pdd-cli 0.0.25__py3-none-any.whl → 0.0.26__py3-none-any.whl

pdd/llm_invoke.py

@@ -1,389 +1,1202 @@
-#!/usr/bin/env python
-"""
-llm_invoke.py
-
-This module provides a single function, llm_invoke, that runs a prompt with a given input
-against a language model (LLM) using Langchain and returns the output, cost, and model name.
-The function supports model selection based on cost/ELO interpolation controlled by the
-"strength" parameter. It also implements a retry mechanism: if a model invocation fails,
-it falls back to the next candidate (cheaper for strength < 0.5, or higher ELO for strength ≥ 0.5).
-
-Usage:
-    from llm_invoke import llm_invoke
-    result = llm_invoke(prompt, input_json, strength, temperature, verbose=True, output_pydantic=MyPydanticClass)
-    # result is a dict with keys: 'result', 'cost', 'model_name'
-    
-Environment:
-    - PDD_MODEL_DEFAULT: if set, used as the base model name. Otherwise defaults to "gpt-4.1-nano".
-    - PDD_PATH: if set, models are loaded from $PDD_PATH/data/llm_model.csv; otherwise from ./data/llm_model.csv.
-    - Models that require an API key will check the corresponding environment variable (name provided in the CSV).
-"""
+# Corrected code_under_test (llm_invoke.py)
+# Added optional debugging prints in _select_model_candidates
 
 import os
-import csv
-import json
+import pandas as pd
+import litellm
+import logging # ADDED FOR DETAILED LOGGING
+
+# --- Configure Detailed Logging for LiteLLM --- MODIFIED SECTION
+PROJECT_ROOT_FOR_LOG = '/Users/gregtanaka/Documents/pdd_cloud/pdd' # Explicit project root
+LOG_FILE_PATH = os.path.join(PROJECT_ROOT_FOR_LOG, 'litellm_debug.log')
+
+# Get the litellm logger specifically
+litellm_logger = logging.getLogger("litellm")
+litellm_logger.setLevel(logging.DEBUG) # Set its level to DEBUG
+
+# Create a file handler
+file_handler = logging.FileHandler(LOG_FILE_PATH, mode='w')
+file_handler.setLevel(logging.DEBUG)
+
+# Create a formatter and add it to the handler
+formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')
+file_handler.setFormatter(formatter)
+
+# Add the handler to the litellm logger
+# Check if handlers are already present to avoid duplication if module is reloaded
+if not litellm_logger.handlers:
+    litellm_logger.addHandler(file_handler)
+
+# Also ensure the root logger has a basic handler if nothing else is configured
+# This can help catch messages if litellm logs to root or other unnamed loggers
+if not logging.getLogger().handlers: # Check root logger
+    logging.basicConfig(level=logging.DEBUG) # Default to console for other logs
+# --- End Detailed Logging Configuration ---
 
-from pydantic import BaseModel, Field
+
+import json
 from rich import print as rprint
-from rich.errors import MarkupError
-
-# Langchain core and community imports
-from langchain_core.prompts import PromptTemplate, ChatPromptTemplate
-from langchain_community.cache import SQLiteCache
-from langchain.globals import set_llm_cache
-from langchain_core.output_parsers import PydanticOutputParser, StrOutputParser
-from langchain_core.runnables import RunnablePassthrough, ConfigurableField
-
-# LLM provider imports
-from langchain_openai import AzureChatOpenAI, ChatOpenAI, OpenAI
-from langchain_fireworks import Fireworks
-from langchain_anthropic import ChatAnthropic
-from langchain_google_genai import ChatGoogleGenerativeAI
-from langchain_google_vertexai import ChatVertexAI
-from langchain_groq import ChatGroq
-from langchain_together import Together
-from langchain_ollama.llms import OllamaLLM
-
-from langchain.callbacks.base import BaseCallbackHandler
-from langchain.schema import LLMResult
-
-# ---------------- Internal Helper Classes and Functions ---------------- #
-
-class CompletionStatusHandler(BaseCallbackHandler):
-    """
-    Callback handler to capture LLM token usage and completion metadata.
-    """
-    def __init__(self):
-        self.is_complete = False
-        self.finish_reason = None
-        self.input_tokens = None
-        self.output_tokens = None
-
-    def on_llm_end(self, response: LLMResult, **kwargs) -> None:
-        self.is_complete = True
-        if response.generations and response.generations[0]:
-            generation = response.generations[0][0]
-            # Safely get generation_info; if it's None, default to {}
-            generation_info = generation.generation_info or {}
-            self.finish_reason = (generation_info.get('finish_reason') or "").lower()
-            
-            # Attempt to get token usage from generation.message if available.
-            if (
-                hasattr(generation, "message")
-                and generation.message is not None
-                and hasattr(generation.message, "usage_metadata")
-                and generation.message.usage_metadata
-            ):
-                usage_metadata = generation.message.usage_metadata
-            else:
-                usage_metadata = generation_info.get("usage_metadata", {})
-            
-            self.input_tokens = usage_metadata.get('input_tokens', 0)
-            self.output_tokens = usage_metadata.get('output_tokens', 0)
+from dotenv import load_dotenv
+from pathlib import Path
+from typing import Optional, Dict, List, Any, Type, Union
+from pydantic import BaseModel, ValidationError
+import openai  # Import openai for exception handling as LiteLLM maps to its types
+from langchain_core.prompts import PromptTemplate
+import warnings
+import time as time_module # Alias to avoid conflict with 'time' parameter
+# Import the default model constant
+from pdd import DEFAULT_LLM_MODEL
 
-class ModelInfo:
-    """
-    Represents information about an LLM model as loaded from the CSV.
-    """
-    def __init__(self, provider, model, input_cost, output_cost, coding_arena_elo,
-                 base_url, api_key, counter, encoder, max_tokens, max_completion_tokens,
-                 structured_output):
-        self.provider = provider.strip() if provider else ""
-        self.model = model.strip() if model else ""
-        self.input_cost = float(input_cost) if input_cost else 0.0
-        self.output_cost = float(output_cost) if output_cost else 0.0
-        self.average_cost = (self.input_cost + self.output_cost) / 2
-        self.coding_arena_elo = float(coding_arena_elo) if coding_arena_elo else 0.0
-        self.base_url = base_url.strip() if base_url else None
-        self.api_key = api_key.strip() if api_key else None
-        self.counter = counter.strip() if counter else None
-        self.encoder = encoder.strip() if encoder else None
-        self.max_tokens = int(max_tokens) if max_tokens else None
-        self.max_completion_tokens = int(max_completion_tokens) if max_completion_tokens else None
-        self.structured_output = (str(structured_output).lower() == 'true') if structured_output else False
-
-def load_models():
+# Opt-in to future pandas behavior regarding downcasting
+pd.set_option('future.no_silent_downcasting', True)
+
+# <<< SET LITELLM DEBUG LOGGING >>>
+# os.environ['LITELLM_LOG'] = 'DEBUG' # Keep commented out unless debugging LiteLLM itself
+
+# --- Constants and Configuration ---
+
+# Determine project root: 1. PDD_PATH env var, 2. Search upwards from script, 3. CWD
+PROJECT_ROOT = None
+PDD_PATH_ENV = os.getenv("PDD_PATH")
+
+if PDD_PATH_ENV:
+    _path_from_env = Path(PDD_PATH_ENV)
+    if _path_from_env.is_dir():
+        PROJECT_ROOT = _path_from_env.resolve()
+        # print(f"[DEBUG] Using PROJECT_ROOT from PDD_PATH: {PROJECT_ROOT}") # Optional debug
+    else:
+        warnings.warn(f"PDD_PATH environment variable ('{PDD_PATH_ENV}') is set but not a valid directory. Attempting auto-detection.")
+
+if PROJECT_ROOT is None: # If PDD_PATH wasn't set or was invalid
+    try:
+        # Start from the directory containing this script
+        current_dir = Path(__file__).resolve().parent
+        # Look for project markers (e.g., .git, pyproject.toml, data/, .env)
+        # Go up a maximum of 5 levels to prevent infinite loops
+        for _ in range(5):
+            has_git = (current_dir / ".git").exists()
+            has_pyproject = (current_dir / "pyproject.toml").exists()
+            has_data = (current_dir / "data").is_dir()
+            has_dotenv = (current_dir / ".env").exists()
+
+            if has_git or has_pyproject or has_data or has_dotenv:
+                PROJECT_ROOT = current_dir
+                # print(f"[DEBUG] Determined PROJECT_ROOT by marker search: {PROJECT_ROOT}") # Optional debug
+                break
+
+            parent_dir = current_dir.parent
+            if parent_dir == current_dir: # Reached filesystem root
+                break
+            current_dir = parent_dir
+
+    except NameError: # __file__ might not be defined (e.g., interactive session)
+        warnings.warn("__file__ not defined. Cannot automatically detect project root from script location.")
+    except Exception as e: # Catch potential permission errors etc.
+        warnings.warn(f"Error during project root auto-detection: {e}")
+
+if PROJECT_ROOT is None: # Fallback to CWD if no method succeeded
+    PROJECT_ROOT = Path.cwd().resolve()
+    warnings.warn(f"Could not determine project root automatically. Using current working directory: {PROJECT_ROOT}. Ensure this is the intended root or set the PDD_PATH environment variable.")
+
+
+ENV_PATH = PROJECT_ROOT / ".env"
+# --- Determine LLM_MODEL_CSV_PATH ---
+# Prioritize ~/.pdd/llm_model.csv
+user_pdd_dir = Path.home() / ".pdd"
+user_model_csv_path = user_pdd_dir / "llm_model.csv"
+
+if user_model_csv_path.is_file():
+    LLM_MODEL_CSV_PATH = user_model_csv_path
+    print(f"[INFO] Using user-specific LLM model CSV: {LLM_MODEL_CSV_PATH}")
+else:
+    LLM_MODEL_CSV_PATH = PROJECT_ROOT / "data" / "llm_model.csv"
+    print(f"[INFO] Using project LLM model CSV: {LLM_MODEL_CSV_PATH}")
+# ---------------------------------
+
+# Load environment variables from .env file
+# print(f"[DEBUG] Attempting to load .env from: {ENV_PATH}") # Optional debug
+if ENV_PATH.exists():
+    load_dotenv(dotenv_path=ENV_PATH)
+    # print(f"[DEBUG] Loaded .env file from: {ENV_PATH}") # Optional debug
+else:
+    # Reduce verbosity if .env is optional or often missing
+    # warnings.warn(f".env file not found at {ENV_PATH}. API keys might be missing.")
+    pass # Silently proceed if .env is optional
+
+# Default model if PDD_MODEL_DEFAULT is not set
+# Use the imported constant as the default
+DEFAULT_BASE_MODEL = os.getenv("PDD_MODEL_DEFAULT", DEFAULT_LLM_MODEL)
+
+# --- LiteLLM Cache Configuration (S3 compatible for GCS, with SQLite fallback) ---
+GCS_BUCKET_NAME = os.getenv("GCS_BUCKET_NAME")
+GCS_ENDPOINT_URL = "https://storage.googleapis.com" # GCS S3 compatibility endpoint
+GCS_REGION_NAME = os.getenv("GCS_REGION_NAME", "auto") # Often 'auto' works for GCS
+GCS_HMAC_ACCESS_KEY_ID = os.getenv("GCS_HMAC_ACCESS_KEY_ID") # Load HMAC Key ID
+GCS_HMAC_SECRET_ACCESS_KEY = os.getenv("GCS_HMAC_SECRET_ACCESS_KEY") # Load HMAC Secret
+
+cache_configured = False
+
+if GCS_BUCKET_NAME and GCS_HMAC_ACCESS_KEY_ID and GCS_HMAC_SECRET_ACCESS_KEY:
+    # Store original AWS credentials before overwriting for GCS cache setup
+    original_aws_access_key_id = os.environ.get('AWS_ACCESS_KEY_ID')
+    original_aws_secret_access_key = os.environ.get('AWS_SECRET_ACCESS_KEY')
+    original_aws_region_name = os.environ.get('AWS_REGION_NAME')
+
+    try:
+        # Temporarily set AWS env vars to GCS HMAC keys for S3 compatible cache
+        os.environ['AWS_ACCESS_KEY_ID'] = GCS_HMAC_ACCESS_KEY_ID
+        os.environ['AWS_SECRET_ACCESS_KEY'] = GCS_HMAC_SECRET_ACCESS_KEY
+        # os.environ['AWS_REGION_NAME'] = GCS_REGION_NAME  # Uncomment if needed
+
+        litellm.cache = litellm.Cache(
+            type="s3",
+            s3_bucket_name=GCS_BUCKET_NAME,
+            s3_region_name=GCS_REGION_NAME, # Pass region explicitly to cache
+            s3_endpoint_url=GCS_ENDPOINT_URL,
+        )
+        print(f"[INFO] LiteLLM cache configured for GCS bucket (S3 compatible): {GCS_BUCKET_NAME}")
+        cache_configured = True
+
+    except Exception as e:
+        warnings.warn(f"Failed to configure LiteLLM S3/GCS cache: {e}. Attempting SQLite cache fallback.")
+        litellm.cache = None # Explicitly disable cache on failure (will try SQLite next)
+
+    finally:
+        # Restore original AWS credentials after cache setup attempt
+        if original_aws_access_key_id is not None:
+            os.environ['AWS_ACCESS_KEY_ID'] = original_aws_access_key_id
+        elif 'AWS_ACCESS_KEY_ID' in os.environ:
+            del os.environ['AWS_ACCESS_KEY_ID']
+
+        if original_aws_secret_access_key is not None:
+            os.environ['AWS_SECRET_ACCESS_KEY'] = original_aws_secret_access_key
+        elif 'AWS_SECRET_ACCESS_KEY' in os.environ:
+            del os.environ['AWS_SECRET_ACCESS_KEY']
+
+        if original_aws_region_name is not None:
+            os.environ['AWS_REGION_NAME'] = original_aws_region_name
+        elif 'AWS_REGION_NAME' in os.environ:
+            pass # Or just leave it if the temporary setting wasn't done/needed
+
+if not cache_configured:
+    try:
+        # Try SQLite-based cache as a fallback
+        sqlite_cache_path = PROJECT_ROOT / "litellm_cache.sqlite"
+        litellm.cache = litellm.Cache(type="sqlite", cache_path=str(sqlite_cache_path))
+        print(f"[INFO] LiteLLM SQLite cache configured at {sqlite_cache_path}")
+        cache_configured = True
+    except Exception as e2:
+        warnings.warn(f"Failed to configure LiteLLM SQLite cache: {e2}. Caching is disabled.")
+        litellm.cache = None
+
+if not cache_configured:
+    warnings.warn("All LiteLLM cache configuration attempts failed. Caching is disabled.")
+    litellm.cache = None
+
+# --- LiteLLM Callback for Success Logging ---
+
+# Module-level storage for last callback data (Use with caution in concurrent environments)
+_LAST_CALLBACK_DATA = {
+    "input_tokens": 0,
+    "output_tokens": 0,
+    "finish_reason": None,
+    "cost": 0.0,
+}
+
+def _litellm_success_callback(
+    kwargs: Dict[str, Any],              # kwargs passed to completion
+    completion_response: Any,            # response object from completion
+    start_time: float, end_time: float   # start/end time
+):
     """
-    Loads model information from llm_model.csv located in either $PDD_PATH/data or ./data.
+    LiteLLM success callback to capture usage and finish reason.
+    Stores data in a module-level variable for potential retrieval.
     """
-    pdd_path = os.environ.get('PDD_PATH', '.')
-    models_file = os.path.join(pdd_path, 'data', 'llm_model.csv')
-    models = []
+    global _LAST_CALLBACK_DATA
+    usage = getattr(completion_response, 'usage', None)
+    input_tokens = getattr(usage, 'prompt_tokens', 0)
+    output_tokens = getattr(usage, 'completion_tokens', 0)
+    finish_reason = getattr(completion_response.choices[0], 'finish_reason', None)
+
+    calculated_cost = 0.0
     try:
-        with open(models_file, newline='') as csvfile:
-            reader = csv.DictReader(csvfile)
-            for row in reader:
-                model_info = ModelInfo(
-                    provider=row.get('provider',''),
-                    model=row.get('model',''),
-                    input_cost=row.get('input','0'),
-                    output_cost=row.get('output','0'),
-                    coding_arena_elo=row.get('coding_arena_elo','0'),
-                    base_url=row.get('base_url',''),
-                    api_key=row.get('api_key',''),
-                    counter=row.get('counter',''),
-                    encoder=row.get('encoder',''),
-                    max_tokens=row.get('max_tokens',''),
-                    max_completion_tokens=row.get('max_completion_tokens',''),
-                    structured_output=row.get('structured_output','False')
+        # Attempt 1: Use the response object directly (works for most single calls)
+        cost_val = litellm.completion_cost(completion_response=completion_response)
+        calculated_cost = cost_val if cost_val is not None else 0.0
+    except Exception as e1:
+        # Attempt 2: If response object failed (e.g., missing provider in model name),
+        # try again using explicit model from kwargs and tokens from usage.
+        # This is often needed for batch completion items.
+        print(f"[DEBUG] Attempting cost calculation with fallback method: {e1}")
+        try:
+            model_name = kwargs.get("model") # Get original model name from input kwargs
+            if model_name and usage:
+                prompt_tokens = getattr(usage, 'prompt_tokens', 0)
+                completion_tokens = getattr(usage, 'completion_tokens', 0)
+                cost_val = litellm.completion_cost(
+                    model=model_name,
+                    prompt_tokens=prompt_tokens,
+                    completion_tokens=completion_tokens
                 )
-                models.append(model_info)
-    except FileNotFoundError:
-        raise FileNotFoundError(f"llm_model.csv not found at {models_file}")
-    return models
+                calculated_cost = cost_val if cost_val is not None else 0.0
+            else:
+                # If we can't get model name or usage, fallback to 0
+                calculated_cost = 0.0
+                # Optional: Log the original error e1 if needed
+                # print(f"[Callback WARN] Failed to calculate cost with response object ({e1}) and fallback failed.")
+        except Exception as e2:
+            # Optional: Log secondary error e2 if needed
+            # print(f"[Callback WARN] Failed to calculate cost with fallback method: {e2}")
+            calculated_cost = 0.0 # Default to 0 on any error
+            print(f"[DEBUG] Cost calculation failed with fallback method: {e2}")
 
-def select_model(models, base_model_name):
-    """
-    Retrieve the base model whose name matches base_model_name. Raises an error if not found.
-    """
-    for model in models:
-        if model.model == base_model_name:
-            return model
-    raise ValueError(f"Base model '{base_model_name}' not found in the models list.")
+    _LAST_CALLBACK_DATA["input_tokens"] = input_tokens
+    _LAST_CALLBACK_DATA["output_tokens"] = output_tokens
+    _LAST_CALLBACK_DATA["finish_reason"] = finish_reason
+    _LAST_CALLBACK_DATA["cost"] = calculated_cost # Store the calculated cost
+
+    # Callback doesn't need to return a value now
+    # return calculated_cost
+
+    # Example of logging within the callback (can be expanded)
+    # print(f"[Callback] Tokens: In={input_tokens}, Out={output_tokens}. Reason: {finish_reason}. Cost: ${calculated_cost:.6f}")
+
+# Register the callback with LiteLLM
+litellm.success_callback = [_litellm_success_callback]
+
+# --- Helper Functions ---
+
+def _load_model_data(csv_path: Path) -> pd.DataFrame:
+    """Loads and preprocesses the LLM model data from CSV."""
+    if not csv_path.exists():
+        raise FileNotFoundError(f"LLM model CSV not found at {csv_path}")
+    try:
+        df = pd.read_csv(csv_path)
+        # Basic validation and type conversion
+        required_cols = ['provider', 'model', 'input', 'output', 'coding_arena_elo', 'api_key', 'structured_output', 'reasoning_type']
+        for col in required_cols:
+            if col not in df.columns:
+                raise ValueError(f"Missing required column in CSV: {col}")
+
+        # Convert numeric columns, handling potential errors
+        numeric_cols = ['input', 'output', 'coding_arena_elo', 'max_tokens',
+                        'max_completion_tokens', 'max_reasoning_tokens']
+        for col in numeric_cols:
+            if col in df.columns:
+                # Use errors='coerce' to turn unparseable values into NaN
+                df[col] = pd.to_numeric(df[col], errors='coerce')
+
+        # Fill NaN in critical numeric columns used for selection/interpolation
+        df['input'] = df['input'].fillna(0.0)
+        df['output'] = df['output'].fillna(0.0)
+        df['coding_arena_elo'] = df['coding_arena_elo'].fillna(0) # Use 0 ELO for missing
+        # Ensure max_reasoning_tokens is numeric, fillna with 0
+        df['max_reasoning_tokens'] = df['max_reasoning_tokens'].fillna(0).astype(int) # Ensure int
+
+        # Calculate average cost (handle potential division by zero if needed, though unlikely with fillna)
+        df['avg_cost'] = (df['input'] + df['output']) / 2
+
+        # Ensure boolean interpretation for structured_output
+        if 'structured_output' in df.columns:
+             df['structured_output'] = df['structured_output'].fillna(False).astype(bool)
+        else:
+             df['structured_output'] = False # Assume false if column missing
+
+        # Ensure reasoning_type is string, fillna with 'none' and lowercase
+        df['reasoning_type'] = df['reasoning_type'].fillna('none').astype(str).str.lower()
+
+        # Ensure api_key is treated as string, fill NaN with empty string ''
+        # This handles cases where read_csv might interpret empty fields as NaN
+        df['api_key'] = df['api_key'].fillna('').astype(str)
+
+        return df
+    except Exception as e:
+        raise RuntimeError(f"Error loading or processing LLM model CSV {csv_path}: {e}") from e
+
+def _select_model_candidates(
+    strength: float,
+    base_model_name: str,
+    model_df: pd.DataFrame
+) -> List[Dict[str, Any]]:
+    """Selects and sorts candidate models based on strength and availability."""
+
+    # 1. Filter by API Key Name Presence (initial availability check)
+    # Keep models with a non-empty api_key field in the CSV.
+    # The actual key value check happens later.
+    # Allow models with empty api_key (e.g., Bedrock using AWS creds, local models)
+    available_df = model_df[model_df['api_key'].notna()].copy()
+
+    # --- Check if the initial DataFrame itself was empty ---
+    if model_df.empty:
+        raise ValueError("Loaded model data is empty. Check CSV file.")
+
+    # --- Check if filtering resulted in empty (might indicate all models had NaN api_key) ---
+    if available_df.empty:
+        # This case is less likely if notna() is the only filter, but good to check.
+        rprint("[WARN] No models found after filtering for non-NaN api_key. Check CSV 'api_key' column.")
+        # Decide if this should be a hard error or allow proceeding if logic permits
+        # For now, let's raise an error as it likely indicates a CSV issue.
+        raise ValueError("No models available after initial filtering (all had NaN 'api_key'?).")
+
+    # 2. Find Base Model
+    base_model_row = available_df[available_df['model'] == base_model_name]
+    if base_model_row.empty:
+        # Try finding base model in the *original* df in case it was filtered out
+        original_base = model_df[model_df['model'] == base_model_name]
+        if not original_base.empty:
+             raise ValueError(f"Base model '{base_model_name}' found in CSV but requires API key '{original_base.iloc[0]['api_key']}' which might be missing or invalid configuration.")
+        else:
+             raise ValueError(f"Specified base model '{base_model_name}' not found in the LLM model CSV.")
+
+    base_model = base_model_row.iloc[0]
+
+    # 3. Determine Target and Sort
+    candidates = []
+    target_metric_value = None # For debugging print
 
-def get_candidate_models(strength, models, base_model):
-    """
-    Returns ordered list of candidate models based on strength parameter.
-    Only includes models with available API keys.
-    """
-    # Filter for models with valid API keys (including test environment)
-    available_models = [m for m in models 
-                       if not m.api_key or 
-                       os.environ.get(m.api_key) or 
-                       m.api_key == "EXISTING_KEY"]
-    
-    if not available_models:
-        raise RuntimeError("No models available with valid API keys")
-
-    # For base model case (strength = 0.5), use base model if available
     if strength == 0.5:
-        base_candidates = [m for m in available_models if m.model == base_model.model]
-        if base_candidates:
-            return base_candidates
-        return [available_models[0]]
-
-    # For strength < 0.5, prioritize cheaper models
-    if strength < 0.5:
-        # Get models cheaper than or equal to base model
-        cheaper_models = [m for m in available_models 
-                         if m.average_cost <= base_model.average_cost]
-        if not cheaper_models:
-            return [available_models[0]]
-            
-        # For test environment, honor the mock model setup
-        test_models = [m for m in cheaper_models if m.api_key == "EXISTING_KEY"]
-        if test_models:
-            return test_models
-
-        # Production path: interpolate based on cost
-        cheapest = min(cheaper_models, key=lambda m: m.average_cost)
-        cost_range = base_model.average_cost - cheapest.average_cost
-        target_cost = cheapest.average_cost + (strength / 0.5) * cost_range
-        return sorted(cheaper_models, key=lambda m: abs(m.average_cost - target_cost))
-
-    # For strength > 0.5, prioritize higher ELO models
-    # Get models with higher or equal ELO than base_model
-    better_models = [m for m in available_models 
-                    if m.coding_arena_elo >= base_model.coding_arena_elo]
-    if not better_models:
-        return [available_models[0]]
-        
-    # For test environment, honor the mock model setup
-    test_models = [m for m in better_models if m.api_key == "EXISTING_KEY"]
-    if test_models:
-        return test_models
-
-    # Production path: interpolate based on ELO
-    highest = max(better_models, key=lambda m: m.coding_arena_elo)
-    elo_range = highest.coding_arena_elo - base_model.coding_arena_elo
-    target_elo = base_model.coding_arena_elo + ((strength - 0.5) / 0.5) * elo_range
-    return sorted(better_models, key=lambda m: abs(m.coding_arena_elo - target_elo))
-
-def create_llm_instance(selected_model, temperature, handler):
-    """
-    Creates an instance of the LLM using the selected_model parameters.
-    Handles provider-specific settings and token limit configurations.
-    """
-    provider = selected_model.provider.lower()
-    model_name = selected_model.model
-    base_url = selected_model.base_url
-    api_key_env = selected_model.api_key
-    max_completion_tokens = selected_model.max_completion_tokens
-    max_tokens = selected_model.max_tokens
-
-    api_key = os.environ.get(api_key_env) if api_key_env else None
-
-    if provider == 'openai':
-        if base_url:
-            llm = ChatOpenAI(model=model_name, temperature=temperature,
-                             openai_api_key=api_key, callbacks=[handler],
-                             openai_api_base=base_url)
+        # target_model = base_model
+        # Sort remaining by ELO descending as fallback
+        available_df['sort_metric'] = -available_df['coding_arena_elo'] # Negative for descending sort
+        candidates = available_df.sort_values(by='sort_metric').to_dict('records')
+        # Ensure base model is first if it exists
+        if any(c['model'] == base_model_name for c in candidates):
+            candidates.sort(key=lambda x: 0 if x['model'] == base_model_name else 1)
+        target_metric_value = f"Base Model ELO: {base_model['coding_arena_elo']}"
+
+    elif strength < 0.5:
+        # Interpolate by Cost (downwards from base)
+        base_cost = base_model['avg_cost']
+        cheapest_model = available_df.loc[available_df['avg_cost'].idxmin()]
+        cheapest_cost = cheapest_model['avg_cost']
+
+        if base_cost <= cheapest_cost: # Handle edge case where base is cheapest
+             target_cost = cheapest_cost + strength * (base_cost - cheapest_cost) # Will be <= base_cost
         else:
-            if model_name.startswith('o'):
-                llm = ChatOpenAI(model=model_name, temperature=temperature,
-                                 openai_api_key=api_key, callbacks=[handler],
-                                 reasoning={"effort": "high","summary": "auto"})
-            else:
-                llm = ChatOpenAI(model=model_name, temperature=temperature,
-                                 openai_api_key=api_key, callbacks=[handler])
-    elif provider == 'anthropic':
-        # Special case for Claude 3.7 Sonnet with thinking token budget
-        if 'claude-3-7-sonnet' in model_name:
-            llm = ChatAnthropic(
-                model=model_name, 
-                temperature=temperature, 
-                callbacks=[handler],
-                thinking={"type": "enabled", "budget_tokens": 4000}  # 32K thinking token budget
-            )
+             # Interpolate between cheapest and base
+             target_cost = cheapest_cost + (strength / 0.5) * (base_cost - cheapest_cost)
+
+        available_df['sort_metric'] = abs(available_df['avg_cost'] - target_cost)
+        candidates = available_df.sort_values(by='sort_metric').to_dict('records')
+        target_metric_value = f"Target Cost: {target_cost:.6f}"
+
+    else: # strength > 0.5
+        # Interpolate by ELO (upwards from base)
+        base_elo = base_model['coding_arena_elo']
+        highest_elo_model = available_df.loc[available_df['coding_arena_elo'].idxmax()]
+        highest_elo = highest_elo_model['coding_arena_elo']
+
+        if highest_elo <= base_elo: # Handle edge case where base has highest ELO
+            target_elo = base_elo + (strength - 0.5) * (highest_elo - base_elo) # Will be >= base_elo
         else:
-            llm = ChatAnthropic(model=model_name, temperature=temperature, callbacks=[handler])
-    elif provider == 'google':
-        llm = ChatGoogleGenerativeAI(model=model_name, temperature=temperature, callbacks=[handler])
-    elif provider == 'googlevertexai':
-        llm = ChatVertexAI(model=model_name, temperature=temperature, callbacks=[handler])
-    elif provider == 'ollama':
-        llm = OllamaLLM(model=model_name, temperature=temperature, callbacks=[handler])
-    elif provider == 'azure':
-        llm = AzureChatOpenAI(model=model_name, temperature=temperature,
-                              callbacks=[handler], openai_api_key=api_key, openai_api_base=base_url)
-    elif provider == 'fireworks':
-        llm = Fireworks(model=model_name, temperature=temperature, callbacks=[handler])
-    elif provider == 'together':
-        llm = Together(model=model_name, temperature=temperature, callbacks=[handler])
-    elif provider == 'groq':
-        llm = ChatGroq(model_name=model_name, temperature=temperature, callbacks=[handler])
+            # Interpolate between base and highest
+            target_elo = base_elo + ((strength - 0.5) / 0.5) * (highest_elo - base_elo)
+
+        available_df['sort_metric'] = abs(available_df['coding_arena_elo'] - target_elo)
+        candidates = available_df.sort_values(by='sort_metric').to_dict('records')
+        target_metric_value = f"Target ELO: {target_elo:.2f}"
+
+
+    if not candidates:
+         # This should ideally not happen if available_df was not empty
+         raise RuntimeError("Model selection resulted in an empty candidate list.")
+
+    # --- DEBUGGING PRINT ---
+    if os.getenv("PDD_DEBUG_SELECTOR"): # Add env var check for debug prints
+        print("\n--- DEBUG: _select_model_candidates ---")
+        print(f"Strength: {strength}, Base Model: {base_model_name}")
+        print(f"Metric: {target_metric_value}")
+        print("Available DF (Sorted by metric):")
+        # Select columns relevant to the sorting metric
+        sort_cols = ['model', 'avg_cost', 'coding_arena_elo', 'sort_metric']
+        print(available_df.sort_values(by='sort_metric')[sort_cols])
+        print("Final Candidates List (Model Names):")
+        print([c['model'] for c in candidates])
+        print("---------------------------------------\n")
+    # --- END DEBUGGING PRINT ---
+
+    return candidates
+
+
+def _ensure_api_key(model_info: Dict[str, Any], newly_acquired_keys: Dict[str, bool], verbose: bool) -> bool:
+    """Checks for API key in env, prompts user if missing, and updates .env."""
+    key_name = model_info.get('api_key')
+
+    if not key_name or key_name == "EXISTING_KEY":
+        if verbose:
+            rprint(f"[INFO] Skipping API key check for model {model_info.get('model')} (key name: {key_name})")
+        return True # Assume key is handled elsewhere or not needed
+
+    key_value = os.getenv(key_name)
+
+    if key_value:
+        if verbose:
+            rprint(f"[INFO] API key '{key_name}' found in environment.")
+        newly_acquired_keys[key_name] = False # Mark as existing
+        return True
     else:
-        raise ValueError(f"Unsupported provider: {selected_model.provider}")
+        rprint(f"[WARN] API key environment variable '{key_name}' for model '{model_info.get('model')}' is not set.")
+        try:
+            # Interactive prompt
+            user_provided_key = input(f"Please enter the API key for {key_name}: ").strip()
+            if not user_provided_key:
+                rprint("[ERROR] No API key provided. Cannot proceed with this model.")
+                return False
+
+            # Set environment variable for the current process
+            os.environ[key_name] = user_provided_key
+            rprint(f"[INFO] API key '{key_name}' set for the current session.")
+            newly_acquired_keys[key_name] = True # Mark as newly acquired
+
+            # Update .env file
+            try:
+                lines = []
+                if ENV_PATH.exists():
+                    with open(ENV_PATH, 'r') as f:
+                        lines = f.readlines()
+
+                new_lines = []
+                # key_updated = False
+                prefix = f"{key_name}="
+                prefix_spaced = f"{key_name} =" # Handle potential spaces
+
+                for line in lines:
+                    stripped_line = line.strip()
+                    if stripped_line.startswith(prefix) or stripped_line.startswith(prefix_spaced):
+                        # Comment out the old key
+                        new_lines.append(f"# {line}")
+                        # key_updated = True # Indicates we found an old line to comment
+                    elif stripped_line.startswith(f"# {prefix}") or stripped_line.startswith(f"# {prefix_spaced}"):
+                         # Keep already commented lines as they are
+                         new_lines.append(line)
+                    else:
+                        new_lines.append(line)
+
+                # Append the new key, ensuring quotes for robustness
+                new_key_line = f'{key_name}="{user_provided_key}"\n'
+                # Add newline before if file not empty and doesn't end with newline
+                if new_lines and not new_lines[-1].endswith('\n'):
+                     new_lines.append('\n')
+                new_lines.append(new_key_line)
 
-    if max_completion_tokens:
-        llm.model_kwargs = {"max_completion_tokens": max_completion_tokens}
-    elif max_tokens:
-        if provider == 'google' or provider == 'googlevertexai':
-            llm.max_output_tokens = max_tokens
+
+                with open(ENV_PATH, 'w') as f:
+                    f.writelines(new_lines)
+
+                rprint(f"[INFO] API key '{key_name}' saved to {ENV_PATH}.")
+                rprint("[bold yellow]SECURITY WARNING:[/bold yellow] The API key has been saved to your .env file. "
+                       "Ensure this file is kept secure and is included in your .gitignore.")
+
+            except IOError as e:
+                rprint(f"[ERROR] Failed to update .env file at {ENV_PATH}: {e}")
+                # Continue since the key is set in the environment for this session
+
+            return True
+
+        except EOFError: # Handle non-interactive environments
+             rprint(f"[ERROR] Cannot prompt for API key '{key_name}' in a non-interactive environment.")
+             return False
+        except Exception as e:
+             rprint(f"[ERROR] An unexpected error occurred during API key acquisition: {e}")
+             return False
+
+
+def _format_messages(prompt: str, input_data: Union[Dict[str, Any], List[Dict[str, Any]]], use_batch_mode: bool) -> Union[List[Dict[str, str]], List[List[Dict[str, str]]]]:
+    """Formats prompt and input into LiteLLM message format."""
+    try:
+        prompt_template = PromptTemplate.from_template(prompt)
+        if use_batch_mode:
+            if not isinstance(input_data, list):
+                raise ValueError("input_json must be a list of dictionaries when use_batch_mode is True.")
+            all_messages = []
+            for item in input_data:
+                if not isinstance(item, dict):
+                     raise ValueError("Each item in input_json list must be a dictionary for batch mode.")
+                formatted_prompt = prompt_template.format(**item)
+                all_messages.append([{"role": "user", "content": formatted_prompt}])
+            return all_messages
         else:
-            llm.max_tokens = max_tokens
-    return llm
+            if not isinstance(input_data, dict):
+                raise ValueError("input_json must be a dictionary when use_batch_mode is False.")
+            formatted_prompt = prompt_template.format(**input_data)
+            return [{"role": "user", "content": formatted_prompt}]
+    except KeyError as e:
+        raise ValueError(f"Prompt formatting error: Missing key {e} in input_json for prompt template.") from e
+    except Exception as e:
+        raise ValueError(f"Error formatting prompt: {e}") from e
 
-def calculate_cost(handler, selected_model):
-    """
-    Calculates the cost of the invoke run based on token usage.
+# --- Main Function ---
+
+def llm_invoke(
+    prompt: Optional[str] = None,
+    input_json: Optional[Union[Dict[str, Any], List[Dict[str, Any]]]] = None,
+    strength: float = 0.5, # Use pdd.DEFAULT_STRENGTH if available, else 0.5
+    temperature: float = 0.1,
+    verbose: bool = False,
+    output_pydantic: Optional[Type[BaseModel]] = None,
+    time: float = 0.25,
+    use_batch_mode: bool = False,
+    messages: Optional[Union[List[Dict[str, str]], List[List[Dict[str, str]]]]] = None,
+) -> Dict[str, Any]:
     """
-    input_tokens = handler.input_tokens or 0
-    output_tokens = handler.output_tokens or 0
-    input_cost = selected_model.input_cost
-    output_cost = selected_model.output_cost
-    total_cost = (input_tokens / 1_000_000) * input_cost + (output_tokens / 1_000_000) * output_cost
-    return total_cost
+    Runs a prompt with given input using LiteLLM, handling model selection,
+    API key acquisition, structured output, batching, and reasoning time.
+    The maximum completion token length defaults to the provider's maximum.
 
-# ---------------- Main Function ---------------- #
+    Args:
+        prompt: Prompt template string (required if messages is None).
+        input_json: Dictionary or list of dictionaries for prompt variables (required if messages is None).
+        strength: Model selection strength (0=cheapest, 0.5=base, 1=highest ELO).
+        temperature: LLM temperature.
+        verbose: Print detailed logs.
+        output_pydantic: Optional Pydantic model for structured output.
+        time: Relative thinking time (0-1, default 0.25).
+        use_batch_mode: Use batch completion if True.
+        messages: Pre-formatted list of messages (or list of lists for batch). If provided, ignores prompt and input_json.
 
-def llm_invoke(prompt, input_json, strength, temperature, verbose=False, output_pydantic=None):
-    """
-    Invokes an LLM chain with the provided prompt and input_json, using a model selected based on the strength parameter.
-    
-    Inputs:
-        prompt (str): The prompt template as a string.
-        input_json (dict): JSON object containing inputs for the prompt.
-        strength (float): 0 (cheapest) to 1 (highest ELO); 0.5 uses the base model.
-        temperature (float): Temperature for the LLM invocation.
-        verbose (bool): When True, prints detailed information.
-        output_pydantic (Optional): A Pydantic model class for structured output.
-    
-    Output (dict): Contains:
-        'result' - LLM output (string or parsed Pydantic object).
-        'cost' - Calculated cost of the invoke run.
-        'model_name' - Name of the selected model that succeeded.
+    Returns:
+        Dictionary containing 'result', 'cost', 'model_name', 'thinking_output'.
+
+    Raises:
+        ValueError: For invalid inputs or prompt formatting errors.
+        FileNotFoundError: If llm_model.csv is missing.
+        RuntimeError: If all candidate models fail.
+        openai.*Error: If LiteLLM encounters API errors after retries.
     """
-    if prompt is None or not isinstance(prompt, str):
-        raise ValueError("Prompt is required.")
-    if input_json is None:
-        raise ValueError("Input JSON is required.")
-    if not isinstance(input_json, dict):
-        raise ValueError("Input JSON must be a dictionary.")
-
-    set_llm_cache(SQLiteCache(database_path=".langchain.db"))
-    base_model_name = os.environ.get('PDD_MODEL_DEFAULT', 'gpt-4.1-nano')
-    models = load_models()
-    
-    try:
-        base_model = select_model(models, base_model_name)
-    except ValueError as e:
-        raise RuntimeError(f"Base model error: {str(e)}") from e
+    if verbose: # Print args early if verbose
+        rprint("[DEBUG llm_invoke start] Arguments received:")
+        rprint(f"  prompt: {'provided' if prompt else 'None'}")
+        rprint(f"  input_json: {'provided' if input_json is not None else 'None'}")
+        rprint(f"  strength: {strength}")
+        rprint(f"  temperature: {temperature}")
+        rprint(f"  verbose: {verbose}")
+        rprint(f"  output_pydantic: {output_pydantic.__name__ if output_pydantic else 'None'}")
+        rprint(f"  time: {time}")
+        rprint(f"  use_batch_mode: {use_batch_mode}")
+        rprint(f"  messages: {'provided' if messages else 'None'}")
 
-    candidate_models = get_candidate_models(strength, models, base_model)
+    # --- 1. Load Environment & Validate Inputs ---
+    # .env loading happens at module level
 
-    if verbose:
-        rprint(f"[bold cyan]Candidate models (in order):[/bold cyan] {[m.model for m in candidate_models]}")
+    if messages:
+        if verbose:
+            rprint("[INFO] Using provided 'messages' input.")
+        # Basic validation of messages format
+        if use_batch_mode:
+            if not isinstance(messages, list) or not all(isinstance(m_list, list) for m_list in messages):
+                 raise ValueError("'messages' must be a list of lists when use_batch_mode is True.")
+            if not all(isinstance(msg, dict) and 'role' in msg and 'content' in msg for m_list in messages for msg in m_list):
+                 raise ValueError("Each message in the lists within 'messages' must be a dictionary with 'role' and 'content'.")
+        else:
+            if not isinstance(messages, list) or not all(isinstance(msg, dict) and 'role' in msg and 'content' in msg for msg in messages):
+                 raise ValueError("'messages' must be a list of dictionaries with 'role' and 'content'.")
+        formatted_messages = messages
+    elif prompt and input_json is not None:
+         if not isinstance(prompt, str) or not prompt:
+             raise ValueError("'prompt' must be a non-empty string when 'messages' is not provided.")
+         formatted_messages = _format_messages(prompt, input_json, use_batch_mode)
+    else:
+        raise ValueError("Either 'messages' or both 'prompt' and 'input_json' must be provided.")
+
+    if not (0.0 <= strength <= 1.0):
+        raise ValueError("'strength' must be between 0.0 and 1.0.")
+    if not (0.0 <= temperature <= 2.0): # Common range for temperature
+        warnings.warn("'temperature' is outside the typical range (0.0-2.0).")
+    if not (0.0 <= time <= 1.0):
+        raise ValueError("'time' must be between 0.0 and 1.0.")
+
+    # --- 2. Load Model Data & Select Candidates ---
+    try:
+        model_df = _load_model_data(LLM_MODEL_CSV_PATH)
+        candidate_models = _select_model_candidates(strength, DEFAULT_BASE_MODEL, model_df)
+    except (FileNotFoundError, ValueError, RuntimeError) as e:
+        rprint(f"[ERROR] Failed during model loading or selection: {e}")
+        raise
 
-    last_error = None
-    for model in candidate_models:
-        handler = CompletionStatusHandler()
+    if verbose:
+        # This print statement is crucial for the verbose test
+        # Calculate and print strength for each candidate model
+        # Find min/max for cost and ELO
+        min_cost = model_df['avg_cost'].min()
+        max_elo = model_df['coding_arena_elo'].max()
+        base_cost = model_df[model_df['model'] == DEFAULT_BASE_MODEL]['avg_cost'].iloc[0] if not model_df[model_df['model'] == DEFAULT_BASE_MODEL].empty else min_cost
+        base_elo = model_df[model_df['model'] == DEFAULT_BASE_MODEL]['coding_arena_elo'].iloc[0] if not model_df[model_df['model'] == DEFAULT_BASE_MODEL].empty else max_elo
+        
+        def calc_strength(candidate):
+            # If strength < 0.5, interpolate by cost (cheaper = 0, base = 0.5)
+            # If strength > 0.5, interpolate by ELO (base = 0.5, highest = 1.0)
+            avg_cost = candidate.get('avg_cost', min_cost)
+            elo = candidate.get('coding_arena_elo', base_elo)
+            if strength < 0.5:
+                # Map cost to [0, 0.5]
+                if base_cost == min_cost:
+                    return 0.5 # Avoid div by zero
+                rel = (avg_cost - min_cost) / (base_cost - min_cost)
+                return max(0.0, min(0.5, rel * 0.5))
+            elif strength > 0.5:
+                # Map ELO to [0.5, 1.0]
+                if max_elo == base_elo:
+                    return 0.5 # Avoid div by zero
+                rel = (elo - base_elo) / (max_elo - base_elo)
+                return max(0.5, min(1.0, 0.5 + rel * 0.5))
+            else:
+                return 0.5
+        
+        model_strengths_formatted = [(c['model'], f"{float(calc_strength(c)):.3f}") for c in candidate_models]
+        rprint("[INFO] Candidate models selected and ordered (with strength):", model_strengths_formatted)
+        rprint(f"[INFO] Strength: {strength}, Temperature: {temperature}, Time: {time}")
+        if use_batch_mode:
+            rprint("[INFO] Batch mode enabled.")
+        if output_pydantic:
+            rprint(f"[INFO] Pydantic output requested: {output_pydantic.__name__}")
         try:
-            try:
-                prompt_template = PromptTemplate.from_template(prompt)
-            except ValueError:
-                raise ValueError("Invalid prompt template")
+            # Only print input_json if it was actually provided (not when messages were used)
+            if input_json is not None:
+                rprint("[INFO] Input JSON:")
+                rprint(input_json)
+            else:
+                 rprint("[INFO] Input: Using pre-formatted 'messages'.")
+        except Exception:
+            print("[INFO] Input JSON/Messages (fallback print):") # Fallback for complex objects rich might fail on
+            print(input_json if input_json is not None else "[Messages provided directly]")
 
-            llm = create_llm_instance(model, temperature, handler)
+
+    # --- 3. Iterate Through Candidates and Invoke LLM ---
+    last_exception = None
+    newly_acquired_keys: Dict[str, bool] = {} # Track keys obtained in this run
+
+    for model_info in candidate_models:
+        model_name_litellm = model_info['model']
+        api_key_name = model_info.get('api_key')
+        provider = model_info.get('provider', '').lower()
+
+        if verbose:
+            rprint(f"\n[ATTEMPT] Trying model: {model_name_litellm} (Provider: {provider})")
+
+        retry_with_same_model = True
+        while retry_with_same_model:
+            retry_with_same_model = False # Assume success unless auth error on new key
+
+            # --- 4. API Key Check & Acquisition ---
+            if not _ensure_api_key(model_info, newly_acquired_keys, verbose):
+                # Problem getting key, break inner loop, try next model candidate
+                if verbose:
+                    rprint(f"[SKIP] Skipping {model_name_litellm} due to API key/credentials issue after prompt.")
+                break # Breaks the 'while retry_with_same_model' loop
+
+            # --- 5. Prepare LiteLLM Arguments ---
+            litellm_kwargs: Dict[str, Any] = {
+                "model": model_name_litellm,
+                "messages": formatted_messages,
+                "temperature": temperature,
+            }
+
+            api_key_name_from_csv = model_info.get('api_key') # From CSV
+            # Determine if it's a Vertex AI model for special handling
+            is_vertex_model = (provider.lower() == 'google') or \
+                              (provider.lower() == 'googlevertexai') or \
+                              (provider.lower() == 'vertex_ai') or \
+                              model_name_litellm.startswith('vertex_ai/')
+
+            if is_vertex_model and api_key_name_from_csv == 'VERTEX_CREDENTIALS':
+                credentials_file_path = os.getenv("VERTEX_CREDENTIALS") # Path from env var
+                vertex_project_env = os.getenv("VERTEX_PROJECT")
+                vertex_location_env = os.getenv("VERTEX_LOCATION")
+
+                if credentials_file_path and vertex_project_env and vertex_location_env:
+                    try:
+                        with open(credentials_file_path, 'r') as f:
+                            loaded_credentials = json.load(f)
+                        vertex_credentials_json_string = json.dumps(loaded_credentials)
+                        
+                        litellm_kwargs["vertex_credentials"] = vertex_credentials_json_string
+                        litellm_kwargs["vertex_project"] = vertex_project_env
+                        litellm_kwargs["vertex_location"] = vertex_location_env
+                        if verbose:
+                            rprint(f"[INFO] For Vertex AI: using vertex_credentials from '{credentials_file_path}', project '{vertex_project_env}', location '{vertex_location_env}'.")
+                    except FileNotFoundError:
+                        if verbose:
+                            rprint(f"[ERROR] Vertex credentials file not found at path specified by VERTEX_CREDENTIALS env var: '{credentials_file_path}'. LiteLLM may try ADC or fail.")
+                    except json.JSONDecodeError:
+                        if verbose:
+                            rprint(f"[ERROR] Failed to decode JSON from Vertex credentials file: '{credentials_file_path}'. Check file content. LiteLLM may try ADC or fail.")
+                    except Exception as e:
+                        if verbose:
+                            rprint(f"[ERROR] Failed to load or process Vertex credentials from '{credentials_file_path}': {e}. LiteLLM may try ADC or fail.")
+                else:
+                    if verbose:
+                        rprint(f"[WARN] For Vertex AI (using '{api_key_name_from_csv}'): One or more required environment variables (VERTEX_CREDENTIALS, VERTEX_PROJECT, VERTEX_LOCATION) are missing.")
+                        if not credentials_file_path: rprint(f"  Reason: VERTEX_CREDENTIALS (path to JSON file) env var not set or empty.")
+                        if not vertex_project_env: rprint(f"  Reason: VERTEX_PROJECT env var not set or empty.")
+                        if not vertex_location_env: rprint(f"  Reason: VERTEX_LOCATION env var not set or empty.")
+                        rprint(f"  LiteLLM may attempt to use Application Default Credentials or the call may fail.")
+
+            elif api_key_name_from_csv: # For other api_key_names specified in CSV (e.g., OPENAI_API_KEY, or a direct VERTEX_AI_API_KEY string)
+                key_value = os.getenv(api_key_name_from_csv)
+                if key_value:
+                    litellm_kwargs["api_key"] = key_value
+                    if verbose:
+                        rprint(f"[INFO] Explicitly passing API key from env var '{api_key_name_from_csv}' as 'api_key' parameter to LiteLLM.")
+                    
+                    # If this model is Vertex AI AND uses a direct API key string (not VERTEX_CREDENTIALS from CSV),
+                    # also pass project and location from env vars.
+                    if is_vertex_model: 
+                        vertex_project_env = os.getenv("VERTEX_PROJECT")
+                        vertex_location_env = os.getenv("VERTEX_LOCATION")
+                        if vertex_project_env and vertex_location_env:
+                            litellm_kwargs["vertex_project"] = vertex_project_env
+                            litellm_kwargs["vertex_location"] = vertex_location_env
+                            if verbose:
+                                rprint(f"[INFO] For Vertex AI model (using direct API key '{api_key_name_from_csv}'), also passing vertex_project='{vertex_project_env}' and vertex_location='{vertex_location_env}' from env vars.")
+                        elif verbose:
+                             rprint(f"[WARN] For Vertex AI model (using direct API key '{api_key_name_from_csv}'), VERTEX_PROJECT or VERTEX_LOCATION env vars not set. This might be required by LiteLLM.")
+                elif verbose: # api_key_name_from_csv was in CSV, but corresponding env var was not set/empty
+                    rprint(f"[WARN] API key name '{api_key_name_from_csv}' found in CSV, but the environment variable '{api_key_name_from_csv}' is not set or empty. LiteLLM will use default authentication if applicable (e.g., other standard env vars or ADC).")
+            
+            elif verbose: # No api_key_name_from_csv in CSV for this model
+                rprint(f"[INFO] No API key name specified in CSV for model '{model_name_litellm}'. LiteLLM will use its default authentication mechanisms (e.g., standard provider env vars or ADC for Vertex AI).")
+
+            # Add api_base if present in CSV
+            api_base = model_info.get('base_url')
+            if pd.notna(api_base) and api_base:
+                litellm_kwargs["api_base"] = str(api_base)
+
+            # Handle Structured Output (JSON Mode / Pydantic)
             if output_pydantic:
-                if model.structured_output:
-                    llm = llm.with_structured_output(output_pydantic)
-                    chain = prompt_template | llm
+                # Check if model supports structured output based on CSV flag or LiteLLM check
+                supports_structured = model_info.get('structured_output', False)
+                # Optional: Add litellm.supports_response_schema check if CSV flag is unreliable
+                # if not supports_structured:
+                #     try: supports_structured = litellm.supports_response_schema(model=model_name_litellm)
+                #     except: pass # Ignore errors in supports_response_schema check
+
+                if supports_structured:
+                    if verbose:
+                        rprint(f"[INFO] Requesting structured output (Pydantic: {output_pydantic.__name__}) for {model_name_litellm}")
+                    # Pass the Pydantic model directly if supported, else use json_object
+                    # LiteLLM handles passing Pydantic models for supported providers
+                    litellm_kwargs["response_format"] = output_pydantic
+                    # As a fallback, one could use:
+                    # litellm_kwargs["response_format"] = {"type": "json_object"}
+                    # And potentially enable client-side validation:
+                    # litellm.enable_json_schema_validation = True # Enable globally if needed
                 else:
-                    parser = PydanticOutputParser(pydantic_object=output_pydantic)
-                    chain = prompt_template | llm | parser
-            else:
-                chain = prompt_template | llm | StrOutputParser()
-
-            result_output = chain.invoke(input_json)
-            cost = calculate_cost(handler, model)
-
-            if verbose:
-                rprint(f"[bold green]Selected model: {model.model}[/bold green]")
-                rprint(f"Per input token cost: ${model.input_cost} per million tokens")
-                rprint(f"Per output token cost: ${model.output_cost} per million tokens")
-                rprint(f"Number of input tokens: {handler.input_tokens}")
-                rprint(f"Number of output tokens: {handler.output_tokens}")
-                rprint(f"Cost of invoke run: ${cost:.0e}")
-                rprint(f"Strength used: {strength}")
-                rprint(f"Temperature used: {temperature}")
-                try:
-                    # Try printing with rich formatting first
-                    rprint(f"Input JSON: {str(input_json)}")
-                except MarkupError:
-                    # Fallback to standard print if rich markup fails
-                    print(f"Input JSON: {str(input_json)}")
-                except Exception:
-                    print(f"Input JSON: {input_json}")
-                if output_pydantic:
-                    rprint(f"Output Pydantic format: {output_pydantic}")
-                try:
-                    # Try printing with rich formatting first
-                    rprint(f"Result: {result_output}")
-                except MarkupError as me:
-                    # Fallback to standard print if rich markup fails
-                    print(f"[bold yellow]Warning:[/bold yellow] Failed to render result with rich markup: {me}")
-                    print(f"Raw Result: {str(result_output)}") # Use standard print
-
-            return {'result': result_output, 'cost': cost, 'model_name': model.model}
+                    if verbose:
+                        rprint(f"[WARN] Model {model_name_litellm} does not support structured output via CSV flag. Output might not be valid {output_pydantic.__name__}.")
+                    # Proceed without forcing JSON mode, parsing will be attempted later
 
-        except Exception as e:
-            last_error = e
-            if verbose:
-                rprint(f"[red]Error with model {model.model}: {str(e)}[/red]")
-            continue
+            # --- NEW REASONING LOGIC ---
+            reasoning_type = model_info.get('reasoning_type', 'none') # Defaults to 'none'
+            max_reasoning_tokens_val = model_info.get('max_reasoning_tokens', 0) # Defaults to 0
+
+            if time > 0: # Only apply reasoning if time is requested
+                if reasoning_type == 'budget':
+                    if max_reasoning_tokens_val > 0:
+                        budget = int(time * max_reasoning_tokens_val)
+                        if budget > 0:
+                            # Currently known: Anthropic uses 'thinking'
+                            # Model name comparison is more robust than provider string
+                            if provider == 'anthropic': # Check provider column instead of model prefix
+                                litellm_kwargs["thinking"] = {"type": "enabled", "budget_tokens": budget}
+                                if verbose:
+                                    rprint(f"[INFO] Requesting Anthropic thinking (budget type) with budget: {budget} tokens for {model_name_litellm}")
+                            else:
+                                # If other providers adopt a budget param recognized by LiteLLM, add here
+                                if verbose:
+                                    rprint(f"[WARN] Reasoning type is 'budget' for {model_name_litellm}, but no specific LiteLLM budget parameter known for this provider. Parameter not sent.")
+                        elif verbose:
+                            rprint(f"[INFO] Calculated reasoning budget is 0 for {model_name_litellm}, skipping reasoning parameter.")
+                    elif verbose:
+                        rprint(f"[WARN] Reasoning type is 'budget' for {model_name_litellm}, but 'max_reasoning_tokens' is missing or zero in CSV. Reasoning parameter not sent.")
+
+                elif reasoning_type == 'effort':
+                    effort = "low"
+                    if time > 0.7:
+                        effort = "high"
+                    elif time > 0.3:
+                        effort = "medium"
+                    # Use the common 'reasoning_effort' param LiteLLM provides
+                    litellm_kwargs["reasoning_effort"] = effort
+                    if verbose:
+                        rprint(f"[INFO] Requesting reasoning_effort='{effort}' (effort type) for {model_name_litellm} based on time={time}")
+
+                elif reasoning_type == 'none':
+                    if verbose:
+                        rprint(f"[INFO] Model {model_name_litellm} has reasoning_type='none'. No reasoning parameter sent.")
+
+                else: # Unknown reasoning_type in CSV
+                     if verbose:
+                         rprint(f"[WARN] Unknown reasoning_type '{reasoning_type}' for model {model_name_litellm} in CSV. No reasoning parameter sent.")
+
+            # --- END NEW REASONING LOGIC ---
+
+            # Add caching control per call if needed (example: force refresh)
+            # litellm_kwargs["cache"] = {"no-cache": True}
+
+            # --- 6. LLM Invocation ---
+            try:
+                start_time = time_module.time()
+
+                # --- ADDED CACHE STATUS DEBUGGING (NOW UNCONDITIONAL) ---
+                print(f"[DEBUG llm_invoke] Cache Check: litellm.cache is None: {litellm.cache is None}") # MODIFIED: unconditional print
+                if litellm.cache is not None:
+                    print(f"[DEBUG llm_invoke] litellm.cache type: {type(litellm.cache)}, ID: {id(litellm.cache)}") # MODIFIED: unconditional print
+                # --- END ADDED CACHE STATUS DEBUGGING ---
 
-    if isinstance(last_error, ValueError) and "Invalid prompt template" in str(last_error):
-        raise ValueError("Invalid prompt template")
-    if last_error:
-        raise RuntimeError(f"Error during LLM invocation: {str(last_error)}")
-    raise RuntimeError("No available models could process the request")
+                # <<< EXPLICITLY ENABLE CACHING >>>
+                # Only add if litellm.cache is configured
+                if litellm.cache is not None:
+                    litellm_kwargs["caching"] = True
+                else: # MODIFIED: unconditional print for this path too
+                    print(f"[DEBUG llm_invoke] NOT ENABLING CACHING: litellm.cache is None at call time.")
 
+
+                if use_batch_mode:
+                    if verbose:
+                        rprint(f"[INFO] Calling litellm.batch_completion for {model_name_litellm}...")
+                    response = litellm.batch_completion(**litellm_kwargs)
+
+
+                else:
+                    if verbose:
+                        rprint(f"[INFO] Calling litellm.completion for {model_name_litellm}...")
+                    response = litellm.completion(**litellm_kwargs)
+
+                end_time = time_module.time()
+
+                if verbose:
+                    rprint(f"[SUCCESS] Invocation successful for {model_name_litellm} (took {end_time - start_time:.2f}s)")
+
+                # --- 7. Process Response ---
+                results = []
+                thinking_outputs = []
+
+                response_list = response if use_batch_mode else [response]
+
+                for i, resp_item in enumerate(response_list):
+                    # Cost calculation is handled entirely by the success callback
+
+                    # Thinking Output
+                    thinking = None
+                    try:
+                        # Attempt 1: Check _hidden_params based on isolated test script
+                        if hasattr(resp_item, '_hidden_params') and resp_item._hidden_params and 'thinking' in resp_item._hidden_params:
+                             thinking = resp_item._hidden_params['thinking']
+                             if verbose:
+                                 rprint("[DEBUG] Extracted thinking output from response._hidden_params['thinking']")
+                        # Attempt 2: Fallback to reasoning_content in message
+                        # Use .get() for safer access
+                        elif hasattr(resp_item, 'choices') and resp_item.choices and hasattr(resp_item.choices[0], 'message') and hasattr(resp_item.choices[0].message, 'get') and resp_item.choices[0].message.get('reasoning_content'):
+                            thinking = resp_item.choices[0].message.get('reasoning_content')
+                            if verbose:
+                                rprint("[DEBUG] Extracted thinking output from response.choices[0].message.get('reasoning_content')")
+
+                    except (AttributeError, IndexError, KeyError, TypeError):
+                        if verbose:
+                            rprint("[DEBUG] Failed to extract thinking output from known locations.")
+                        pass # Ignore if structure doesn't match or errors occur
+                    thinking_outputs.append(thinking)
+
+                    # Result (String or Pydantic)
+                    try:
+                        raw_result = resp_item.choices[0].message.content
+                        if output_pydantic:
+                            parsed_result = None
+                            json_string_to_parse = None
+
+                            try:
+                                # Attempt 1: Check if LiteLLM already parsed it
+                                if isinstance(raw_result, output_pydantic):
+                                    parsed_result = raw_result
+                                    if verbose:
+                                        rprint("[DEBUG] Pydantic object received directly from LiteLLM.")
+
+                                # Attempt 2: Check if raw_result is dict-like and validate
+                                elif isinstance(raw_result, dict):
+                                    parsed_result = output_pydantic.model_validate(raw_result)
+                                    if verbose:
+                                        rprint("[DEBUG] Validated dictionary-like object directly.")
+
+                                # Attempt 3: Process as string (if not already parsed/validated)
+                                elif isinstance(raw_result, str):
+                                    json_string_to_parse = raw_result # Start with the raw string
+                                    try:
+                                        # Look for first { and last }
+                                        start_brace = json_string_to_parse.find('{')
+                                        end_brace = json_string_to_parse.rfind('}')
+                                        if start_brace != -1 and end_brace != -1 and end_brace > start_brace:
+                                            potential_json = json_string_to_parse[start_brace:end_brace+1]
+                                            # Basic check if it looks like JSON
+                                            if potential_json.strip().startswith('{') and potential_json.strip().endswith('}'):
+                                                if verbose:
+                                                    rprint(f"[DEBUG] Attempting to parse extracted JSON block: '{potential_json}'")
+                                                parsed_result = output_pydantic.model_validate_json(potential_json)
+                                            else:
+                                                # If block extraction fails, try cleaning markdown next
+                                                raise ValueError("Extracted block doesn't look like JSON")
+                                        else:
+                                             # If no braces found, try cleaning markdown next
+                                            raise ValueError("Could not find enclosing {}")
+                                    except (json.JSONDecodeError, ValidationError, ValueError) as extraction_error:
+                                        if verbose:
+                                            rprint(f"[DEBUG] JSON block extraction/validation failed ('{extraction_error}'). Trying markdown cleaning.")
+                                        # Fallback: Clean markdown fences and retry JSON validation
+                                        cleaned_result_str = raw_result.strip()
+                                        if cleaned_result_str.startswith("```json"):
+                                            cleaned_result_str = cleaned_result_str[7:]
+                                        elif cleaned_result_str.startswith("```"):
+                                            cleaned_result_str = cleaned_result_str[3:]
+                                        if cleaned_result_str.endswith("```"):
+                                            cleaned_result_str = cleaned_result_str[:-3]
+                                        cleaned_result_str = cleaned_result_str.strip()
+                                        # Check again if it looks like JSON before parsing
+                                        if cleaned_result_str.startswith('{') and cleaned_result_str.endswith('}'):
+                                            if verbose:
+                                                rprint(f"[DEBUG] Attempting parse after cleaning markdown fences. Cleaned string: '{cleaned_result_str}'")
+                                            json_string_to_parse = cleaned_result_str # Update string for error reporting
+                                            parsed_result = output_pydantic.model_validate_json(json_string_to_parse)
+                                        else:
+                                            # If still doesn't look like JSON, raise error
+                                            raise ValueError("Content after cleaning markdown doesn't look like JSON")
+
+
+                                # Check if any parsing attempt succeeded
+                                if parsed_result is None:
+                                    # This case should ideally be caught by exceptions above, but as a safeguard:
+                                    raise TypeError(f"Raw result type {type(raw_result)} or content could not be validated/parsed against {output_pydantic.__name__}.")
+
+                            except (ValidationError, json.JSONDecodeError, TypeError, ValueError) as parse_error:
+                                rprint(f"[ERROR] Failed to parse response into Pydantic model {output_pydantic.__name__} for item {i}: {parse_error}")
+                                # Use the string that was last attempted for parsing in the error message
+                                error_content = json_string_to_parse if json_string_to_parse is not None else raw_result
+                                rprint("[ERROR] Content attempted for parsing:", repr(error_content)) # Use repr for clarity
+                                results.append(f"ERROR: Failed to parse Pydantic. Raw: {repr(raw_result)}")
+                                continue # Skip appending result below if parsing failed
+
+                            # If parsing succeeded, append the parsed_result
+                            results.append(parsed_result)
+
+                        else:
+                            # If output_pydantic was not requested, append the raw result
+                            results.append(raw_result)
+
+                    except (AttributeError, IndexError) as e:
+                         rprint(f"[ERROR] Could not extract result content from response item {i}: {e}")
+                         results.append(f"ERROR: Could not extract result content. Response: {resp_item}")
+
+                # --- Retrieve Cost from Callback Data --- (Reinstated)
+                # For batch, this will reflect the cost associated with the *last* item processed by the callback.
+                # A fully accurate batch total would require a more complex callback class to aggregate.
+                total_cost = _LAST_CALLBACK_DATA.get("cost", 0.0)
+                # ----------------------------------------
+
+                final_result = results if use_batch_mode else results[0]
+                final_thinking = thinking_outputs if use_batch_mode else thinking_outputs[0]
+
+                # --- Verbose Output for Success ---
+                if verbose:
+                    # Get token usage from the *last* callback data (might not be accurate for batch)
+                    input_tokens = _LAST_CALLBACK_DATA.get("input_tokens", 0)
+                    output_tokens = _LAST_CALLBACK_DATA.get("output_tokens", 0)
+
+                    cost_input_pm = model_info.get('input', 0.0) if pd.notna(model_info.get('input')) else 0.0
+                    cost_output_pm = model_info.get('output', 0.0) if pd.notna(model_info.get('output')) else 0.0
+
+                    rprint(f"[RESULT] Model Used: {model_name_litellm}")
+                    rprint(f"[RESULT] Cost (Input): ${cost_input_pm:.2f}/M tokens")
+                    rprint(f"[RESULT] Cost (Output): ${cost_output_pm:.2f}/M tokens")
+                    rprint(f"[RESULT] Tokens (Prompt): {input_tokens}")
+                    rprint(f"[RESULT] Tokens (Completion): {output_tokens}")
+                    # Display the cost captured by the callback
+                    rprint(f"[RESULT] Total Cost (from callback): ${total_cost:.6g}") # Renamed label for clarity
+                    rprint("[RESULT] Max Completion Tokens: Provider Default") # Indicate default limit
+                    if final_thinking:
+                        rprint("[RESULT] Thinking Output:")
+                        rprint(final_thinking) # Rich print should handle the thinking output format
+
+                # --- Print raw output before returning if verbose ---
+                if verbose:
+                    rprint("[DEBUG] Raw output before return:")
+                    print(f"  Raw Result (repr): {repr(final_result)}")
+                    print(f"  Raw Thinking (repr): {repr(final_thinking)}")
+                    rprint("-" * 20) # Separator
+
+                # --- Return Success ---
+                return {
+                    'result': final_result,
+                    'cost': total_cost,
+                    'model_name': model_name_litellm, # Actual model used
+                    'thinking_output': final_thinking if final_thinking else None
+                }
+
+            # --- 6b. Handle Invocation Errors ---
+            except openai.AuthenticationError as e:
+                last_exception = e
+                if newly_acquired_keys.get(api_key_name):
+                    rprint(f"[AUTH ERROR] Authentication failed for {model_name_litellm} with the newly provided key for '{api_key_name}'. Please check the key and try again.")
+                    # Invalidate the key in env for this session to force re-prompt on retry
+                    if api_key_name in os.environ:
+                         del os.environ[api_key_name]
+                    # Clear the 'newly acquired' status for this key so the next attempt doesn't trigger immediate retry loop
+                    newly_acquired_keys[api_key_name] = False
+                    retry_with_same_model = True # Set flag to retry the same model after re-prompt
+                    # Go back to the start of the 'while retry_with_same_model' loop
+                else:
+                    rprint(f"[AUTH ERROR] Authentication failed for {model_name_litellm} using existing key '{api_key_name}'. Trying next model.")
+                    break # Break inner loop, try next model candidate
+
+            except (openai.RateLimitError, openai.APITimeoutError, openai.APIConnectionError,
+                    openai.APIStatusError, openai.BadRequestError, openai.InternalServerError,
+                    Exception) as e:
+                last_exception = e
+                error_type = type(e).__name__
+                rprint(f"[ERROR] Invocation failed for {model_name_litellm} ({error_type}): {e}. Trying next model.")
+                # Log more details in verbose mode
+                if verbose:
+                    import traceback
+                    traceback.print_exc()
+                break # Break inner loop, try next model candidate
+
+        # If the inner loop was broken (not by success), continue to the next candidate model
+        continue
+
+    # --- 8. Handle Failure of All Candidates ---
+    error_message = "All candidate models failed."
+    if last_exception:
+        error_message += f" Last error ({type(last_exception).__name__}): {last_exception}"
+    rprint(f"[FATAL] {error_message}")
+    raise RuntimeError(error_message) from last_exception
+
+# --- Example Usage (Optional) ---
 if __name__ == "__main__":
-    example_prompt = "Tell me a joke about {topic}"
-    example_input = {"topic": "programming"}
+    # This block allows running the file directly for testing.
+    # Ensure you have a ./data/llm_model.csv file and potentially a .env file.
+
+    # Set PDD_DEBUG_SELECTOR=1 to see model selection details
+    # os.environ["PDD_DEBUG_SELECTOR"] = "1"
+
+    # Example 1: Simple text generation
+    print("\n--- Example 1: Simple Text Generation (Strength 0.5) ---")
     try:
-        output = llm_invoke(example_prompt, example_input, strength=0.5, temperature=0.7, verbose=True)
-        rprint("[bold magenta]Invocation succeeded:[/bold magenta]", output)
-    except Exception as err:
-        rprint(f"[bold red]Invocation failed:[/bold red] {err}")
+        response = llm_invoke(
+            prompt="Tell me a short joke about {topic}.",
+            input_json={"topic": "programmers"},
+            strength=0.5, # Use base model (gpt-4.1-nano)
+            temperature=0.7,
+            verbose=True
+        )
+        rprint("\nExample 1 Response:")
+        rprint(response)
+    except Exception as e:
+        rprint(f"\nExample 1 Failed: {e}")
+
+    # Example 1b: Simple text generation (Strength 0.3)
+    print("\n--- Example 1b: Simple Text Generation (Strength 0.3) ---")
+    try:
+        response = llm_invoke(
+            prompt="Tell me a short joke about {topic}.",
+            input_json={"topic": "keyboards"},
+            strength=0.3, # Should select gemini-pro based on cost interpolation
+            temperature=0.7,
+            verbose=True
+        )
+        rprint("\nExample 1b Response:")
+        rprint(response)
+    except Exception as e:
+        rprint(f"\nExample 1b Failed: {e}")
+
+    # Example 2: Structured output (requires a Pydantic model)
+    print("\n--- Example 2: Structured Output (Pydantic, Strength 0.8) ---")
+    class JokeStructure(BaseModel):
+        setup: str
+        punchline: str
+        rating: Optional[int] = None
+
+    try:
+        # Use a model known to support structured output (check your CSV)
+        # Strength 0.8 should select gemini-pro based on ELO interpolation
+        response_structured = llm_invoke(
+            prompt="Create a joke about {topic}. Output ONLY the JSON object with 'setup' and 'punchline'.",
+            input_json={"topic": "data science"},
+            strength=0.8, # Try a higher ELO model (gemini-pro expected)
+            temperature=1,
+            output_pydantic=JokeStructure,
+            verbose=True
+        )
+        rprint("\nExample 2 Response:")
+        rprint(response_structured)
+        if isinstance(response_structured.get('result'), JokeStructure):
+             rprint("\nPydantic object received successfully:", response_structured['result'].model_dump())
+        else:
+             rprint("\nResult was not the expected Pydantic object:", response_structured.get('result'))
+
+    except Exception as e:
+        rprint(f"\nExample 2 Failed: {e}")
+
+
+    # Example 3: Batch processing
+    print("\n--- Example 3: Batch Processing (Strength 0.3) ---")
+    try:
+        batch_input = [
+            {"animal": "cat", "adjective": "lazy"},
+            {"animal": "dog", "adjective": "energetic"},
+        ]
+        # Strength 0.3 should select gemini-pro
+        response_batch = llm_invoke(
+            prompt="Describe a {adjective} {animal} in one sentence.",
+            input_json=batch_input,
+            strength=0.3, # Cheaper model maybe (gemini-pro expected)
+            temperature=0.5,
+            use_batch_mode=True,
+            verbose=True
+        )
+        rprint("\nExample 3 Response:")
+        rprint(response_batch)
+    except Exception as e:
+        rprint(f"\nExample 3 Failed: {e}")
+
+    # Example 4: Using 'messages' input
+    print("\n--- Example 4: Using 'messages' input (Strength 0.5) ---")
+    try:
+        custom_messages = [
+            {"role": "system", "content": "You are a helpful assistant."},
+            {"role": "user", "content": "What is the capital of France?"}
+        ]
+        # Strength 0.5 should select gpt-4.1-nano
+        response_messages = llm_invoke(
+            messages=custom_messages,
+            strength=0.5,
+            temperature=0.1,
+            verbose=True
+        )
+        rprint("\nExample 4 Response:")
+        rprint(response_messages)
+    except Exception as e:
+        rprint(f"\nExample 4 Failed: {e}")
+
+    # Example 5: Requesting thinking time (e.g., for Anthropic)
+    print("\n--- Example 5: Requesting Thinking Time (Strength 1.0, Time 0.5) ---")
+    try:
+        # Ensure your CSV has max_reasoning_tokens for an Anthropic model
+        # Strength 1.0 should select claude-3 (highest ELO)
+        # Time 0.5 with budget type should request thinking
+        response_thinking = llm_invoke(
+            prompt="Explain the theory of relativity simply, taking some time to think.",
+            input_json={},
+            strength=1.0, # Try to get highest ELO model (claude-3)
+            temperature=1,
+            time=0.5, # Request moderate thinking time
+            verbose=True
+        )
+        rprint("\nExample 5 Response:")
+        rprint(response_thinking)
+    except Exception as e:
+        rprint(f"\nExample 5 Failed: {e}")
+
+    # Example 6: Pydantic Fallback Parsing (Strength 0.3)
+    print("\n--- Example 6: Pydantic Fallback Parsing (Strength 0.3) ---")
+    # This requires mocking litellm.completion to return a JSON string
+    # even when gemini-pro (which supports structured output) is selected.
+    # This is hard to demonstrate cleanly in the __main__ block without mocks.
+    # The unit test test_llm_invoke_output_pydantic_unsupported_parses covers this.
+    print("(Covered by unit tests)")