output": 0.00099}, "gpt-oss-120B": {"input": 0.00075, "output": 0.00099}, "gpt-oss-20B": {"input": 0.00075, "output": 0.00099} }, "gemini": { "gemini-3-pro": {"input": 0.0002, "output": 0.001}, "gemini-3-flash": {"input": 0.0001, "output": 0.0005} } } # Helper functions # Provider for every model in the registry, derived from MODEL_COSTS so the two # cannot drift apart. MODEL_PROVIDERS = {model: provider for provider, models in MODEL_COSTS.items() for model in models} # Checked in order; "gpt-oss" must precede "gpt" since the gpt-oss models are GROQ's. _PROVIDER_PREFIXES = ( ("claude", "anthropic"), ("gpt-oss", "groq"), ("deepseek", "groq"), ("gpt", "openai"), ("gemini", "gemini"), ) def get_provider_for_model(model_name): """Determine the provider based on model name""" if not model_name: return "Unknown" model_name = str(model_name).lower() provider = MODEL_PROVIDERS.get(model_name) if provider: return provider # Retired and aliased IDs are absent from MODEL_COSTS, but the family prefix # still identifies the provider. Callers use this to choose provider-specific # request parameters, where guessing beats returning "Unknown". for prefix, prefix_provider in _PROVIDER_PREFIXES: if model_name.startswith(prefix): return prefix_provider return "Unknown" def get_model_tier(model_name): """Get the tier of a model""" for tier_id, tier_info in MODEL_TIERS.items(): if model_name in tier_info["models"]: return tier_id return "tier1" # Default to tier1 if not found def calculate_cost(model_name, input_tokens, output_tokens): """Calculate the cost for using the model based on tokens""" if not model_name: return 0 # Convert tokens to thousands input_tokens_in_thousands = input_tokens / 1000 output_tokens_in_thousands = output_tokens / 1000 # Get model provider model_provider = get_provider_for_model(model_name) # Handle case where model is not found if model_provider == "Unknown" or model_name not in MODEL_COSTS.get(model_provider, {}): return 0 return (input_tokens_in_thousands * MODEL_COSTS[model_provider][model_name]["input"] + output_tokens_in_thousands * MODEL_COSTS[model_provider][model_name]["output"]) def get_credit_cost(model_name): """Get the credit cost for a model""" tier_id = get_model_tier(model_name) return MODEL_TIERS[tier_id]["credits"] def get_display_name(model_name): """Get the display name for a model""" return MODEL_METADATA.get(model_name, {}).get("display_name", model_name) def get_context_window(model_name): """Get the context window size for a model""" return MODEL_METADATA.get(model_name, {}).get("context_window", 4096) def get_all_models_for_provider(provider): """Get all models for a specific provider""" if provider not in MODEL_COSTS: return [] return list(MODEL_COSTS[provider].keys()) def get_models_by_tier(tier_id): """Get all models for a specific tier""" return MODEL_TIERS.get(tier_id, {}).get("models", [])