ibm-watsonx-orchestrate-evaluation-framework 1.1.3__py3-none-any.whl → 1.1.4__py3-none-any.whl

{ibm_watsonx_orchestrate_evaluation_framework-1.1.3.dist-info → ibm_watsonx_orchestrate_evaluation_framework-1.1.4.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ibm-watsonx-orchestrate-evaluation-framework
-Version: 1.1.3
+Version: 1.1.4
 Summary: The WxO evaluation framework
 Author-email: Haode Qi <Haode.Qi@ibm.com>
 License: MIT
@@ -33,3 +33,9 @@ Requires-Dist: notebook~=7.4.1; extra == "rag-eval"
 Requires-Dist: ipywidgets~=8.1.6; extra == "rag-eval"
 Requires-Dist: jupyter_contrib_nbextensions; extra == "rag-eval"
 Requires-Dist: jupyter~=1.1.1; extra == "rag-eval"
+Provides-Extra: script
+Requires-Dist: langchain~=0.3.27; extra == "script"
+Requires-Dist: langchain-core~=0.3.75; extra == "script"
+Requires-Dist: langchain-ibm~=0.3.17; extra == "script"
+Requires-Dist: langchain-ollama~=0.3.7; extra == "script"
+Requires-Dist: rich~=13.9.4; extra == "script"

{ibm_watsonx_orchestrate_evaluation_framework-1.1.3.dist-info → ibm_watsonx_orchestrate_evaluation_framework-1.1.4.dist-info}/RECORD

@@ -1,21 +1,23 @@
 wxo_agentic_evaluation/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-wxo_agentic_evaluation/analyze_run.py,sha256=Ji3aVrEJoF47nkFHdJWp_j3JSqzYAmnLJAg_H2Y-Qgs,13295
+wxo_agentic_evaluation/analyze_run.py,sha256=T1fgy1TY-KnIMB2k6cMNcr0_1eGpX3IYAc8BG_KwkeA,26615
 wxo_agentic_evaluation/annotate.py,sha256=PwgRBAIVBW_yEoOLYNHg9-XVo78zzTXb68kzR2JbtCM,1230
-wxo_agentic_evaluation/arg_configs.py,sha256=KttX3LFPXjg4qRlbeQ-fQ4Qp5-9_Uz5tt4TCx93KRAY,3028
+wxo_agentic_evaluation/arg_configs.py,sha256=417vaTlkPV1VQ57yLY-EnPSdxXizjl6ETZlF76sy190,3048
 wxo_agentic_evaluation/batch_annotate.py,sha256=cZWhDt4k1Tap-e7Mw48tOXKwlHk-JxUvDT6N6n_A7PA,6694
 wxo_agentic_evaluation/data_annotator.py,sha256=ovrymEn2Jlivffg9_mtW7sT73_iOBLU5mX9n5hQQWPo,8398
 wxo_agentic_evaluation/description_quality_checker.py,sha256=Skmt_X-z5rJ9-rBXu5acp0sxq_LyjL0sOOYQVcn25K4,6163
 wxo_agentic_evaluation/evaluation.py,sha256=ZeMmxSbJyA86CVjX8EUdMsVrn25MMqMYO91DZqbe7f0,1090
-wxo_agentic_evaluation/evaluation_package.py,sha256=Ud1h7HDr47Gs4XPUoPagm6oS54Iqb_UWGlcyKoCLnfE,24319
-wxo_agentic_evaluation/inference_backend.py,sha256=mG7Z-Hi63znfJ7vzwCCYNPMc6AHgu7Codnw4puoAM3U,33004
+wxo_agentic_evaluation/evaluation_package.py,sha256=xEtx26FY7w1IqJkoWOS0VngjYyArKESlLhzwLH-vci8,28575
+wxo_agentic_evaluation/inference_backend.py,sha256=P8IzkmFAQoiTPfbuHPUPdE-npK8yNBzsAUJDtc9i9uE,33270
 wxo_agentic_evaluation/llm_matching.py,sha256=HY_4T_4-JXr08Z8o0XWcZfyrzxM0hBpCYGbwh7uSOkw,1479
 wxo_agentic_evaluation/llm_rag_eval.py,sha256=cMUutCpmR1Zg5MM7KbHjHkF42FRBBACFNc-MzwxAw9M,1655
+wxo_agentic_evaluation/llm_safety_eval.py,sha256=pNuq4xLxkImyksGmsQire_nIQWOEoGqCc-Z3ZCSrcTQ,2268
 wxo_agentic_evaluation/llm_user.py,sha256=-DezpQKYoWuN5kQBAx5zwE3Qd_OaxfizZQU7MeqScoM,1519
-wxo_agentic_evaluation/main.py,sha256=5yfynZkzYl52by-7xNMuNdN2FKGEamM-6k-w6fkg6ew,13574
+wxo_agentic_evaluation/main.py,sha256=5WDJN-cpK0Dt0niVKg7b_f9CNTDC54g1psN20MYyGEw,18100
+wxo_agentic_evaluation/main_v2.py,sha256=96pujdcfZJyDo1naGlLAk5mFrrSY0hIxrlH4qTdSCSs,14896
 wxo_agentic_evaluation/quick_eval.py,sha256=7JWBB4Q5psZi2O26wZkQgPudu3uNZZBFrZSYou2ivgw,12876
 wxo_agentic_evaluation/record_chat.py,sha256=uDnc0r5rZdy-KZv36ntBMZMzTiv2pcbHayakg_seZGg,8660
 wxo_agentic_evaluation/resource_map.py,sha256=fmkLcQEx4_tpc46rkSoEOsvmd5WMkxaJpIfgqaR31Ms,1646
-wxo_agentic_evaluation/service_instance.py,sha256=lAwfIRJD20vOZFsmtqBt7z4-AmIWE-Fu5VGjmVeyoso,8506
+wxo_agentic_evaluation/service_instance.py,sha256=Mgr4UjnwYts91J_iLyygubsZw3aLenPnIfKcqz8OrRU,8515
 wxo_agentic_evaluation/test_prompt.py,sha256=ksteXCs9iDQPMETc4Hb7JAXHhxz2r678U6-sgZJAO28,3924
 wxo_agentic_evaluation/tool_planner.py,sha256=RohospVdfYURyFVETgjm1EukmgpNBvBJopUs6obdhn0,14111
 wxo_agentic_evaluation/type.py,sha256=wAqE7sHEOuAD6s-GxLzdPdMyyjNqh-jOuV-KJR5zH5U,4047
@@ -28,10 +30,9 @@ wxo_agentic_evaluation/external_agent/external_validate.py,sha256=gBnizwTIYRHjkV
 wxo_agentic_evaluation/external_agent/performance_test.py,sha256=mLiUsgZlpj6sKZl2lOjb04ts6UOTf7PoOmvLMZrTN1M,2494
 wxo_agentic_evaluation/external_agent/types.py,sha256=Fu_hPBk-vhJ5kAAi6nwVRTrnr0oaxcoV7aXHsJwxYlg,1653
 wxo_agentic_evaluation/metrics/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-wxo_agentic_evaluation/metrics/llm_as_judge.py,sha256=2GvvenWwWn-PV6HAwqL6-L-Wt6jCE8AthQTrtFAh8f4,1218
+wxo_agentic_evaluation/metrics/llm_as_judge.py,sha256=-JtRcCSYIafMRAL1W7mz0oLRySD1Thje8ankbFmCoMQ,1755
 wxo_agentic_evaluation/metrics/metrics.py,sha256=X2Bapjc7aGwU8--evEYOr2x-CNGsMMBTmMP1dXnURUo,6336
 wxo_agentic_evaluation/otel_support/evaluate_tau.py,sha256=RGfaM0jx5WmF4y1EnVsE2FWpfDQEoL3_sIMNcMUbZuQ,2023
-wxo_agentic_evaluation/otel_support/evaluate_tau_traces.py,sha256=gY5m5INv0IQrA4Xi2wigAUI1cnxzGPYtMLWCIo9pubQ,5602
 wxo_agentic_evaluation/otel_support/otel_message_conversion.py,sha256=6fU2nXtMZUiQuhp2w2ByYigeo7wlOmUSo1CEHcb1iqE,649
 wxo_agentic_evaluation/otel_support/tasks_test.py,sha256=Z7NglyL-uI4vzb1Lum9aEdPZ7x_J2g1A-MKEABRX3nU,67543
 wxo_agentic_evaluation/prompt/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -39,23 +40,25 @@ wxo_agentic_evaluation/prompt/answer_relevancy_prompt.jinja2,sha256=vLrMWce-5Hlv
 wxo_agentic_evaluation/prompt/args_extractor_prompt.jinja2,sha256=0qBicXFcc6AA3mQNLPVRmFsnuYaCABJXgZkIH9fO0Js,952
 wxo_agentic_evaluation/prompt/bad_tool_descriptions_prompt.jinja2,sha256=_Ty6QDcQcbde2ZP2HVvFtOCm_2mFu_1cUM6qj11MvcU,8085
 wxo_agentic_evaluation/prompt/batch_testcase_prompt.jinja2,sha256=QXuk2ecnEPPRCPoWZJyrtb1gAVuIPljB91YoqPBp2Dk,1896
+wxo_agentic_evaluation/prompt/derailment_prompt.jinja2,sha256=Q77FVf0-TixFz0_i2-YEh6UwrP0DRNz-cP9NDcDlqpY,1802
 wxo_agentic_evaluation/prompt/faithfulness_prompt.jinja2,sha256=DW9OdjeZJbOWrngRqTAVD4w0va_HtA2FR4G1POIIamM,2524
 wxo_agentic_evaluation/prompt/keyword_matching_prompt.jinja2,sha256=7mTkSrppjgPluUAIMTWaT30K7M4J4hyR_LjSjW1Ofq0,1290
 wxo_agentic_evaluation/prompt/keywords_generation_prompt.jinja2,sha256=PiCjr1ag44Jk5xD3F24fLD_bOGYh2sF0i5miY4OrVlc,1890
-wxo_agentic_evaluation/prompt/llama_user_prompt.jinja2,sha256=GAHtEJvFNtgWBQma1I9KJdhXdhmqbEQf7JY66Z1JLMU,1113
-wxo_agentic_evaluation/prompt/off_policy_attack_generation_prompt.jinja2,sha256=yrLtXfmVIJ_C3XIaTvpqlQGlg9kKIibrVR3UzpzBEmo,1288
+wxo_agentic_evaluation/prompt/llama_user_prompt.jinja2,sha256=o1OfN9ltWUzSyisZBJNdYC3PCI5pImPLgjQD1iOf8UQ,4651
+wxo_agentic_evaluation/prompt/off_policy_attack_generation_prompt.jinja2,sha256=4ZzRfyXbyGbos_XpN_YhvbMFSzlyWxlPtG3qYfqBYbM,1289
 wxo_agentic_evaluation/prompt/on_policy_attack_generation_prompt.jinja2,sha256=90KF7fXW5PPwY8IkPqA6ZflDMkr_KFDpO9H_mVGdGf8,2212
 wxo_agentic_evaluation/prompt/semantic_matching_prompt.jinja2,sha256=MltPfEXYyOwEC2xNLl7UsFTxNbr8CwHaEcPqtvKE2r8,2749
 wxo_agentic_evaluation/prompt/starting_sentence_generation_prompt.jinja2,sha256=m_l6f7acfnWJmGQ0mXAy85oLGLgzhVhoz7UL1FVYq8A,4908
 wxo_agentic_evaluation/prompt/story_generation_prompt.jinja2,sha256=_DxjkFoHpNTmdVSUzUrUdwn4Cng7nAGqkMnm0ScOH1w,4191
-wxo_agentic_evaluation/prompt/template_render.py,sha256=xVy7NOeGk5_XxzTT-YIY4HVAseQFU2SbRMSdvQGa-FE,4829
+wxo_agentic_evaluation/prompt/template_render.py,sha256=gQy3j3RkoPJjlCip2-sq2vz0b7_adCpqnZ-uKGDU-gQ,5282
 wxo_agentic_evaluation/prompt/tool_chain_agent.jinja2,sha256=9RcIjLYoOvtFsf-RgyMfMcj2Fe8fq1wGkE4nG1zamYY,297
 wxo_agentic_evaluation/prompt/tool_planner.jinja2,sha256=Ln43kwfSX50B1VBsT-MY1TCE0o8pGFh8aQJAzZfGkpI,3239
+wxo_agentic_evaluation/prompt/unsafe_topic_prompt.jinja2,sha256=swxhd_9mxoRSNtvumup40bKdKDb8O_YMv6unytGJxdc,2447
 wxo_agentic_evaluation/prompt/examples/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 wxo_agentic_evaluation/prompt/examples/data_simple.json,sha256=XXF-Pn-mosklC9Ch7coyaJxosFNnl3OkHSW3YPuiKMM,2333
-wxo_agentic_evaluation/red_teaming/attack_evaluator.py,sha256=pfhMUjddv32pIRewea7o1vn_xrV_LuyC8vRlJ7qVyO8,5267
+wxo_agentic_evaluation/red_teaming/attack_evaluator.py,sha256=bG3VBqmWoGySHfnkfv-2PPnwaktFQ9jeC3kCJzSMQ4k,8315
 wxo_agentic_evaluation/red_teaming/attack_generator.py,sha256=Sz9zB5O1ct7EoZCog8GNdwj8yWFZo7HJLPbA9HvelZc,11886
-wxo_agentic_evaluation/red_teaming/attack_list.py,sha256=edphWARWqDtXFtcHTVbRXngvO0YfG5SgrfPtrBRXuFw,4734
+wxo_agentic_evaluation/red_teaming/attack_list.py,sha256=yRnxwebrr7cl_3Rr3e3Xgdt0luCEYICvg00kutj47fo,9478
 wxo_agentic_evaluation/red_teaming/attack_runner.py,sha256=XXNP43mEneuDBo_zGPdCVNRdUNy-KGd7kbIKYwKhKJQ,4477
 wxo_agentic_evaluation/referenceless_eval/__init__.py,sha256=lijXMgQ8nQe-9eIfade2jLfHMlXfYafMZIwXtC9KDZo,106
 wxo_agentic_evaluation/referenceless_eval/referenceless_eval.py,sha256=G2b7rwN0VTLBVGwU1VXKUl4eqT8Ya8zCcOorwkZwrZA,4354
@@ -95,8 +98,8 @@ wxo_agentic_evaluation/utils/__init__.py,sha256=QMxk6hx1CDvCBLFh40WpPZmqFNJtDqwX
 wxo_agentic_evaluation/utils/open_ai_tool_extractor.py,sha256=kJUuprXfY5IfCRIvLT4oSvMf-Ly7RYNo4if-Lb6yGiA,6080
 wxo_agentic_evaluation/utils/rich_utils.py,sha256=pJ43hwvSZRJwckPOeUhqGdw4_RwxLsYO02dDt6PLqxA,6351
 wxo_agentic_evaluation/utils/rouge_score.py,sha256=WvcGh6mwF4rWH599J9_lAt3BfaHbAZKtKEJBsC61iKo,692
-wxo_agentic_evaluation/utils/utils.py,sha256=8PUpmOoPrEG5xBDOWMsaKanYsnZV5-UZWQa7x8P-J2g,11634
-ibm_watsonx_orchestrate_evaluation_framework-1.1.3.dist-info/METADATA,sha256=SRO-KH4zJYQhHMhyhDIqrkeoELwrDnTvYbwcIZT9i9w,1435
-ibm_watsonx_orchestrate_evaluation_framework-1.1.3.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-ibm_watsonx_orchestrate_evaluation_framework-1.1.3.dist-info/top_level.txt,sha256=2okpqtpxyqHoLyb2msio4pzqSg7yPSzwI7ekks96wYE,23
-ibm_watsonx_orchestrate_evaluation_framework-1.1.3.dist-info/RECORD,,
+wxo_agentic_evaluation/utils/utils.py,sha256=yDPF0hsd_ypMUanf4AZOQbbBh5KhjTc_VOgIcQ-6htI,12682
+ibm_watsonx_orchestrate_evaluation_framework-1.1.4.dist-info/METADATA,sha256=3HNLootsqOLwTThaqwx33iaBDlbSh1UgoUEBRRra1LE,1728
+ibm_watsonx_orchestrate_evaluation_framework-1.1.4.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+ibm_watsonx_orchestrate_evaluation_framework-1.1.4.dist-info/top_level.txt,sha256=2okpqtpxyqHoLyb2msio4pzqSg7yPSzwI7ekks96wYE,23
+ibm_watsonx_orchestrate_evaluation_framework-1.1.4.dist-info/RECORD,,

wxo_agentic_evaluation/analyze_run.py

@@ -1,6 +1,7 @@
 import csv
 import json
 import os
+import re
 from pathlib import Path
 from typing import Dict, List, Optional, Set
 
@@ -15,7 +16,10 @@ from wxo_agentic_evaluation.arg_configs import AnalyzeConfig
 from wxo_agentic_evaluation.description_quality_checker import (
     DescriptionQualityInspector,
 )
-from wxo_agentic_evaluation.metrics.metrics import ToolCallAndRoutingMetrics
+from wxo_agentic_evaluation.metrics.metrics import (
+    TextMatchType,
+    ToolCallAndRoutingMetrics,
+)
 from wxo_agentic_evaluation.type import (
     ContentType,
     ExtendedMessage,
@@ -28,7 +32,11 @@ from wxo_agentic_evaluation.utils.rich_utils import (
     print_done,
     warn,
 )
-from wxo_agentic_evaluation.utils.utils import add_line_seperator
+from wxo_agentic_evaluation.utils.utils import (
+    add_line_seperator,
+    list_run_files,
+    load_run_metrics,
+)
 
 
 class Analyzer:
@@ -136,7 +144,10 @@ class Analyzer:
         return analysis_for_display
 
     def render(
-        self, data: List[ExtendedMessage], tool_definition_path: Optional[str]
+        self,
+        data: List[ExtendedMessage],
+        tool_definition_path: Optional[str],
+        meta: Optional[dict] = None,
     ) -> Group:
         """
         Render the conversation history and analysis results.
@@ -147,6 +158,8 @@ class Analyzer:
         conversation_lines = []
         reason_lines = []
         failing_tools = []
+        added_errors_header = False
+        added_missed_header = False
 
         for entry in data:
             msg = entry.message
@@ -179,6 +192,11 @@ class Analyzer:
 
             text_line = Text(f"{label}: {content}\n")
             if reason:
+                if not added_errors_header:
+                    reason_lines.append(
+                        Text("\nTool Call Errors:\n", style="bold red")
+                    )
+                    added_errors_header = True
                 text_line.stylize("bold red")
                 reason_text = f"❌ {tool_name}: {json.dumps(reason)}\n\n"
                 reason_lines.append(Text(reason_text, style="red"))
@@ -199,6 +217,17 @@ class Analyzer:
             if description_quality_inspection_lines:
                 reason_lines.extend(description_quality_inspection_lines)
 
+        if meta:
+            missed = meta.get("missed_tool_calls") or []
+            if missed:
+                if not added_missed_header:
+                    reason_lines.append(
+                        Text("\nMissed Calls:\n", style="bold red")
+                    )
+                    added_missed_header = True
+                for tool in missed:
+                    reason_lines.append(Text(f"❌ {tool}\n", style="red"))
+
         conversation_panel = Panel(
             Text().join(conversation_lines),
             title="Conversation History",
@@ -238,6 +267,7 @@ class Analyzer:
 
         def get_test_messages(test_case_name):
             test_messages = []
+            meta = {}
 
             test_case_path = os.path.join(
                 config.data_path,
@@ -247,11 +277,15 @@ class Analyzer:
 
             with open(test_case_path, "r", encoding="utf-8") as f:
                 temp = json.load(f)
+                if temp and isinstance(temp[-1], dict) and "meta" in temp[-1]:
+                    meta = temp[-1]["meta"]
+                    temp = temp[:-1]
+
                 for entry in temp:
                     msg = ExtendedMessage(**entry)
                     test_messages.append(msg)
 
-            return test_messages
+            return test_messages, meta
 
         def get_metrics(test_case_name):
             test_metrics_path = os.path.join(
@@ -281,32 +315,324 @@ class Analyzer:
 
             pretty_print(panel)
 
-        for test_case_entry in test_case_with_failed_tools:
-            test_case_name = test_case_entry["dataset_name"]
+        messages_dir = os.path.join(config.data_path, "messages")
 
-            test_messages = get_test_messages(test_case_name=test_case_name)
+        RUN_NAME_ONLY_RE = re.compile(r"^(?P<parent>.+)\.run(?P<id>\d+)$")
+        processed_parents: Set[str] = set()
 
-            metrics: ToolCallAndRoutingMetrics = get_metrics(
-                test_case_name=test_case_name
-            )
+        overall_runs_performed = 0
+        overall_runs_problematic = 0
+        overall_text_match_hits = 0
+        overall_text_match_den = 0
+        overall_journey_vals = []
 
-            header_panel = self._create_header_analysis_panel(
-                test_case_name, metrics
+        for test_case_entry in summary:
+            dataset_base = test_case_entry["dataset_name"]
+
+            # If CSV row looks like "<parent>.runN" and we have runs on disk for <parent>, skip the per-run row.
+            m = RUN_NAME_ONLY_RE.match(dataset_base)
+            if m:
+                parent = m.group("parent")
+                if list_run_files(messages_dir, parent):
+                    continue
+
+            # Avoid processing a parent twice if it appears multiple times in CSV.
+            if dataset_base in processed_parents:
+                continue
+
+            run_map = list_run_files(messages_dir, dataset_base)
+
+            # ---- SINGLE RUN (legacy or run1 only) ----
+            if not run_map or len(run_map) == 1:
+                if not run_map:
+                    # Legacy single-run files
+                    test_messages, meta = get_test_messages(
+                        test_case_name=dataset_base
+                    )
+                    metrics: ToolCallAndRoutingMetrics = get_metrics(
+                        test_case_name=dataset_base
+                    )
+                    runs_performed = 1
+                else:
+                    run_id = next(iter(run_map))
+                    paths = run_map[run_id]
+                    runs_performed = 1
+                    if not paths["metrics"]:
+                        pretty_print(
+                            f"❌ {dataset_base}.run{run_id} — metrics file missing.",
+                            style="bold red",
+                        )
+                        # Count it as analyzed & problematic
+                        processed_parents.add(dataset_base)
+                        ds_table = Table(show_header=False, box=None)
+                        ds_table.add_row("Type: Single-run")
+                        ds_table.add_row("Status: ❌ Problematic")
+                        pretty_print(
+                            Panel(
+                                ds_table,
+                                title=f"📋 Analysis Summary — {dataset_base}",
+                                border_style="green",
+                            )
+                        )
+                        overall_runs_performed += 1
+                        overall_runs_problematic += 1
+                        add_line_seperator(self._generate_style_config())
+                        continue
+
+                    metrics = load_run_metrics(paths["metrics"])
+                    meta = {}
+
+                    if paths["analyze"]:
+                        with open(paths["analyze"], "r", encoding="utf-8") as f:
+                            raw = json.load(f)
+                        if (
+                            raw
+                            and isinstance(raw[-1], dict)
+                            and "meta" in raw[-1]
+                        ):
+                            meta = raw[-1]["meta"]
+                            raw = raw[:-1]
+                        test_messages = [
+                            ExtendedMessage(**entry) for entry in raw
+                        ]
+                    else:
+                        test_messages, meta = [], {}
+
+                # --- compute status uniformly (legacy & run1) ---
+                had_incorrect_param = (
+                    hasattr(metrics, "tool_calls_with_incorrect_parameter")
+                    and float(metrics.tool_calls_with_incorrect_parameter or 0)
+                    > 0
+                )
+                low_precision = (
+                    hasattr(metrics, "tool_call_precision")
+                    and float(
+                        metrics.tool_call_precision
+                        if metrics.tool_call_precision is not None
+                        else 1.0
+                    )
+                    < 1.0
+                )
+                low_recall = (
+                    hasattr(metrics, "tool_call_recall")
+                    and float(
+                        metrics.tool_call_recall
+                        if metrics.tool_call_recall is not None
+                        else 1.0
+                    )
+                    < 1.0
+                )
+                runs_problematic = (
+                    1
+                    if (
+                        (
+                            hasattr(metrics, "is_success")
+                            and not metrics.is_success
+                        )
+                        or had_incorrect_param
+                        or low_precision
+                        or low_recall
+                    )
+                    else 0
+                )
+
+                processed_parents.add(dataset_base)
+
+                # ✅ Dataset-level panel (print BEFORE details)
+                ds_table = Table(show_header=False, box=None)
+                ds_table.add_row("Type: Single-run")
+                status = (
+                    "❌ Problematic" if runs_problematic else "✅ No problems"
+                )
+                ds_table.add_row(f"Status: {status}")
+                pretty_print(
+                    Panel(
+                        ds_table,
+                        title=f"📋 Analysis Summary — {dataset_base}",
+                        border_style="green",
+                    )
+                )
+
+                # Update overall counters/averages
+                overall_runs_performed += runs_performed
+                overall_runs_problematic += runs_problematic
+                tm = getattr(metrics, "text_match", None)
+                tm_val = getattr(tm, "value", None) if tm else None
+
+                if tm_val is not None and tm_val != TextMatchType.na:
+                    overall_text_match_den += 1
+                    overall_text_match_hits += (
+                        tm_val == TextMatchType.text_match
+                    )
+                if getattr(metrics, "is_success", None) is not None:
+                    overall_journey_vals.append(
+                        1 if bool(metrics.is_success) else 0
+                    )
+
+                # Replay details only if problematic
+                if runs_problematic:
+                    pretty_print(
+                        self._create_header_analysis_panel(
+                            dataset_base, metrics
+                        )
+                    )
+                    pretty_print(
+                        self.render(
+                            test_messages, config.tool_definition_path, meta
+                        )
+                    )
+                    add_line_seperator(self._generate_style_config())
+
+                continue
+
+            # ---- MULTI RUN (two-pass: compute first, then print summary, then details) ----
+            processed_parents.add(dataset_base)
+            runs_performed = len(run_map)
+            runs_problematic = 0
+            text_match_hits = 0
+            text_match_den = 0
+            journey_vals = []
+
+            # First pass: compute aggregates and collect problematic runs to replay later
+            deferred_runs = []
+            for run_id in sorted(run_map):
+                paths = run_map[run_id]
+                if not paths["metrics"]:
+                    runs_problematic += 1
+                    # no analyze file to replay; still counted as problematic
+                    continue
+
+                metrics = load_run_metrics(paths["metrics"])
+
+                # Aggregate for per-dataset
+                tm = getattr(metrics, "text_match", None)
+                tm_val = getattr(tm, "value", None) if tm is not None else None
+                if tm_val is not None and tm_val != TextMatchType.na.value:
+                    text_match_den += 1
+                    text_match_hits += tm_val == TextMatchType.text_match.value
+
+                if getattr(metrics, "is_success", None) is not None:
+                    journey_vals.append(1 if bool(metrics.is_success) else 0)
+
+                # Decide if problematic
+                had_incorrect_param = (
+                    hasattr(metrics, "tool_calls_with_incorrect_parameter")
+                    and float(metrics.tool_calls_with_incorrect_parameter or 0)
+                    > 0
+                )
+                low_precision = (
+                    hasattr(metrics, "tool_call_precision")
+                    and float(
+                        metrics.tool_call_precision
+                        if metrics.tool_call_precision is not None
+                        else 1.0
+                    )
+                    < 1.0
+                )
+                low_recall = (
+                    hasattr(metrics, "tool_call_recall")
+                    and float(
+                        metrics.tool_call_recall
+                        if metrics.tool_call_recall is not None
+                        else 1.0
+                    )
+                    < 1.0
+                )
+
+                is_problem = (
+                    (hasattr(metrics, "is_success") and not metrics.is_success)
+                    or had_incorrect_param
+                    or low_precision
+                    or low_recall
+                )
+                if is_problem:
+                    runs_problematic += 1
+                    deferred_runs.append(
+                        {
+                            "title": f"{dataset_base}.run{run_id}",
+                            "metrics": metrics,
+                            "analyze_path": paths.get("analyze"),
+                        }
+                    )
+
+            # Print the dataset panel FIRST with both lines inside
+            ds_table = Table(show_header=False, box=None)
+            ds_table.add_row(f"Type: Multi-run ({runs_performed} runs)")
+            ds_table.add_row(
+                f"Runs with problems: {runs_problematic} / {runs_performed}"
+            )
+            status = (
+                "❌ Problematic" if runs_problematic > 0 else "✅ No problems"
+            )
+            ds_table.add_row(f"Status: {status}")
+            pretty_print(
+                Panel(
+                    ds_table,
+                    title=f"📋 Analysis Summary — {dataset_base}",
+                    border_style="green",
+                )
             )
-            pretty_print(header_panel)
 
-            tool_definition_path = (
-                config.tool_definition_path
-                if config.tool_definition_path
-                else None
+            # Second pass: now replay only the problematic runs (so summary stays at the top)
+            for item in deferred_runs:
+                pretty_print(
+                    self._create_header_analysis_panel(
+                        item["title"], item["metrics"]
+                    )
+                )
+                if item["analyze_path"]:
+                    with open(item["analyze_path"], "r", encoding="utf-8") as f:
+                        raw = json.load(f)
+                    meta = {}
+                    if raw and isinstance(raw[-1], dict) and "meta" in raw[-1]:
+                        meta = raw[-1]["meta"]
+                        raw = raw[:-1]
+                    test_messages = [ExtendedMessage(**entry) for entry in raw]
+
+                    pretty_print(
+                        self.render(
+                            test_messages, config.tool_definition_path, meta
+                        )
+                    )
+                add_line_seperator(self._generate_style_config())
+
+            # Update overall aggregates
+            overall_runs_performed += runs_performed
+            overall_runs_problematic += runs_problematic
+            overall_text_match_hits += text_match_hits
+            overall_text_match_den += text_match_den
+            overall_journey_vals.extend(journey_vals)
+
+        # --- Overall summary ---
+        overall_lines = [
+            f"Test cases analyzed: {len(processed_parents)}",
+            f"Total runs executed: {overall_runs_performed}",
+            f"Problematic runs: {overall_runs_problematic} ({round((overall_runs_problematic/overall_runs_performed)*100,1) if overall_runs_performed else 0}%)",
+        ]
+
+        if overall_text_match_den:
+            tm_pct = round(
+                (overall_text_match_hits / overall_text_match_den) * 100, 2
             )
+            overall_lines.append(f"Avg text-match success: {tm_pct}%")
+        else:
+            overall_lines.append("Avg text-match success: N/A")
 
-            rendered_content = self.render(
-                data=test_messages, tool_definition_path=tool_definition_path
+        if overall_journey_vals:
+            js_pct = round(
+                (sum(overall_journey_vals) / len(overall_journey_vals)) * 100, 2
             )
-            pretty_print(rendered_content)
+            overall_lines.append(f"Avg journey success: {js_pct}%")
+        else:
+            overall_lines.append("Avg journey success: N/A")
 
-            add_line_seperator(self._generate_style_config())
+        pretty_print(
+            Panel(
+                Text("\n".join(overall_lines)),
+                title="📋 Overall Summary",
+                border_style="cyan",
+            )
+        )
 
     def _create_header_analysis_panel(
         self, test_case_name: str, metrics: ToolCallAndRoutingMetrics
@@ -324,7 +650,7 @@ class Analyzer:
         header_table.add_row(f"Journey Success: {metrics.is_success}")
 
         header_panel = Panel(
-            header_table, title="[bold green]📋 Analysis Summary[/bold green]"
+            header_table, title="[bold green]Test Case Summary[/bold green]"
         )
 
         return header_panel
@@ -339,13 +665,16 @@ class Analyzer:
             if test_case_name.lower().strip() == "summary (average)":
                 continue
 
-            if (
-                not entry["is_success"]
-                or float(entry["tool_calls_with_incorrect_parameter"]) > 0
-                or float(entry["tool_call_precision"]) < 1.0
-                or float(entry["tool_call_recall"]) < 1.0
-            ):
+            is_success = str(entry["is_success"]).strip().lower() == "true"
+
+            tip = float(
+                entry.get("tool_calls_with_incorrect_parameter", 0) or 0
+            )
+            tcp = float(entry.get("tool_call_precision", 1) or 1)
+            tcr = float(entry.get("tool_call_recall", 1) or 1)
 
+            # Apply the 4 checks
+            if (not is_success) or (tip > 0) or (tcp < 1.0) or (tcr < 1.0):
                 test_case_with_failed_tools.append(entry)
 
         return test_case_with_failed_tools

wxo_agentic_evaluation/arg_configs.py

@@ -46,6 +46,7 @@ class TestConfig:
     skip_available_results: bool = False
     data_annotation_run: bool = False
     num_workers: int = 2
+    n_runs: int = 1
 
 
 @dataclass