@microsoft/m365-copilot-eval 1.2.1-preview.1 → 1.3.0-preview.1

package/src/clients/cli/main.py

@@ -4,6 +4,7 @@ import argparse
 import sys
 import csv
 import functools
+import logging
 import webbrowser
 import urllib.request
 import urllib.error
@@ -21,16 +22,40 @@ from azure.ai.evaluation import (
 from dotenv import load_dotenv
 from auth.auth_handler import AuthHandler
 from custom_evaluators.CitationsEvaluator import CitationsEvaluator, CitationFormat
-#from custom_evaluators.ConcisenessNonLLMEvaluator import ConcisenessNonLLMEvaluator
-#from custom_evaluators.PII.PII import PIIEvaluator
+from custom_evaluators.ExactMatchEvaluator import ExactMatchEvaluator
+from custom_evaluators.PartialMatchEvaluator import PartialMatchEvaluator
 from generate_report import generate_html_report, calculate_aggregate_statistics
 from response_extractor import extract_enhanced_responses, get_response_text_for_evaluation
 from schema_handler import DocumentUpgrader, SchemaVersionManager
+from common import (
+    RELEVANCE,
+    COHERENCE,
+    GROUNDEDNESS,
+    TOOL_CALL_ACCURACY,
+    CITATIONS,
+    EXACT_MATCH,
+    PARTIAL_MATCH,
+    REQUIRES_AZURE_OPENAI,
+    REQUIRES_TOOL_DEFINITIONS,
+    METRIC_IDS,
+    pascal_case_to_title,
+)
+from evaluator_resolver import (
+    EVALUATOR_REGISTRY,
+    validate_evaluator_names,
+    check_prerequisites,
+    resolve_default_evaluators,
+    resolve_evaluators_for_prompt,
+    get_evaluator_threshold,
+)
 from version_check import check_min_version, get_cli_version
 from datetime import datetime, timezone
 from pathlib import Path
 import tzlocal
 
+from cli_logging.console_diagnostics import render_diagnostic, serialize_diagnostic_record
+from cli_logging.logging_utils import LOG_LEVEL_MAP, LogLevel, Operation, format_structured_log_entry, resolve_log_level
+
 # Allowed endpoints for URL validation
 ALLOWED_ENDPOINTS = [
     'substrate.office.com'
@@ -48,20 +73,63 @@ VERSION_CHECK_BYPASS_FLAGS = (
     "signout",
 )
 
+CLI_LOGGER_NAME = "m365.eval.cli"
+CLI_LOGGER = logging.getLogger(CLI_LOGGER_NAME)
+DIAGNOSTIC_RECORDS: List[Dict[str, Any]] = []
+
+
+def _ensure_logger_handler() -> None:
+    if CLI_LOGGER.handlers:
+        return
+    handler = logging.StreamHandler(sys.stdout)
+    handler.setFormatter(logging.Formatter("%(message)s"))
+    CLI_LOGGER.addHandler(handler)
+    CLI_LOGGER.propagate = False
+
+
+def configure_cli_logging(effective_log_level: str) -> None:
+    _ensure_logger_handler()
+    CLI_LOGGER.setLevel(LOG_LEVEL_MAP[effective_log_level])
+
+
+def emit_structured_log(level: str, message: str, operation: str = Operation.EVALUATE) -> None:
+    _ensure_logger_handler()
+    context = {
+        "request-id": None,
+        "conversation-id": None,
+        "message-id": None,
+        "operation": operation,
+    }
+    entry = format_structured_log_entry(
+        level=level,
+        message=message,
+        logger_name=CLI_LOGGER_NAME,
+        run_context=context,
+    )
+    DIAGNOSTIC_RECORDS.append(entry)
+
+    try:
+        CLI_LOGGER.log(LOG_LEVEL_MAP.get(level, logging.INFO), render_diagnostic(entry))
+    except Exception:
+        pass
+
 
 def should_bypass_min_version_check(args: argparse.Namespace) -> bool:
     """Return True if the current invocation should skip min-version checks."""
     return any(getattr(args, flag, False) for flag in VERSION_CHECK_BYPASS_FLAGS)
 
-def write_results_to_html(results: List[Dict], output_file: str):
+def write_results_to_html(results: List[Dict], output_file: str,
+                          agent_name: Optional[str] = None, agent_id: Optional[str] = None,
+                          cli_version: Optional[str] = None):
     """Write results to HTML file using generate_html_report from generate_report.py."""
     try:
-        html = generate_html_report(results) 
+        html = generate_html_report(results, agent_name=agent_name, agent_id=agent_id,
+                                    cli_version=cli_version)
         with open(output_file, 'w', encoding='utf-8') as f:
             f.write(html)
-        print(f"HTML report saved to {output_file}")
+        emit_structured_log("info", f"HTML report saved to {output_file}", operation=Operation.WRITE_OUTPUT)
     except Exception as e:
-        print(f"Error writing to HTML file: {e}")
+        emit_structured_log("error", f"Error writing to HTML file: {e}", operation=Operation.WRITE_OUTPUT)
         sys.exit(1)
 
 def get_default_prompts_and_responses():
@@ -74,7 +142,7 @@ def get_default_prompts_and_responses():
     ]
     return prompts, expected_responses
 
-def load_prompts_from_file(file_path: str) -> Tuple[List[str], List[str]]:
+def load_prompts_from_file(file_path: str) -> Tuple[List[Dict], Optional[Dict]]:
     """Load prompts and expected responses from a JSON file.
 
     Supports three formats:
@@ -84,6 +152,10 @@ def load_prompts_from_file(file_path: str) -> Tuple[List[str], List[str]]:
 
     For eval documents (format 1) and array format (format 2), schema validation
     and auto-upgrade are applied via DocumentUpgrader.
+
+    Returns:
+        Tuple of (eval_items, default_evaluators). Items are dicts with prompt,
+        expected_response, and optional evaluators/evaluators_mode fields.
     """
     try:
         with open(file_path, 'r', encoding='utf-8') as f:
@@ -100,18 +172,18 @@ def load_prompts_from_file(file_path: str) -> Tuple[List[str], List[str]]:
                 upgrader = DocumentUpgrader()
             except Exception as e:
                 # Schema infrastructure not available (missing files, etc.) — skip
-                print(f"Warning: Unable to initialize document upgrader: {e}")
+                emit_structured_log("warning", f"Unable to initialize document upgrader: {e}", operation=Operation.LOAD_PROMPTS)
                 upgrader = None
 
             if upgrader is not None:
                 result = upgrader.upgrade(Path(file_path))
 
                 if result.error:
-                    print(f"Schema validation error: {result.error}")
+                    emit_structured_log("error", f"Schema validation error: {result.error}", operation=Operation.LOAD_PROMPTS)
                     sys.exit(1)
 
                 if result.upgraded and result.message:
-                    print(result.message)
+                    emit_structured_log("info", result.message, operation=Operation.LOAD_PROMPTS)
 
                 # Use the parsed document from the upgrade result
                 if result.document is not None:
@@ -119,26 +191,26 @@ def load_prompts_from_file(file_path: str) -> Tuple[List[str], List[str]]:
 
         if isinstance(data, list):
             # Format: [{"prompt": "...", "expected_response": "..."}, ...]
-            prompts = [item.get("prompt", "") for item in data]
-            expected_responses = [item.get("expected_response", "") for item in data]
+            return data, None
         elif isinstance(data, dict):
             if "items" in data:
                 # Eval document format: {"schemaVersion": "...", "items": [...]}
-                items = data["items"]
-                prompts = [item.get("prompt", "") for item in items]
-                expected_responses = [item.get("expected_response", "") for item in items]
+                return data["items"], data.get("default_evaluators")
             else:
                 # Format: {"prompts": [...], "expected_responses": [...]}
                 prompts = data.get("prompts", [])
                 expected_responses = data.get("expected_responses", [])
+                eval_items = [
+                    {"prompt": p, "expected_response": e}
+                    for p, e in zip(prompts, expected_responses)
+                ]
+                return eval_items, None
         else:
             raise ValueError("Invalid file format")
-
-        return prompts, expected_responses
     except SystemExit:
         raise
     except Exception as e:
-        print(f"Error loading prompts from file: {e}")
+        emit_structured_log("error", f"Error loading prompts from file: {e}", operation=Operation.LOAD_PROMPTS)
         sys.exit(1)
 
 def get_interactive_prompts() -> Tuple[List[str], List[str]]:
@@ -165,116 +237,168 @@ def get_interactive_prompts() -> Tuple[List[str], List[str]]:
     
     return prompts, expected_responses
 
-def run_evaluations(args, responses: dict, expected_responses: list) -> list:
-    """Run evaluations against the responses."""
+def run_evaluations(args, responses: List[Dict[str, Any]], eval_items: List[Dict],
+                     default_evaluators: Dict[str, Any]) -> list:
+    """Run evaluations against the responses using per-prompt evaluator resolution.
+
+    Args:
+        args: CLI arguments.
+        responses: List of enhanced response dicts (one per prompt, aligned with eval_items by index).
+        eval_items: List of item dicts (prompt, expected_response, evaluators, evaluators_mode).
+        default_evaluators: Resolved default evaluators (from resolve_default_evaluators).
+    """
+    if len(responses) != len(eval_items):
+        raise ValueError(
+            f"Mismatch between number of responses ({len(responses)}) and evaluation items ({len(eval_items)})."
+        )
+
     model_config = AzureOpenAIModelConfiguration(
         azure_endpoint=os.environ.get("AZURE_AI_OPENAI_ENDPOINT"),
         api_key=os.environ.get("AZURE_AI_API_KEY"),
         api_version=os.environ.get("AZURE_AI_API_VERSION"),
         azure_deployment=os.environ.get("AZURE_AI_MODEL_NAME"),
     )
-    
-    # Initialize evaluators
-    relevance_evaluator = RelevanceEvaluator(model_config=model_config) # Evaluate relevance for a given response. Range is 1 - 5.
-    coherence_evaluator = CoherenceEvaluator(model_config=model_config) # Measures the coherence (human-like quality) of the response. Range is 1 - 5.
-    groundedness_evaluator = GroundednessEvaluator(model_config=model_config) # Evaluates the response for for factuality and groundedness against provided context. Range is 1 - 5.
-    #concisenessnonllm_evaluator = ConcisenessNonLLMEvaluator() # Evaluates the response for conciseness. Range is 1 - 5.
-    #pii_evaluator = PIIEvaluator(model_config=model_config) # Evaluates the response for presence of PII. Range
-    # Parse citation format from args
-    citation_format = CitationFormat.OAI_UNICODE if args.citation_format == 'oai_unicode' else CitationFormat.LEGACY_BRACKET
-    citations_evaluator = CitationsEvaluator(citation_format=citation_format) # Evaluates citations present in the response using regex pattern matching
-
-    tool_call_accuracy_evaluator = ToolCallAccuracyEvaluator(model_config) # Evaluate tool call accuracy if tool definitions are present in response
-    
 
-    PASS_THRESHOLD = 3  # All evaluators must meet or exceed this value (out of 5) to pass
+    # Build available context for prerequisite checks
+    has_azure_openai = bool(
+        os.environ.get("AZURE_AI_OPENAI_ENDPOINT")
+        and os.environ.get("AZURE_AI_API_KEY")
+    )
 
-    def decorate_metric(metric_id: str, data):
+    DEFAULT_PASS_THRESHOLD = 3
+
+    def decorate_metric(metric_id: str, data, threshold: Optional[int] = None):
         """Augment raw evaluator output with standardized threshold + pass/fail result."""
+        pass_threshold = threshold if threshold is not None else DEFAULT_PASS_THRESHOLD
         payload = {}
-        # Preserve original structure if dict
         if isinstance(data, dict):
             payload.update(data)
         else:
             payload['raw'] = data
 
-        # Try to extract a numeric score
         score_val = None
         if isinstance(data, dict):
-            for k in (metric_id, f"{metric_id}_score", 'score', 'value'):
-                if k in data:
-                    score_val = data[k]
-                    break
+            if metric_id in data:
+                score_val = data[metric_id]
         if isinstance(score_val, (int, float)):
-            payload['threshold'] = PASS_THRESHOLD
-            payload['result'] = 'pass' if score_val >= PASS_THRESHOLD else 'fail'
+            payload['threshold'] = pass_threshold
+            payload['result'] = 'pass' if score_val >= pass_threshold else 'fail'
         else:
-            # If we cannot determine score, mark unknown (no pass/fail)
-            payload['threshold'] = PASS_THRESHOLD
+            payload['threshold'] = pass_threshold
             payload.setdefault('result', 'unknown')
         return json.dumps(payload, indent=4)
 
+    # Validate all evaluator names upfront (across defaults and all items)
+    all_evaluator_maps = [default_evaluators]
+    for eval_item in eval_items:
+        if "evaluators" in eval_item:
+            all_evaluator_maps.append(eval_item["evaluators"])
+    for emap in all_evaluator_maps:
+        validate_evaluator_names(emap)
+
     evaluation_results = []
-    for prompt, expected_response in zip(responses.keys(), expected_responses):
-        # Extract text response for evaluation (backward compatibility)
-        enhanced_response = responses[prompt]
+    for enhanced_response, eval_item in zip(responses, eval_items):
         actual_response_text = get_response_text_for_evaluation(enhanced_response)
-        
-        # Run evaluations using text response
-        relevance_score = relevance_evaluator(
-            query=prompt,
-            response=actual_response_text
+        prompt = eval_item.get("prompt", "")
+        expected_response = eval_item.get("expected_response", "")
+        prompt_evaluators = eval_item.get("evaluators")
+        evaluators_mode = eval_item.get("evaluators_mode", "extend")
+
+        # Resolve evaluators for this prompt
+        resolved = resolve_evaluators_for_prompt(
+            prompt_evaluators, evaluators_mode, prompt, default_evaluators,
         )
-        coherence_score = coherence_evaluator(
-            query=prompt,
-            response=actual_response_text
-        )
-
-        groundedness_score = groundedness_evaluator(
-            response=actual_response_text,
-            context=expected_response
-        )
-
-        #PII_score = pii_evaluator(response=actual_response_text)
-        #concisenessNonLLM_score = concisenessnonllm_evaluator(response=actual_response_text)
 
-        citations_score = citations_evaluator(
-            response=actual_response_text
+        # Build runtime context for prerequisite checks
+        has_tool_defs = bool(
+            args.m365_agent_id and enhanced_response.get("tool_definitions")
         )
+        available_context = {
+            REQUIRES_AZURE_OPENAI: has_azure_openai,
+            REQUIRES_TOOL_DEFINITIONS: has_tool_defs,
+        }
 
-        tool_call_accuracy = None
-        if args.m365_agent_id and enhanced_response.get("tool_definitions"):
-            tool_call_accuracy = tool_call_accuracy_evaluator(
-                query=prompt,
-                response=enhanced_response.get("response", actual_response_text),
-                tool_definitions=enhanced_response["tool_definitions"]
-            )
+        results_dict: Dict[str, Optional[str]] = {}
+        evaluators_ran: List[str] = []
+
+        for eval_name, eval_options in resolved.items():
+            # Check prerequisites
+            can_run, warn_msg = check_prerequisites(eval_name, available_context)
+            if not can_run:
+                if warn_msg:
+                    emit_structured_log("warning", f"Evaluator '{eval_name}' prerequisite check failed: {warn_msg}", operation=Operation.EVALUATE)
+                results_dict[eval_name] = None
+                continue
+
+            evaluators_ran.append(eval_name)
+            threshold = get_evaluator_threshold(eval_name, eval_options)
+
+            if eval_name == RELEVANCE:
+                raw_score = RelevanceEvaluator(model_config=model_config)(query=prompt, response=actual_response_text)
+                results_dict[RELEVANCE] = decorate_metric(METRIC_IDS[RELEVANCE], raw_score, threshold)
+            elif eval_name == COHERENCE:
+                raw_score = CoherenceEvaluator(model_config=model_config)(query=prompt, response=actual_response_text)
+                results_dict[COHERENCE] = decorate_metric(METRIC_IDS[COHERENCE], raw_score, threshold)
+            elif eval_name == GROUNDEDNESS:
+                raw_score = GroundednessEvaluator(model_config=model_config)(response=actual_response_text, context=expected_response)
+                results_dict[GROUNDEDNESS] = decorate_metric(METRIC_IDS[GROUNDEDNESS], raw_score, threshold)
+            elif eval_name == TOOL_CALL_ACCURACY:
+                raw_score = ToolCallAccuracyEvaluator(model_config)(
+                    query=prompt,
+                    response=enhanced_response.get("response", actual_response_text),
+                    tool_definitions=enhanced_response["tool_definitions"],
+                )
+                results_dict[TOOL_CALL_ACCURACY] = decorate_metric(METRIC_IDS[TOOL_CALL_ACCURACY], raw_score, threshold)
+            elif eval_name == CITATIONS:
+                fmt_str = eval_options.get("citation_format", "oai_unicode")
+                fmt_map = {
+                    "oai_unicode": CitationFormat.OAI_UNICODE,
+                    "bracket": CitationFormat.LEGACY_BRACKET,
+                    "mixed": CitationFormat.AUTO,
+                }
+                raw_score = CitationsEvaluator(citation_format=fmt_map.get(fmt_str, CitationFormat.OAI_UNICODE))(response=actual_response_text)
+                results_dict[CITATIONS] = decorate_metric(METRIC_IDS[CITATIONS], raw_score, threshold)
+            elif eval_name == EXACT_MATCH:
+                # ExactMatch is binary (match/no-match) — it includes its own result
+                # field, so we skip decorate_metric which assumes a numeric score.
+                case_sensitive = eval_options.get("case_sensitive", False)
+                raw_score = ExactMatchEvaluator(case_sensitive=case_sensitive)(response=actual_response_text, expected_answer=expected_response)
+                results_dict[EXACT_MATCH] = json.dumps(raw_score, indent=4)
+            elif eval_name == PARTIAL_MATCH:
+                case_sensitive = eval_options.get("case_sensitive", False)
+                raw_score = PartialMatchEvaluator(case_sensitive=case_sensitive)(response=actual_response_text, expected_answer=expected_response)
+                results_dict[PARTIAL_MATCH] = decorate_metric(METRIC_IDS[PARTIAL_MATCH], raw_score, threshold)
 
         evaluation_result = {
             "prompt": prompt,
-            "response": actual_response_text,  # Keep simple text for backward compatibility
+            "response": actual_response_text,
             "expected_response": expected_response,
-            "results": {
-                "relevance_score": decorate_metric("relevance", relevance_score),
-                "coherence_score": decorate_metric("coherence", coherence_score),
-                "groundedness_score": decorate_metric("groundedness", groundedness_score),
-                #"concisenessnonllm_score": decorate_metric("concisenessnonllm", concisenessNonLLM_score),
-                #"pii_score": decorate_metric("pii", PII_score),
-                "citations_score": json.dumps(citations_score, indent=4),
-                "tool_call_accuracy_score": json.dumps(tool_call_accuracy, indent=4) if tool_call_accuracy else None
-            }
+            "evaluators_ran": evaluators_ran,
+            "results": results_dict,
         }
 
-        if args.verbose:
-            print(f".................................. Evaluation for prompt: {evaluation_result['prompt']} ..................................")
-            print(f"Scores: {evaluation_result['results']}")
-            print("...........................................................................................................................")
+        # Preserve evaluator config metadata for output
+        if "evaluators" in eval_item:
+            evaluation_result["evaluators"] = eval_item["evaluators"]
+        if "evaluators_mode" in eval_item:
+            evaluation_result["evaluators_mode"] = eval_item["evaluators_mode"]
+
+        if getattr(args, "effective_log_level", "info") == "debug":
+            emit_structured_log(
+                "debug",
+                f"Evaluation completed for prompt='{evaluation_result['prompt']}'. "
+                f"Evaluators: {', '.join(evaluators_ran)}. "
+                f"Scores: {evaluation_result['results']}",
+                operation=Operation.EVALUATE,
+            )
 
         evaluation_results.append(evaluation_result)
-    
+
     return evaluation_results
 
-def write_results_to_console(results):
+def write_results_to_console(results, agent_name: Optional[str] = None,
+                             agent_id: Optional[str] = None,
+                             cli_version: Optional[str] = None):
     """Write the response to console."""
     # ANSI color codes
     BOLD = '\033[1m'
@@ -286,47 +410,66 @@ def write_results_to_console(results):
     ORANGE = '\033[38;5;208m'
     RED = '\033[91m'
     RESET = '\033[0m'
-    
+
+    # Show metadata
+    metadata_parts = []
+    if agent_name:
+        metadata_parts.append(f"Agent Name: {agent_name}")
+    if agent_id:
+        metadata_parts.append(f"Agent ID: {agent_id}")
+    if cli_version:
+        metadata_parts.append(f"CLI Version: {cli_version}")
+    if metadata_parts:
+        print(f"{BOLD}{CYAN}{' | '.join(metadata_parts)}{RESET}")
+        print()
+
     # Show aggregate statistics if multiple results
     if len(results) > 1:
         aggregates = calculate_aggregate_statistics(results)
         if aggregates:
-            print(f"{BOLD}{BLUE}📊 Aggregate Statistics ({len(results)} prompts):{RESET}")
+            print(f"{BOLD}{BLUE}Aggregate Statistics ({len(results)} prompts):{RESET}")
             print(f"{BLUE}{'=' * 60}{RESET}")
-            
+
             for metric_name, stats in aggregates.items():
                 pass_color = GREEN if stats['pass_rate'] >= 80 else YELLOW if stats['pass_rate'] >= 60 else RED
-                print(f"{BOLD}{CYAN}{metric_name}:{RESET}")
+                prompts_evaluated = stats.get('prompts_evaluated', stats['total_evaluated'])
+                total_prompts = stats.get('total_prompts', len(results))
+                print(f"{BOLD}{CYAN}{metric_name}:{RESET} ({prompts_evaluated}/{total_prompts} prompts)")
                 print(f"  Pass Rate: {pass_color}{stats['pass_rate']:.1f}%{RESET} ({stats['pass_count']}/{stats['total_evaluated']} passed)")
                 print(f"  Avg Score: {MAGENTA}{stats['avg_score']:.2f}{RESET}")
                 if stats.get('threshold') is not None:
                     print(f"  Threshold: {YELLOW}{stats['threshold']}{RESET}")
                 print()
-            
+
             print(f"{BLUE}{'=' * 60}{RESET}")
             print()
-    
-    print(f"{BOLD}{BLUE}📝 Individual Results:{RESET}")
+
+    print(f"{BOLD}{BLUE}Individual Results:{RESET}")
     print(f"{BLUE}{'=' * 50}{RESET}")
     for i, result in enumerate(results, 1):
         print(f"{BOLD}{GREEN}Prompt {i}:{RESET} {result['prompt']}")
+
+        # Show which evaluators ran for this prompt
+        evaluators_ran = result.get('evaluators_ran', [])
+        if evaluators_ran:
+            print(f"{BOLD}{CYAN}Evaluators:{RESET} {', '.join(evaluators_ran)}")
+
         print(f"{BOLD}{CYAN}Response:{RESET} {result['response']}")
         print(f"{BOLD}{YELLOW}Expected Response:{RESET} {result['expected_response']}")
 
-        # Print metric scores generically from nested results (fallback to flat keys for back-compat)
-        metrics = result.get('results') or {k: v for k, v in result.items() if isinstance(k, str) and k.endswith('_score')}
-        if metrics:
-            for k, v in metrics.items():
-                name = k.replace('_', ' ')
-                if 'relevance' in k:
-                    color = MAGENTA
-                elif 'coherence' in k:
-                    color = ORANGE
-                elif 'fluency' in k:
-                    color = GREEN
-                else:
-                    color = BLUE
-                print(f"{BOLD}{color}{name}:{RESET} {v}")
+        # Print metric scores from results
+        metrics = result.get('results', {})
+        for eval_name, v in metrics.items():
+            if v is None:
+                continue  # Skip null/N/A scores from skipped evaluators
+            display_name = pascal_case_to_title(eval_name)
+            if eval_name == RELEVANCE:
+                color = MAGENTA
+            elif eval_name == COHERENCE:
+                color = ORANGE
+            else:
+                color = BLUE
+            print(f"{BOLD}{color}{display_name}:{RESET} {v}")
         print(f"{BLUE}{'-' * 30}{RESET}")
 
 def extract_eval_score(data: dict, metric_id: str) -> Optional[Dict]:
@@ -338,10 +481,8 @@ def extract_eval_score(data: dict, metric_id: str) -> Optional[Dict]:
     DEFAULT_THRESHOLD = 3  # fallback; decorate_metric should always set this
 
     score_val = None
-    for k in (metric_id, f"{metric_id}_score", "score", "value"):
-        if k in data and isinstance(data[k], (int, float)):
-            score_val = data[k]
-            break
+    if metric_id in data and isinstance(data[metric_id], (int, float)):
+        score_val = data[metric_id]
     if score_val is None:
         return None
 
@@ -364,9 +505,11 @@ def convert_result_to_eval_item(result: Dict) -> Dict:
     """Convert an internal evaluation result dict to a schema-compliant EvalItem.
 
     Internal format (from run_evaluations):
-        {prompt, response, expected_response, results: {relevance_score: "JSON", ...}}
+        {prompt, response, expected_response, results: {Relevance: "JSON", ...},
+         evaluators_ran: [...], evaluators: {...}, evaluators_mode: "..."}
     Schema EvalItem format:
-        {prompt, response, expected_response, scores: {relevance: EvalScore, ...}}
+        {prompt, response, expected_response, scores: {relevance: EvalScore, ...},
+         evaluators: {...}, evaluators_mode: "..."}
     """
     item: Dict[str, Any] = {
         "prompt": result["prompt"],
@@ -374,30 +517,35 @@ def convert_result_to_eval_item(result: Dict) -> Dict:
         "expected_response": result["expected_response"],
     }
 
+    # Preserve evaluator config in output
+    if "evaluators" in result:
+        item["evaluators"] = result["evaluators"]
+    if "evaluators_mode" in result:
+        item["evaluators_mode"] = result["evaluators_mode"]
+
     scores: Dict[str, Any] = {}
     results_dict = result.get("results", {})
 
     # EvalScore metrics (all share the same schema shape: {score, result, threshold})
-    # Tuple: (internal results key, metric ID for score lookup, schema output key)
-    for internal_key, metric_id, schema_key in [
-        ("relevance_score", "relevance", "relevance"),
-        ("coherence_score", "coherence", "coherence"),
-        ("groundedness_score", "groundedness", "groundedness"),
-        ("tool_call_accuracy_score", "tool_call_accuracy", "toolCallAccuracy"),
+    for eval_key, schema_key in [
+        (RELEVANCE, "relevance"),
+        (COHERENCE, "coherence"),
+        (GROUNDEDNESS, "groundedness"),
+        (TOOL_CALL_ACCURACY, "toolCallAccuracy"),
     ]:
-        raw = results_dict.get(internal_key)
+        raw = results_dict.get(eval_key)
         if not raw:
             continue
         data = json.loads(raw) if isinstance(raw, str) else raw
-        eval_score = extract_eval_score(data, metric_id)
+        eval_score = extract_eval_score(data, METRIC_IDS[eval_key])
         if eval_score:
             scores[schema_key] = eval_score
 
-    # Citations → CitationScore (different schema shape: {count, result, threshold} + format)
-    raw_citations = results_dict.get("citations_score")
+    # Citations → CitationScore
+    raw_citations = results_dict.get(CITATIONS)
     if raw_citations:
         data = json.loads(raw_citations) if isinstance(raw_citations, str) else raw_citations
-        count = data.get("score", 0)
+        count = data.get("citations", 0)
         cit_result = data.get("result")
         if cit_result not in ("pass", "fail"):
             cit_result = "pass" if count >= data.get("threshold", 1) else "fail"
@@ -411,17 +559,42 @@ def convert_result_to_eval_item(result: Dict) -> Dict:
             citation_score["format"] = data["citation_format"]
         scores["citations"] = citation_score
 
+    # ExactMatch → ExactMatchScore
+    raw_exact = results_dict.get(EXACT_MATCH)
+    if raw_exact:
+        data = json.loads(raw_exact) if isinstance(raw_exact, str) else raw_exact
+        is_match = data.get("exact_match", 0.0) == 1.0
+        scores["exactMatch"] = {
+            "match": is_match,
+            "result": data.get("result", "pass" if is_match else "fail"),
+            "reason": data.get("exact_match_reason", ""),
+        }
+
+    # PartialMatch → PartialMatchScore
+    raw_partial = results_dict.get(PARTIAL_MATCH)
+    if raw_partial:
+        data = json.loads(raw_partial) if isinstance(raw_partial, str) else raw_partial
+        scores["partialMatch"] = {
+            "score": data.get("partial_match", 0.0),
+            "result": data.get("result", "fail"),
+            "threshold": data.get("threshold", 0.5),
+            "reason": data.get("partial_match_reason", ""),
+        }
+
     if scores:
         item["scores"] = scores
 
     return item
 
 
-def write_results_to_json(results: List[Dict], output_file: str, agent_id: Optional[str] = None):
+def write_results_to_json(results: List[Dict], output_file: str, agent_id: Optional[str] = None,
+                          default_evaluators: Optional[Dict[str, Any]] = None,
+                          agent_name: Optional[str] = None,
+                          cli_version: Optional[str] = None):
     """Write results to a schema-compliant eval document JSON file.
 
     Output follows the eval-document.schema.json format:
-    {schemaVersion, metadata, items: [EvalItem]}
+    {schemaVersion, metadata, default_evaluators?, items: [EvalItem]}
     """
     try:
         try:
@@ -436,43 +609,68 @@ def write_results_to_json(results: List[Dict], output_file: str, agent_id: Optio
         }
         if agent_id:
             metadata["agentId"] = agent_id
+        if agent_name:
+            metadata["agentName"] = agent_name
+        if cli_version:
+            metadata["cliVersion"] = cli_version
 
         output_data: Dict[str, Any] = {
             "schemaVersion": current_version,
             "metadata": metadata,
-            "items": items,
         }
 
+        if default_evaluators is not None:
+            output_data["default_evaluators"] = default_evaluators
+
+        output_data["items"] = items
+
         with open(output_file, 'w', encoding='utf-8') as f:
             json.dump(output_data, f, indent=2, ensure_ascii=False)
-        print(f"Results saved to {output_file}")
+        emit_structured_log("info", f"Results saved to {output_file}", operation=Operation.WRITE_OUTPUT)
     except Exception as e:
-        print(f"Error writing to JSON file: {e}")
+        emit_structured_log("error", f"Error writing to JSON file: {e}", operation=Operation.WRITE_OUTPUT)
         sys.exit(1)
 
-def write_results_to_csv(results: List[Dict], output_file: str):
+def write_results_to_csv(results: List[Dict], output_file: str,
+                         agent_name: Optional[str] = None, agent_id: Optional[str] = None,
+                         cli_version: Optional[str] = None):
     """Write results to CSV file."""
     try:
         with open(output_file, 'w', newline='', encoding='utf-8') as f:
             if results:
+                # Write metadata header
+                metadata_parts = []
+                if agent_name:
+                    metadata_parts.append(f"Agent Name: {agent_name}")
+                if agent_id:
+                    metadata_parts.append(f"Agent ID: {agent_id}")
+                if cli_version:
+                    metadata_parts.append(f"CLI Version: {cli_version}")
+                if metadata_parts:
+                    f.write(f"# {' | '.join(metadata_parts)}\n")
+
                 # Write aggregate statistics first if multiple results
                 if len(results) > 1:
                     aggregates = calculate_aggregate_statistics(results)
                     if aggregates:
                         f.write("# AGGREGATE STATISTICS\n")
-                        f.write("Metric,Pass Rate (%),Passed,Failed,Avg Score,Threshold\n")
+                        f.write("Metric,Prompts Evaluated,Total Prompts,Pass Rate (%),Passed,Failed,Avg Score,Threshold\n")
                         for metric_name, stats in aggregates.items():
                             threshold_str = str(stats.get('threshold', 'N/A'))
-                            f.write(f"{metric_name},{stats['pass_rate']:.1f},{stats['pass_count']},{stats['fail_count']},{stats['avg_score']:.2f},{threshold_str}\n")
+                            prompts_evaluated = stats.get('prompts_evaluated', stats['total_evaluated'])
+                            total_prompts = stats.get('total_prompts', len(results))
+                            f.write(f"{metric_name},{prompts_evaluated},{total_prompts},{stats['pass_rate']:.1f},{stats['pass_count']},{stats['fail_count']},{stats['avg_score']:.2f},{threshold_str}\n")
                         f.write("\n# INDIVIDUAL RESULTS\n")
-                
-                # Write individual results
-                writer = csv.DictWriter(f, fieldnames=results[0].keys())
+
+                # Write individual results (exclude internal fields)
+                exclude_keys = {'evaluators_ran', 'evaluators', 'evaluators_mode'}
+                fieldnames = [k for k in results[0].keys() if k not in exclude_keys]
+                writer = csv.DictWriter(f, fieldnames=fieldnames, extrasaction='ignore')
                 writer.writeheader()
                 writer.writerows(results)
-        print(f"Results saved to {output_file}")
+            emit_structured_log("info", f"Results saved to {output_file}", operation=Operation.WRITE_OUTPUT)
     except Exception as e:
-        print(f"Error writing to CSV file: {e}")
+        emit_structured_log("error", f"Error writing to CSV file: {e}", operation=Operation.WRITE_OUTPUT)
         sys.exit(1)
 
 def parse_arguments():
@@ -503,8 +701,8 @@ Examples:
   # Save results to HTML and open in browser
   python main.py --output report.html
 
-  # Verbose output
-  python main.py --verbose
+  # Debug-level diagnostics
+  python main.py --log-level debug
 
   # Sign out and clear cached authentication tokens
   python main.py --signout
@@ -553,21 +751,13 @@ Examples:
     
     # Behavior options
     parser.add_argument(
-        '--verbose', 
-        action='store_true', 
-        help='Enable verbose output'
-    )
-    parser.add_argument(
-        '--quiet', 
-        action='store_true', 
-        help='Suppress non-essential output'
-    )
-    parser.add_argument(
-        '--citation-format',
-        choices=['oai_unicode', 'legacy_bracket'],
-        default='oai_unicode',
-        help='Citation format to detect. "oai_unicode" for new OAI format (default), "legacy_bracket" for old [^i^] format'
+        '--log-level',
+        nargs='?',
+        const='info',
+        action='append',
+        help='Set log verbosity: debug, info, warning, error. Bare --log-level resolves to info.'
     )
+
     parser.add_argument(
         '--signout',
         action='store_true',
@@ -600,8 +790,13 @@ def validate_environment() -> CallPath:
     
     missing_vars = [var for var in required_env_vars if not os.environ.get(var)]
     if missing_vars:
-        print(f"Error: Missing required environment variables: {', '.join(missing_vars)}")
-        print("Please ensure your .env file contains all required Azure configuration.")
+        emit_structured_log(
+            "error",
+            "Missing required environment variables: "
+            f"{', '.join(missing_vars)}. Please ensure your .env file contains "
+            "all required Azure configuration.",
+            operation=Operation.VALIDATE_ENV,
+        )
         sys.exit(1)
     return call_path
 
@@ -635,23 +830,42 @@ def validate_endpoint_url(url: str, allowed_domains: List[str]) -> bool:
         # Convert other parsing errors to ValueError
         raise ValueError(f"Invalid URL format: {url}") from e
 
-def get_prompt_datasets(args) -> Tuple[List[str], List[str]]:
-    """Get prompts and expected responses based on command line arguments."""
+def get_prompt_datasets(args) -> Tuple[List[Dict], Optional[Dict]]:
+    """Get prompts and expected responses based on command line arguments.
+
+    Returns:
+        Tuple of (eval_items, default_evaluators).
+    """
     if args.prompts:
         if args.expected and len(args.prompts) != len(args.expected):
-            print("Error: Number of prompts must match number of expected responses")
+            emit_structured_log(
+                "error",
+                "Number of prompts must match number of expected responses. "
+                "Update --expected values to match the prompt count.",
+            )
             sys.exit(1)
-        prompts = args.prompts
-        expected_responses = args.expected or [""] * len(prompts)
+        expected_responses = args.expected or [""] * len(args.prompts)
+        eval_items = [
+            {"prompt": p, "expected_response": e}
+            for p, e in zip(args.prompts, expected_responses)
+        ]
+        return eval_items, None
     elif args.prompts_file:
-        prompts, expected_responses = load_prompts_from_file(args.prompts_file)
+        return load_prompts_from_file(args.prompts_file)
     elif args.interactive:
         prompts, expected_responses = get_interactive_prompts()
+        eval_items = [
+            {"prompt": p, "expected_response": e}
+            for p, e in zip(prompts, expected_responses)
+        ]
+        return eval_items, None
     else:
-        # Use default prompts
         prompts, expected_responses = get_default_prompts_and_responses()
-    
-    return prompts, expected_responses
+        eval_items = [
+            {"prompt": p, "expected_response": e}
+            for p, e in zip(prompts, expected_responses)
+        ]
+        return eval_items, None
 
 def fetch_available_agents(copilot_api_endpoint: str, access_token: str, user_oid: str) -> List[Dict[str, Any]]:
     """
@@ -687,26 +901,27 @@ def fetch_available_agents(copilot_api_endpoint: str, access_token: str, user_oi
             return agents
     except urllib.error.HTTPError as e:
         # If endpoint doesn't exist or returns error, return empty list
-        print(f"Warning: Unable to fetch agents list (HTTP {e.code}).")
+        emit_structured_log("warning", f"Unable to fetch agents list (HTTP {e.code}).", operation=Operation.FETCH_AGENTS)
         return []
     except Exception as e:
-        print(f"Warning: Error fetching agents: {e}")
+        emit_structured_log("warning", f"Error fetching agents: {e}", operation=Operation.FETCH_AGENTS)
         return []
 
-def select_agent_interactively(agents: List[Dict[str, Any]]) -> Optional[str]:
+def select_agent_interactively(agents: List[Dict[str, Any]]) -> Tuple[Optional[str], Optional[str]]:
     """
     Display an interactive agent selector using questionary.
-    
+
     Args:
         agents: List of agent dictionaries.
-        
+
     Returns:
-        Selected agent ID or None if cancelled/skipped
+        Tuple of (agent_id, agent_name) or (None, None) if cancelled/skipped
     """
     if not agents:
-        return None
-    
-    # Create choices for questionary
+        return None, None
+
+    # Build id→name lookup and choices
+    id_to_name: Dict[str, str] = {}
     choices = []
     sorted_agents = sorted(agents, key=lambda x: (not x.get('isOwner', False), x.get('name', '')))
     for agent in sorted_agents:
@@ -714,12 +929,13 @@ def select_agent_interactively(agents: List[Dict[str, Any]]) -> Optional[str]:
         agent_name = agent.get("name", "Unknown")
         agent_description = agent.get("description", "Unknown")
         agent_is_owner = agent.get('isOwner')
-        
+        id_to_name[agent_id] = agent_name
+
         # Format the display text
         display_text = f"{agent_name} ({agent_id}, IsOwner: {agent_is_owner}) - {agent_description}"
-        
+
         choices.append(questionary.Choice(title=display_text, value=agent_id))
-    
+
     # Display the selection prompt
     selected_agent = questionary.select(
         "Select an agent to evaluate:",
@@ -727,8 +943,8 @@ def select_agent_interactively(agents: List[Dict[str, Any]]) -> Optional[str]:
         use_shortcuts=True,
         use_arrow_keys=True
     ).ask()
-    
-    return selected_agent
+
+    return selected_agent, id_to_name.get(selected_agent) if selected_agent else None
 
 @functools.lru_cache(maxsize=1)
 def _get_iana_timezone_name() -> str:
@@ -787,7 +1003,7 @@ def build_chat_payload(prompt: str, user_oid: str, agent_id: str|None) -> bytes:
   
     return json.dumps(message).encode("utf-8")
 
-def send_prompt_to_agent_in_sydney(prompts: List[str], copilot_api_endpoint: str, access_token: str, user_oid: str, args) -> Dict[str, Dict[str, any]]:
+def send_prompt_to_agent_in_sydney(prompts: List[str], copilot_api_endpoint: str, access_token: str, user_oid: str, args) -> List[Dict[str, Any]]:
     """ Send prompts to the chat API and return enhanced responses. """
 
     request_headers = {
@@ -795,15 +1011,15 @@ def send_prompt_to_agent_in_sydney(prompts: List[str], copilot_api_endpoint: str
         "X-Scenario": os.environ.get("X_SCENARIO_HEADER"),
         "Authorization": f"Bearer {access_token}"
         }
-    raw_responses: Dict[str, str] = {}
+    raw_responses: List[Tuple[str, str]] = []
     for i, prompt in enumerate(prompts, 1):
-        if not args.quiet:
-            print(f"Processing prompt {i}/{len(prompts)}...")
+        if getattr(args, "effective_log_level", "info") in ("info", "debug"):
+            emit_structured_log("info", f"Processing prompt {i}/{len(prompts)}.", operation=Operation.SEND_PROMPT)
 
         # Build the payload
         payload = build_chat_payload(prompt, user_oid, args.m365_agent_id)
-        if args.verbose:
-            print(f"[Sydney] Sending payload: {payload.decode('utf-8')}")
+        if getattr(args, "effective_log_level", "info") == "debug":
+            emit_structured_log("debug", f"[Sydney] Sending payload: {payload.decode('utf-8')}", operation=Operation.SEND_PROMPT)
 
         # Send the request to /chat
         req = urllib.request.Request(f"{copilot_api_endpoint}/chat", data=payload, headers=request_headers, method="POST")
@@ -822,42 +1038,83 @@ def send_prompt_to_agent_in_sydney(prompts: List[str], copilot_api_endpoint: str
             raise RuntimeError(msg) from e
         except urllib.error.URLError as e:
             raise RuntimeError(f"Chat API connection error: {getattr(e, 'reason', str(e))}") from e
-        
-        if args.verbose:
-            print(f"[Sydney] Raw response: {raw}")
-        
+
+        if getattr(args, "effective_log_level", "info") == "debug":
+            emit_structured_log("debug", f"[Sydney] Raw response: {raw}", operation=Operation.SEND_PROMPT)
+
         # Store raw response for enhancement
-        raw_responses[prompt] = raw.strip()
-    
+        raw_responses.append((prompt, raw.strip()))
+
     # Extract enhanced responses using the new extractor
-    enhanced_responses = extract_enhanced_responses(raw_responses)
+    enhanced_responses = extract_enhanced_responses(raw_responses, log_level=getattr(args, "effective_log_level", "info"))
+
+    if getattr(args, "effective_log_level", "info") == "debug":
+        for idx, enhanced in enumerate(enhanced_responses, 1):
+            metadata = enhanced.get("metadata", {})
+            context = {
+                "request-id": metadata.get("request_id"),
+                "conversation-id": metadata.get("conversation_id"),
+                "message-id": metadata.get("message_id"),
+                "operation": Operation.SEND_PROMPT,
+            }
+            entry = format_structured_log_entry(
+                level="debug",
+                message=f"Response IDs for prompt {idx}/{len(enhanced_responses)}.",
+                logger_name=CLI_LOGGER_NAME,
+                run_context=context,
+            )
+            DIAGNOSTIC_RECORDS.append(entry)
+            CLI_LOGGER.log(logging.DEBUG, render_diagnostic(entry))
+
     return enhanced_responses
 
-def output_results(results: List[Dict], args):
+def output_results(results: List[Dict], args, default_evaluators: Optional[Dict[str, Any]] = None,
+                    agent_name: Optional[str] = None, cli_version: Optional[str] = None):
     """Output results based on specified format."""
+    metadata_kwargs = dict(
+        agent_name=agent_name,
+        agent_id=getattr(args, 'm365_agent_id', None),
+        cli_version=cli_version,
+    )
     if args.output:
         output_lower = args.output.lower()
         if output_lower.endswith('.json'):
-            write_results_to_json(results, args.output, agent_id=getattr(args, 'm365_agent_id', None))
+            write_results_to_json(results, args.output, default_evaluators=default_evaluators,
+                                  **metadata_kwargs)
         elif output_lower.endswith('.csv'):
-            write_results_to_csv(results, args.output)
+            write_results_to_csv(results, args.output, **metadata_kwargs)
         elif output_lower.endswith('.html'):
-            write_results_to_html(results, args.output)
+            write_results_to_html(results, args.output, **metadata_kwargs)
             abs_path = os.path.abspath(args.output)
             webbrowser.open(f'file://{abs_path}')
         else:
-            write_results_to_json(results, args.output, agent_id=getattr(args, 'm365_agent_id', None))
+            write_results_to_json(results, args.output, default_evaluators=default_evaluators,
+                                  **metadata_kwargs)
     else:
-        write_results_to_console(results)
+        write_results_to_console(results, **metadata_kwargs)
 
 def main():
     """Main function to orchestrate the evaluation process."""
     load_dotenv()
     args = parse_arguments()
 
+    effective_log_level, error_message = resolve_log_level(args.log_level)
+    if error_message:
+        print(error_message)
+        print(
+            "Next step: rerun with --log-level {debug|info|warning|error}. "
+            "For support, share the console diagnostics output from this run."
+        )
+        sys.exit(2)
+
+    args.effective_log_level = effective_log_level
+    configure_cli_logging(effective_log_level)
+    emit_structured_log("info", f"Log level set to '{effective_log_level}'.", operation=Operation.SETUP)
+
     # Check minimum version before proceeding
-    cli_version = get_cli_version(quiet=args.quiet)
-    if not should_bypass_min_version_check(args) and not check_min_version(cli_version, quiet=args.quiet):
+    quiet_for_version = effective_log_level in ("warning", "error")
+    cli_version = get_cli_version(quiet=quiet_for_version)
+    if not should_bypass_min_version_check(args) and not check_min_version(cli_version, quiet=quiet_for_version):
         sys.exit(1)
 
     # Validate environment variables required for evaluation
@@ -885,7 +1142,7 @@ def main():
           try:
               auth_handler.clear_cache()
           except Exception as e:
-              print(f"Error during signout: {e}")
+              emit_structured_log("error", f"Error during signout: {e}", operation=Operation.AUTHENTICATE)
               sys.exit(1)
           sys.exit(0)
 
@@ -898,67 +1155,87 @@ def main():
 
         id_token_claims = auth_result.get("id_token_claims")
         if not isinstance(id_token_claims, dict):
-            print("id_token_claims is missing or invalid in authentication result")
+            emit_structured_log(
+                "warning", "id_token_claims is missing or invalid in authentication result",
+                operation=Operation.AUTHENTICATE,
+            )
         else:
             user_oid = id_token_claims.get("oid") or ""
         
       except Exception as e:
-        print(f"\033[91mError during authentication: {e}\033[0m")
-        if args.verbose:
-            import traceback
-            traceback.print_exc()
-        sys.exit(1)
+          emit_structured_log("error", f"Error during authentication: {e}", operation=Operation.AUTHENTICATE)
+          if effective_log_level == "debug":
+              import traceback
+              traceback.print_exc()
+          sys.exit(1)
 
     if not user_oid and access_token:
         # Fallback: extract from access token.
         user_oid = AuthHandler.extract_user_oid_from_access_token(access_token)
 
-    # 1. Load evaluation datasets (prompts and expected_responses)
-    prompts, expected_responses = get_prompt_datasets(args)
+    # 1. Load evaluation datasets
+    eval_items, file_default_evaluators = get_prompt_datasets(args)
+    default_evaluators = resolve_default_evaluators(file_default_evaluators)
+    prompts = [eval_item.get("prompt", "") for eval_item in eval_items]
 
-    if not args.quiet:
-        print(f"Running evaluation on {len(prompts)} prompt(s)...")
+    if effective_log_level in ("info", "debug"):
+        emit_structured_log("info", f"Running evaluation on {len(prompts)} prompt(s).", operation=Operation.SETUP)
 
+    agent_name = None
     try:
         # 3. Agent selection - if no agent ID provided, prompt user to select
         if not args.m365_agent_id:
-            if not args.quiet:
-                print("No agent ID provided. Fetching available agents...")
+            if effective_log_level in ("info", "debug"):
+                emit_structured_log("info", "No agent ID provided. Fetching available agents.", operation=Operation.FETCH_AGENTS)
             
             available_agents = fetch_available_agents(copilot_api_endpoint, access_token, user_oid)
             if not available_agents:
-              print("No agents are available for interactive selection. Please re-run with --m365-agent-id or set the M365_AGENT_ID environment variable.")
-              sys.exit(1)
-            
-            if available_agents:
-                selected_agent_id = select_agent_interactively(available_agents)
-                if selected_agent_id:
-                    args.m365_agent_id = selected_agent_id
-                    if not args.quiet:
-                        print(f"Selected agent: {args.m365_agent_id}")
-                else:
-                    print("No agent selected. Please re-run with --m365-agent-id or set the M365_AGENT_ID environment variable.")
-                    sys.exit(1)
+                emit_structured_log(
+                  "error",
+                  "No agents are available for interactive selection. Re-run with "
+                  "--m365-agent-id or set M365_AGENT_ID.",
+                  operation=Operation.FETCH_AGENTS,
+                )
+                sys.exit(1)
+
+            selected_agent_id, agent_name = select_agent_interactively(available_agents)
+            if selected_agent_id:
+                args.m365_agent_id = selected_agent_id
+                if effective_log_level in ("info", "debug"):
+                    emit_structured_log("info", f"Selected agent: {args.m365_agent_id}", operation=Operation.FETCH_AGENTS)
+            else:
+                emit_structured_log(
+                        "error",
+                        "No agent selected. Re-run with --m365-agent-id or set M365_AGENT_ID.",
+                        operation=Operation.FETCH_AGENTS,
+                )
+                sys.exit(1)
 
         # 4. Send prompts to chat API
         responses = send_prompt_to_agent_in_sydney(prompts, copilot_api_endpoint, access_token, user_oid, args)
     except Exception as e:
-        print(f"\033[91mError sending prompts to chat API: {e}\033[0m")
-        if args.verbose:
+        emit_structured_log("error", f"Error sending prompts to chat API: {e}", operation=Operation.SEND_PROMPT)
+        if effective_log_level == "debug":
             import traceback
             traceback.print_exc()
         sys.exit(1)
+
     
     # 5. Run evaluations
-    if not args.quiet:
-        print("Running evaluations...")
-    results = run_evaluations(args, responses, expected_responses)
+    if effective_log_level in ("info", "debug"):
+        emit_structured_log("info", "Running evaluations.", operation=Operation.EVALUATE)
+    results = run_evaluations(args, responses, eval_items, default_evaluators)
     
     # 6. Output results
-    output_results(results, args)
+    output_results(results, args, default_evaluators=default_evaluators,
+                   agent_name=agent_name, cli_version=str(cli_version) if cli_version else None)
     
-    if not args.quiet:
-        print(f"\nEvaluation completed successfully! Processed {len(prompts)} prompt(s).")
+    if effective_log_level in ("info", "debug"):
+        emit_structured_log(
+            "info",
+            f"Evaluation completed successfully. Processed {len(prompts)} prompt(s).",
+            operation=Operation.EVALUATE,
+        )
 
 # Call the main function when script is run directly
 if __name__ == "__main__":