ibm-watsonx-orchestrate-evaluation-framework 1.0.2__py3-none-any.whl → 1.0.4__py3-none-any.whl

{ibm_watsonx_orchestrate_evaluation_framework-1.0.2.dist-info → ibm_watsonx_orchestrate_evaluation_framework-1.0.4.dist-info}/METADATA

@@ -1,12 +1,11 @@
 Metadata-Version: 2.4
 Name: ibm-watsonx-orchestrate-evaluation-framework
-Version: 1.0.2
+Version: 1.0.4
 Summary: The WxO evaluation framework
 Author-email: Haode Qi <Haode.Qi@ibm.com>
 License: MIT
 Requires-Python: <3.14,>=3.11
 Description-Content-Type: text/markdown
-License-File: LICENSE
 Requires-Dist: rich~=13.9.4
 Requires-Dist: pydantic~=2.10.6
 Requires-Dist: pyyaml~=6.0.2
@@ -32,7 +31,6 @@ 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"
-Dynamic: license-file
 
 # WXO-agent evaluation framework
 
@@ -46,7 +44,8 @@ Dynamic: license-file
 - The agent calls the `runs/` endpoint of the wxo-lite server instance, and the actual tool code is executed on the server side. The server database is not visible to our framework. 
 
 ## prerequisite
-Follow the [SDK setup guide](https://github.ibm.com/WatsonOrchestrate/wxo-clients/tree/main) to install the SDK. Make sure you are using version 1.2.0 of the SDK as this is the version this framework requires.
+Follow the [SDK setup guide](https://github.ibm.com/WatsonOrchestrate/wxo-clients/tree/main) to install the SDK.    
+The current framework is compatible with ADK version >= 1.20, <= 1.6.0
 
 ## setup for evaluation framework
 Run the following command to install evaluation framework in the same env:
@@ -59,9 +58,11 @@ pip install -e .
 ```bash
 orchestrate server start
 export WATSONX_SPACE_ID=""
-export WATSONX_API_KEY=""
+export WATSONX_APIKEY=""
 ```
 
+NOTE: If you want to use `WO_INSTANCE` and `WO_API_KEY` instead, follow the [model proxy section](#using-model-proxy-provider).
+
 Import sample hr tools and agent into your default `wxo-dev` env:
 ```bash
 orchestrate tools import -f benchmarks/hr_sample/tools.py -k python
@@ -97,7 +98,6 @@ Note:
       ]
     }
   ],
-  "mine_fields": [],
   "story": "Your username is nwaters and you want to find out timeoff schedule from 20250101 to 20250303."
 }
 ```
@@ -124,7 +124,7 @@ NOTE: run `orchestrate env list` to find the name of the active tenant. for defa
 4. Run the test:
 ```bash
 export WATSONX_SPACE_ID=""
-export WATSONX_API_KEY=""
+export WATSONX_APIKEY=""
 python -m wxo_agentic_evaluation.main --config benchmarks/hr_sample/config.yaml
 ```
 
@@ -237,6 +237,69 @@ python -m wxo_agentic_evaluation.main --config benchmarks/hr_sample/config.yaml
 
 For full instructions on setting up tools, writing stories, configuring the pipeline, and generating batch test cases, see the [Batch Test case Generation Guide](./benchmarks/batch_sample/README.MD).
 
+## Using Model Proxy Provider
+
+To use the model proxy provider (which allows direct access to LLM models), follow these steps:
+
+1. Set up environment variables:
+   ```sh
+   export WO_INSTANCE=<your-instance-url>
+   export WO_API_KEY=<your-api-key>
+   ```
+
+2. Create a configuration file similar to [benchmarks/hr_sample/config_model_proxy.yaml](benchmarks/hr_sample/config_model_proxy.yaml):
+   ```yaml
+   test_paths:
+     - <your-test-path>
+
+   auth_config:
+     url: http://localhost:4321
+     tenant_name: wxo-dev
+
+   provider_config:
+     provider: "model_proxy"
+     model_id: "<model-id>"
+
+   output_dir: "<output-dir>"
+   ```
+
+3. Run the evaluation:
+   ```sh
+   python -m wxo_agentic_evaluation.main --config path/to/your/config.yaml
+   ```
+
+## Using Ollama
+
+To use model from Ollama (local LLM deployment), follow these steps:
+
+1. Make sure you have [Ollama](https://ollama.com) installed and running on your system.
+
+2. Pull your desired model using Ollama (e.g. llama3.1:8b):
+   ```sh
+   ollama pull <model-id>
+   ```
+
+3. Create a configuration file similar to [benchmarks/hr_sample/config_ollama.yaml](benchmarks/hr_sample/config_ollama.yaml):
+   ```yaml
+   test_paths:
+     - <your-test-path>
+
+   auth_config:
+     url: http://localhost:4321
+     tenant_name: wxo-dev
+
+   provider_config:
+     provider: "ollama"
+     model_id: "<model-id>"
+
+   output_dir: "results/ollama/<model-name>"
+   ```
+
+4. Run the evaluation:
+   ```sh
+   python -m wxo_agentic_evaluation.main --config path/to/your/config.yaml
+   ```
+
 ## Workflow diagram
 
 To help better understand the workflow, this is a diagram of how this repo works together with wxO lite python SDK and a wxO runtime.

ibm_watsonx_orchestrate_evaluation_framework-1.0.4.dist-info/RECORD

@@ -0,0 +1,56 @@
+wxo_agentic_evaluation/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+wxo_agentic_evaluation/analyze_run.py,sha256=C4HowEukNMM-H8FkRcHRqkiNYIQVCoTKbBLiqr1cFRM,4332
+wxo_agentic_evaluation/annotate.py,sha256=nxYMc6gwfQ-GuNjCPFtbX_-Es5-9XDdbXpMH89yRDdc,1228
+wxo_agentic_evaluation/arg_configs.py,sha256=UCrGcakFaAM3reFquMn03qNtKe7Pg8ScbOF0K7o8VDU,2240
+wxo_agentic_evaluation/batch_annotate.py,sha256=44K4DUI498uaLIWUn3nz82AKcU6VnCjrExoG6GpPHoM,6323
+wxo_agentic_evaluation/data_annotator.py,sha256=DJVG2CdhJRAJ3X1ARbrsn9bPjTuytCDGIBM4PEexfnk,8214
+wxo_agentic_evaluation/evaluation_package.py,sha256=jOSe-TCJdAWCk1sWpRYfi_EMkZERrVf5swm-bxfozzc,21333
+wxo_agentic_evaluation/inference_backend.py,sha256=fhEB1kaNN-A08RtJglBiv3QL_8nq8m-g7xbF4WbHAvU,25691
+wxo_agentic_evaluation/llm_matching.py,sha256=l010exoMmsvTIAVHCm-Ok0diyeQogjCmemUb9rJLe6A,1477
+wxo_agentic_evaluation/llm_rag_eval.py,sha256=vsNGz1cFE5QGdhnfrx-iJq1r6q8tSI9Ef1mzuhoHElg,1642
+wxo_agentic_evaluation/llm_user.py,sha256=0zSsyEM7pYQtLcfbnu0gEIkosHDwntOZY84Ito6__SM,1407
+wxo_agentic_evaluation/main.py,sha256=tRXVle2o1JhwJZOTpqdsOzBOpxPYxAH5ziZkbCmzfyU,11470
+wxo_agentic_evaluation/record_chat.py,sha256=ZaOxIabDcE_CzZjKJESgh8LY7pK9UT4OvqQMFVdTG7A,8102
+wxo_agentic_evaluation/resource_map.py,sha256=-dIWQdpEpPeSCbDeYfRupG9KV1Q4NlHGb5KXywjkulM,1645
+wxo_agentic_evaluation/service_instance.py,sha256=yt7XpwheaRRG8Ri4TFIS5G2p5mnCwvNgj6T7bDF5uTU,6494
+wxo_agentic_evaluation/test_prompt.py,sha256=ksteXCs9iDQPMETc4Hb7JAXHhxz2r678U6-sgZJAO28,3924
+wxo_agentic_evaluation/tool_planner.py,sha256=e-lBb4w1klT1HOL9BTwae3lkGv5VBuYC397mSJgOhus,12622
+wxo_agentic_evaluation/type.py,sha256=uVKim70XgPW-3L7Z0yRO07wAH9xa-NcjfaiIyPhYMR0,3413
+wxo_agentic_evaluation/analytics/tools/analyzer.py,sha256=IPX_lAFujjPVI9fhXTNohXTxTmpqRhfzQygCWDYHBHg,18125
+wxo_agentic_evaluation/analytics/tools/main.py,sha256=ocwPUlEjyK7PMdXBg5OM2DVDQBcaHT4UjR4ZmEhR0C4,6567
+wxo_agentic_evaluation/analytics/tools/types.py,sha256=IFLKI1CCQwPR2iWjif8AqL_TEq--VbLwdwnMqfJujBw,4461
+wxo_agentic_evaluation/analytics/tools/ux.py,sha256=EaWNvsq68X_i2H4pQ2fABtXEEmk3ZXqaMrTs42_7MwE,18347
+wxo_agentic_evaluation/external_agent/__init__.py,sha256=LY3gMNzfIEwjpQkx5_2iZFHGQiUL4ymEkKL1dc2uKq4,1491
+wxo_agentic_evaluation/external_agent/external_validate.py,sha256=xW8tqPcm8JYvveSxf-oFCajvF5J8ORaK23YXu-LuFmc,4142
+wxo_agentic_evaluation/external_agent/performance_test.py,sha256=bCXUsW0OeUzwfSSYObgfAmEU5vARkD-PblYU-mU9aPY,2507
+wxo_agentic_evaluation/external_agent/types.py,sha256=4kfWD_ZyGZmpbib33gCxEuKS4HLb7CEtferlQgQe7uk,1624
+wxo_agentic_evaluation/metrics/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+wxo_agentic_evaluation/metrics/llm_as_judge.py,sha256=bybJQfVWiVh3BoFEZjdBmU9EQO9Ukheu3YWmkI9b1ks,1218
+wxo_agentic_evaluation/metrics/metrics.py,sha256=9O2m6T2iW-PMjGrTdMbOHP2Pr4RN0NwbEp6YgFpTi3I,5572
+wxo_agentic_evaluation/prompt/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+wxo_agentic_evaluation/prompt/answer_relevancy_prompt.jinja2,sha256=vLrMWce-5HlvniCQdtifnl-YdbJfT8-oixzfwulZs98,3839
+wxo_agentic_evaluation/prompt/args_extractor_prompt.jinja2,sha256=0qBicXFcc6AA3mQNLPVRmFsnuYaCABJXgZkIH9fO0Js,952
+wxo_agentic_evaluation/prompt/batch_testcase_prompt.jinja2,sha256=QXuk2ecnEPPRCPoWZJyrtb1gAVuIPljB91YoqPBp2Dk,1896
+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=nDfCD0o9cRYmsgIjzD-RZNQxotlvuqrzdsZIY-vT794,684
+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=FVH5ew2TofC5LGqQzqNj90unrxooUZv_5XxJzVdz8uM,3563
+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/examples/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+wxo_agentic_evaluation/prompt/examples/data_simple.json,sha256=XXF-Pn-mosklC9Ch7coyaJxosFNnl3OkHSW3YPuiKMM,2333
+wxo_agentic_evaluation/service_provider/__init__.py,sha256=EaY4jjKp58M3W8N3b3a8PNC2S81xA7YV2_QkTIy9DfI,1600
+wxo_agentic_evaluation/service_provider/model_proxy_provider.py,sha256=X5tiE0IKCR2CqhwEGm91LOdzFZQWSXzXQgLOtzi6ng0,4002
+wxo_agentic_evaluation/service_provider/ollama_provider.py,sha256=HMHQVUGFbLSQI1dhysAn70ozJl90yRg-CbNd4vsz-Dc,1116
+wxo_agentic_evaluation/service_provider/provider.py,sha256=MsnRzLYAaQiU6y6xf6eId7kn6-CetQuNZl00EP-Nl28,417
+wxo_agentic_evaluation/service_provider/watsonx_provider.py,sha256=iKVkWs4PRTM_S0TIdPgQ9NFQWPlDvcEvuHpQlIPzO10,6216
+wxo_agentic_evaluation/utils/__init__.py,sha256=QMxk6hx1CDvCBLFh40WpPZmqFNJtDqwXP7S7cXD6NQE,145
+wxo_agentic_evaluation/utils/utils.py,sha256=JYZQZ-OBy43gAWg9S7duJi9StRApGJATs2JUsW1l30M,6057
+ibm_watsonx_orchestrate_evaluation_framework-1.0.4.dist-info/METADATA,sha256=uhmuzKUbgWgKDNayG2dAc-YYvZ_ypeVY4onrcomv0Co,17667
+ibm_watsonx_orchestrate_evaluation_framework-1.0.4.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+ibm_watsonx_orchestrate_evaluation_framework-1.0.4.dist-info/top_level.txt,sha256=2okpqtpxyqHoLyb2msio4pzqSg7yPSzwI7ekks96wYE,23
+ibm_watsonx_orchestrate_evaluation_framework-1.0.4.dist-info/RECORD,,

wxo_agentic_evaluation/analytics/tools/analyzer.py

@@ -1,9 +1,9 @@
-from type import Message, ContentType, EvaluationData
+from wxo_agentic_evaluation.type import Message, ContentType, EvaluationData
 from typing import List, Optional
 import json
 import rich
 from collections import defaultdict
-from analytics.tools.types import (
+from wxo_agentic_evaluation.analytics.tools.types import (
     ErrorPatterns,
     ToolFailure,
     HallucinatedParameter,
@@ -15,7 +15,7 @@ from analytics.tools.types import (
     AnalysisResults,
     ErrorType,
 )
-from data_annotator import ERROR_KEYWORDS
+from wxo_agentic_evaluation.data_annotator import ERROR_KEYWORDS
 from http import HTTPStatus
 
 

wxo_agentic_evaluation/analytics/tools/ux.py

@@ -5,7 +5,7 @@ from rich.table import Table
 from rich.panel import Panel
 from rich.align import Align
 from rich.console import Group
-from type import Message, ContentType
+from wxo_agentic_evaluation.type import Message, ContentType
 from typing import List, Dict, Optional
 from analytics.tools.types import (
     ToolDefinitionRecommendation,

wxo_agentic_evaluation/analyze_run.py

@@ -9,9 +9,9 @@ from rich.table import Table
 from typing import List
 from wxo_agentic_evaluation.type import (
     ExtendedMessage,
-    ContentType,
-    ToolCallAndRoutingMetrics,
+    ContentType
 )
+from wxo_agentic_evaluation.metrics.metrics import ToolCallAndRoutingMetrics
 from wxo_agentic_evaluation.arg_configs import AnalyzeConfig
 from jsonargparse import CLI
 
@@ -71,10 +71,11 @@ def analyze(config: AnalyzeConfig):
 
     test_case_with_failed_tools = []
     for entry in summary:
-        test_case_name = entry["test_case"]
+        test_case_name = entry["dataset_name"]
         if test_case_name.lower().strip() == "summary (average)":
             continue
-        if int(entry["Wrong Function Calls"]) > 0 or int(entry["Wrong Parameters"]) > 0:
+        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:
             test_case_with_failed_tools.append(entry)
     if len(test_case_with_failed_tools) == 0:
         header_table = Table(show_header=False, box=None)
@@ -85,7 +86,7 @@ def analyze(config: AnalyzeConfig):
         rich.print(header_panel)
 
     for test_case_entry in test_case_with_failed_tools:
-        test_case_name = test_case_entry["test_case"]
+        test_case_name = test_case_entry["dataset_name"]
 
         test_case_path = os.path.join(
             config.data_path, "messages", f"{test_case_name}.messages.analyze.json"
@@ -94,7 +95,8 @@ def analyze(config: AnalyzeConfig):
         with open(test_case_path, "r", encoding="utf-8") as f:
             temp = json.load(f)
             for entry in temp:
-                test_messages.append(ExtendedMessage(**entry))
+                msg = ExtendedMessage(**entry)
+                test_messages.append(msg)
 
         test_metrics_path = os.path.join(
             config.data_path, "messages", f"{test_case_name}.metrics.json"
@@ -105,11 +107,9 @@ def analyze(config: AnalyzeConfig):
         header_table.add_row(f"Test Case Name: {test_case_name}")
         header_table.add_row((f"Expected Tool Calls: {metrics.expected_tool_calls}"))
         header_table.add_row(f"Correct Tool Calls: {metrics.correct_tool_calls}")
-        irrelevant_tool_calls = test_case_entry["Wrong Function Calls"]
-        header_table.add_row(f"Irrelevant Tool Call: {irrelevant_tool_calls}")
-        tool_call_with_incorrect_parameters = test_case_entry["Wrong Parameters"]
+        header_table.add_row(f"Text Match: {metrics.text_match.value}")
         header_table.add_row(
-            f"Tool Call with incorrect parameters: {tool_call_with_incorrect_parameters}"
+            f"Journey Success: {metrics.is_success}"
         )
         header_panel = Panel(
             header_table, title="[bold green]📋 Analysis Summary[/bold green]"

wxo_agentic_evaluation/arg_configs.py

@@ -22,12 +22,19 @@ class LLMUserConfig:
     user_response_style: List[str] = field(default_factory=list)
 
 
+@dataclass
+class ProviderConfig:
+    model_id: str = field(default="meta-llama/llama-3-405b-instruct")
+    provider: str = field(default="watsonx")
+
+
 @dataclass
 class TestConfig:
     test_paths: List[str]
     output_dir: str
     auth_config: AuthConfig
     wxo_lite_version: str
+    provider_config: ProviderConfig = field(default_factory=ProviderConfig)
     llm_user_config: LLMUserConfig = field(default_factory=LLMUserConfig)
     enable_verbose_logging: bool = True
     enable_manual_user_input: bool = False
@@ -65,7 +72,7 @@ class ChatRecordingConfig:
         default_factory=KeywordsGenerationConfig
     )
     service_url: str = "http://localhost:4321"
-    tenant_name: str = "wxo-dev"
+    tenant_name: str = "local"
     token: str = None
 
 

wxo_agentic_evaluation/batch_annotate.py

@@ -5,7 +5,7 @@ import os
 from pathlib import Path
 from jsonargparse import CLI
 
-from wxo_agentic_evaluation.watsonx_provider import WatsonXProvider
+from wxo_agentic_evaluation.service_provider import get_provider
 from wxo_agentic_evaluation.prompt.template_render import BatchTestCaseGeneratorTemplateRenderer
 from wxo_agentic_evaluation.arg_configs import BatchAnnotateConfig
 from wxo_agentic_evaluation import __file__
@@ -71,7 +71,6 @@ def extract_inputs_from_snapshot(snapshot_path: Path) -> dict:
 def load_example(example_path: Path):
     with example_path.open("r", encoding="utf-8") as f:
         data = json.load(f)
-    data.pop("mine_fields", None)
     return data
 
 
@@ -98,13 +97,9 @@ def build_prompt_for_story(agent, tools, tool_inputs, example_case: dict, story:
 def generate_multiple_in_one(prompt, output_dir, starting_index, model_id="meta-llama/llama-3-405b-instruct", ):
     output_dir.mkdir(parents=True, exist_ok=True)
 
-    provider = WatsonXProvider(
+    provider = get_provider(
         model_id=model_id,
-        llm_decode_parameter={
-            "min_new_tokens": 50,
-            "decoding_method": "greedy",
-            "max_new_tokens": 3000
-        }
+        params={"min_new_tokens": 50, "decoding_method": "greedy", "max_new_tokens": 3000},
     )
 
     response = provider.query(prompt)
@@ -119,7 +114,6 @@ def generate_multiple_in_one(prompt, output_dir, starting_index, model_id="meta-
         assert isinstance(test_cases, list), "Expected list of test cases"
 
         for i, case in enumerate(test_cases, start=starting_index):
-            case["mine_fields"] = []  # ✅ Add the field here
             out_file = output_dir / f"synthetic_test_case_{i}.json"
             with out_file.open("w", encoding="utf-8") as f:
                 json.dump(case, f, indent=2)

wxo_agentic_evaluation/data_annotator.py

@@ -1,5 +1,6 @@
 from wxo_agentic_evaluation.type import Message, EvaluationData
-from wxo_agentic_evaluation.watsonx_provider import WatsonXProvider
+from wxo_agentic_evaluation.service_provider.watsonx_provider import Provider
+from wxo_agentic_evaluation.service_provider import get_provider
 from wxo_agentic_evaluation.prompt.template_render import (
     LlamaKeywordsGenerationTemplateRenderer,
 )
@@ -94,16 +95,16 @@ ERROR_KEYWORDS = [
 class KeywordsGenerationLLM:
     def __init__(
         self,
-        wai_client: WatsonXProvider,
+        provider: Provider,
         template: LlamaKeywordsGenerationTemplateRenderer,
     ):
-        self.wai_client = wai_client
+        self.provider = provider
         self.prompt_template = template
 
     def genereate_keywords(self, response) -> Message | None:
         prompt = self.prompt_template.render(response=response)
-        res = self.wai_client.query(prompt)
-        keywords = ast.literal_eval(res["generated_text"].strip())
+        res: str = self.provider.query(prompt)
+        keywords = ast.literal_eval(res.strip())
         return keywords
 
 
@@ -120,7 +121,6 @@ class DataAnnotator:
             agent="",
             story="",
             starting_sentence=messages[0].content if messages else "",
-            mine_fields=[],
             goals={},
             goal_details=[],
         )
@@ -145,29 +145,48 @@ class DataAnnotator:
 
     def _process_tool_call_order(self, wrong_tool_response_id: list[str]) -> list[str]:
         """Process and order tool calls, skipping failed ones"""
+        # gather all call ids that actually got a response
+        valid_call_ids = {
+            json.loads(m.content)["tool_call_id"]
+            for m in self.messages
+            if m.type == "tool_response"
+        }
+
         order = []
-        for message in self.messages:
-            if message.type == "tool_call":
-                content = json.loads(message.content)
-                # skip all the tool calls that fail
-                if (
-                    content.get("tool_call_id", "") in wrong_tool_response_id
-                    or content.get("id", "") in wrong_tool_response_id
-                ):
-                    continue
-
-                if "tool_call_id" in content:
-                    del content["tool_call_id"]
-                if "id" in content:
-                    del content["id"]
-
-                content = json.dumps(content, sort_keys=True)
-                # for a given tool call signature - function name + args only keep the most recent one
-                if content in order:
-                    idx = order.index(content)
-                    order = order[:idx] + order[idx + 1 :] + [content]
-                else:
-                    order.append(content)
+        for idx, message in enumerate(self.messages):
+            if message.type != "tool_call":
+                continue
+
+            content = json.loads(message.content)
+            call_id = content.get("tool_call_id") or content.get("id")
+
+            # skip any calls that errored
+            if call_id in wrong_tool_response_id:
+                continue
+
+            # skip calls that never produced a tool_response
+            if call_id not in valid_call_ids:
+                continue
+
+            # skip the "reflection" copy that the LLM emits right after a response
+            prev = self.messages[idx - 1] if idx > 0 else None
+            if (
+                prev is not None
+                and prev.type == "tool_response"
+                and json.loads(prev.content).get("tool_call_id") == call_id
+            ):
+                continue
+
+            # normalize ids so json dumps only reflects name-args
+            content.pop("tool_call_id", None)
+            content.pop("id", None)
+
+            signature = json.dumps(content, sort_keys=True)
+            # if we’ve seen that exact (name-args) before, drop the old one
+            if signature in order:
+                order.remove(signature)
+            order.append(signature)
+
         return order
 
     def _process_tool_calls(self) -> tuple[Dict, List, str]:
@@ -209,16 +228,12 @@ class DataAnnotator:
         # we assume single summary step at the end
         for message in self.messages[::-1]:
             if message.role == "assistant":
-                wai_client = WatsonXProvider(
+                provider = get_provider(
                     model_id=self.keywords_generation_config.model_id,
-                    llm_decode_parameter={
-                        "min_new_tokens": 0,
-                        "decoding_method": "greedy",
-                        "max_new_tokens": 256,
-                    },
+                    params={"min_new_tokens": 0, "decoding_method": "greedy", "max_new_tokens": 256},
                 )
                 kw_generator = KeywordsGenerationLLM(
-                    wai_client=wai_client,
+                    provider=provider,
                     template=LlamaKeywordsGenerationTemplateRenderer(
                         self.keywords_generation_config.prompt_config
                     ),
@@ -247,7 +262,6 @@ class DataAnnotator:
             "agent": self.initial_data.agent,
             "goals": goals,
             "goal_details": goal_details,
-            "mine_fields": [],
             "story": self.initial_data.story,
             "starting_sentence": self.initial_data.starting_sentence,
         }