aiqa-client 0.5.2__py3-none-any.whl → 0.6.1__py3-none-any.whl

aiqa/__init__.py

@@ -26,8 +26,8 @@ Example:
     result = my_function()
 """
 
-from .tracing import (
-    WithTracing,
+from .tracing import WithTracing
+from .span_helpers import (
     flush_tracing,
     set_span_attribute,
     set_span_name,
@@ -39,7 +39,10 @@ from .tracing import (
     extract_trace_context,
     set_conversation_id,
     set_component_tag,
+    set_token_usage,
+    set_provider_and_model,
     get_span,
+    submit_feedback,
 )
 from .client import get_aiqa_client
 from .experiment_runner import ExperimentRunner
@@ -60,7 +63,10 @@ __all__ = [
     "extract_trace_context",
     "set_conversation_id",
     "set_component_tag",
+    "set_token_usage",
+    "set_provider_and_model",
     "get_span",
+    "submit_feedback",
     "VERSION",
 ]
 

aiqa/client.py

@@ -5,8 +5,10 @@ from functools import lru_cache
 from typing import Optional, TYPE_CHECKING, Any, Dict
 from opentelemetry import trace
 from opentelemetry.sdk.trace import TracerProvider
-from opentelemetry.sdk.trace.export import BatchSpanProcessor
+from opentelemetry.sdk.trace.export import BatchSpanProcessor, SpanExporter, SpanExportResult, SpanExporter as SpanExporterBase
 from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
+from opentelemetry.sdk.trace import ReadableSpan
+from opentelemetry.trace import SpanContext
 import requests
 
 from .constants import AIQA_TRACER_NAME, LOG_TAG
@@ -254,11 +256,24 @@ def _attach_aiqa_processor(provider: TracerProvider) -> None:
         else:
             endpoint = f"{base_url}/v1/traces"
         
-        # Create OTLP exporter with authentication headers only
+        # Get timeout from environment variable (in seconds)
+        # Supports OTEL_EXPORTER_OTLP_TIMEOUT (standard) or AIQA_EXPORT_TIMEOUT (custom)
+        # Default is 30 seconds (more generous than OTLP default of 10s)
+        timeout = 30.0
+        otlp_timeout = os.getenv("OTEL_EXPORTER_OTLP_TIMEOUT")
+        
+        if otlp_timeout:
+            try:
+                timeout = float(otlp_timeout)
+            except ValueError:
+                logger.warning(f"Invalid OTEL_EXPORTER_OTLP_TIMEOUT value '{otlp_timeout}', using default 30.0")
+        
+        # Create OTLP exporter with authentication headers and timeout
         # The exporter will set Content-Type and other headers automatically
         exporter = OTLPSpanExporter(
             endpoint=endpoint,
             headers=auth_headers if auth_headers else None,
+            timeout=timeout,
         )
         
         provider.add_span_processor(BatchSpanProcessor(exporter))

aiqa/constants.py

@@ -3,6 +3,6 @@ Constants used across the AIQA client package.
 """
 
 AIQA_TRACER_NAME = "aiqa-tracer"
-VERSION = "0.5.2" # automatically updated by set-version-json.sh
+VERSION = "0.6.1" # automatically updated by set-version-json.sh
 
 LOG_TAG = "AIQA" # Used in all logging output to identify AIQA messages

aiqa/experiment_runner.py

@@ -4,10 +4,52 @@ ExperimentRunner - runs experiments on datasets and scores results
 
 import os
 import time
+import asyncio
 from .constants import LOG_TAG
 from .http_utils import build_headers, get_server_url, get_api_key, format_http_error
 from typing import Any, Dict, List, Optional, Callable, Awaitable, Union
+from .tracing import WithTracing
+from .span_helpers import set_span_attribute, flush_tracing
+from .llm_as_judge import score_llm_metric_local, get_model_from_server, call_llm_fallback
 import requests
+from .types import MetricResult, ScoreThisInputOutputMetricType, Example, Result, Metric, CallLLMType
+
+# Type aliases for engine/scoring functions to improve code completion and clarity
+from typing import TypedDict
+
+# Function that processes input and parameters to produce an output (sync or async)
+CallMyCodeType = Callable[[Any, Dict[str, Any]], Union[Any, Awaitable[Any]]]
+
+# Function that scores a given output, using input, example, and parameters (usually async)
+# Returns a dictionary with score/message/etc.
+ScoreThisOutputType = Callable[[Any, Any, Dict[str, Any], Dict[str, Any]], Awaitable[Dict[str, Any]]]
+
+
+
+def _filter_input_for_run(input_data: Any) -> Dict[str, Any]:
+    """Tracing:Filter input - drop most, keep just ids"""
+    if not isinstance(input_data, dict):
+        return {}
+    self_obj = input_data.get("self")
+    if not self_obj:
+        return {}
+    return {
+        "dataset": getattr(self_obj, "dataset_id", None),
+        "experiment": getattr(self_obj, "experiment_id", None),
+    }
+
+
+def _filter_input_for_run_example(
+    self: "ExperimentRunner", 
+    example: Dict[str, Any],
+    call_my_code: Any = None,
+    score_this_output: Any = None,
+) -> Dict[str, Any]:
+    """Filter input for run_example method to extract dataset, experiment, and example IDs."""
+    result = _filter_input_for_run({"self": self})
+    if isinstance(example, dict):
+        result["example"] = example.get("id")
+    return result
 
 
 class ExperimentRunner:
@@ -24,6 +66,7 @@ class ExperimentRunner:
         server_url: Optional[str] = None,
         api_key: Optional[str] = None,
         organisation_id: Optional[str] = None,
+        llm_call_fn: Optional[CallLLMType] = None,
     ):
         """
         Initialize the ExperimentRunner.
@@ -33,7 +76,11 @@ class ExperimentRunner:
             experiment_id: Usually unset, and a fresh experiment is created with a random ID
             server_url: URL of the AIQA server (defaults to AIQA_SERVER_URL env var)
             api_key: API key for authentication (defaults to AIQA_API_KEY env var)
-            organisation_id: Organisation ID for the experiment
+            organisation_id: Optional organisation ID for the experiment. If not provided, will be
+                            derived from the dataset when needed.
+            llm_call_fn: Optional async function that takes (system_prompt, user_message) and returns
+                        raw content string (typically JSON). If not provided, will check for OPENAI_API_KEY
+                        or ANTHROPIC_API_KEY environment variables.
         """
         self.dataset_id = dataset_id
         self.experiment_id = experiment_id
@@ -42,6 +89,8 @@ class ExperimentRunner:
         self.organisation = organisation_id
         self.experiment: Optional[Dict[str, Any]] = None
         self.scores: List[Dict[str, Any]] = []
+        self.llm_call_fn = llm_call_fn
+        self._dataset_cache: Optional[Dict[str, Any]] = None
 
     def _get_headers(self) -> Dict[str, str]:
         """Build HTTP headers for API requests."""
@@ -54,6 +103,9 @@ class ExperimentRunner:
         Returns:
             The dataset object with metrics and other information
         """
+        if self._dataset_cache is not None:
+            return self._dataset_cache
+
         response = requests.get(
             f"{self.server_url}/dataset/{self.dataset_id}",
             headers=self._get_headers(),
@@ -62,7 +114,14 @@ class ExperimentRunner:
         if not response.ok:
             raise Exception(format_http_error(response, "fetch dataset"))
 
-        return response.json()
+        dataset = response.json()
+        self._dataset_cache = dataset
+        
+        # If organisation_id wasn't set, derive it from the dataset
+        if not self.organisation and dataset.get("organisation"):
+            self.organisation = dataset.get("organisation")
+        
+        return dataset
 
     def get_example_inputs(self, limit: int = 10000) -> List[Dict[str, Any]]:
         """
@@ -108,8 +167,13 @@ class ExperimentRunner:
         Returns:
             The created experiment object
         """
+        # Ensure we have the organisation ID - try to get it from the dataset if not set
+        if not self.organisation:
+            dataset = self.get_dataset()
+            self.organisation = dataset.get("organisation")
+        
         if not self.organisation or not self.dataset_id:
-            raise Exception("Organisation and dataset ID are required to create an experiment")
+            raise Exception("Organisation and dataset ID are required to create an experiment. Organisation can be derived from the dataset or set via organisation_id parameter.")
 
         if not experiment_setup:
             experiment_setup = {}
@@ -138,19 +202,19 @@ class ExperimentRunner:
         self.experiment = experiment
         return experiment
 
-    def score_and_store(
+    async def score_and_store(
         self,
-        example: Dict[str, Any],
-        result: Any,
-        scores: Optional[Dict[str, Any]] = None,
-    ) -> Dict[str, Any]:
+        example: Example,
+        output: Any,
+        result: Result,
+    ) -> Result:
         """
         Ask the server to score an example result. Stores the score for later summary calculation.
 
         Args:
             example: The example object
-            result: The output from running the engine on the example
-            scores: Optional pre-computed scores
+            output: The output from running the engine on the example
+            result: The result object for locally calculated scores
 
         Returns:
             The score result from the server
@@ -158,22 +222,34 @@ class ExperimentRunner:
         # Do we have an experiment ID? If not, we need to create the experiment first
         if not self.experiment_id:
             self.create_experiment()
-
-        if scores is None:
-            scores = {}
-
-        print(f"Scoring and storing example: {example['id']}")
+        example_id = example.get("id")
+        if not example_id:
+            raise ValueError("Example must have an 'id' field")
+        if result is None:
+            example_id = example.get("id")
+            if not example_id:
+                raise ValueError("Example must have an 'id' field")
+            result = Result(exampleId=example_id, scores={}, messages={}, errors={})
+        scores = result.get("scores") or {}
+
+        
+        
+        print(f"Scoring and storing example: {example_id}")
         print(f"Scores: {scores}")
 
-        response = requests.post(
-            f"{self.server_url}/experiment/{self.experiment_id}/example/{example['id']}/scoreAndStore",
-            json={
-                "output": result,
-                "traceId": example.get("traceId"),
-                "scores": scores,
-            },
-            headers=self._get_headers(),
-        )
+        # Run synchronous requests.post in a thread pool to avoid blocking
+        def _do_request():
+            return requests.post(
+                f"{self.server_url}/experiment/{self.experiment_id}/example/{example_id}/scoreAndStore",
+                json={
+                    "output": result,
+                    "traceId": example.get("traceId"),
+                    "scores": scores,
+                },
+                headers=self._get_headers(),
+            )
+
+        response = await asyncio.to_thread(_do_request)
 
         if not response.ok:
             raise Exception(format_http_error(response, "score and store"))
@@ -182,12 +258,11 @@ class ExperimentRunner:
         print(f"scoreAndStore response: {json_result}")
         return json_result
 
+    @WithTracing(filter_input=_filter_input_for_run)
     async def run(
         self,
-        engine: Callable[[Any], Union[Any, Awaitable[Any]]],
-        scorer: Optional[
-            Callable[[Any, Dict[str, Any]], Awaitable[Dict[str, Any]]]
-        ] = None,
+        call_my_code: CallMyCodeType,
+        scorer_for_metric_id: Optional[Dict[str, ScoreThisInputOutputMetricType]] = None,
     ) -> None:
         """
         Run an engine function on all examples and score the results.
@@ -199,42 +274,43 @@ class ExperimentRunner:
         examples = self.get_example_inputs()
 
         # Wrap engine to match run_example signature (input, parameters)
-        def wrapped_engine(input_data, parameters):
-            return engine(input_data)
-
-        # Wrap scorer to match run_example signature (output, example, parameters)
-        async def wrapped_scorer(output, example, parameters):
-            if scorer:
-                return await scorer(output, example)
-            return {}
+        async def wrapped_engine(input_data, parameters):
+            result = call_my_code(input_data, parameters)
+            # Handle async functions
+            if hasattr(result, "__await__"):
+                result = await result
+            return result
 
         for example in examples:
-            scores = await self.run_example(example, wrapped_engine, wrapped_scorer)
-            if scores:
-                self.scores.append(
-                    {
-                        "example": example,
-                        "result": scores,
-                        "scores": scores,
-                    }
-                )
-
+            try:
+                scores = await self.run_example(example, wrapped_engine, scorer_for_metric_id)
+                if scores:
+                    self.scores.append(
+                        {
+                            "example": example,
+                            "result": scores,
+                            "scores": scores,
+                        }
+                    )
+            except Exception as e:
+                print(f"Error processing example {example.get('id', 'unknown')}: {e}")
+                # Continue with next example instead of failing entire run
+
+    @WithTracing(filter_input=_filter_input_for_run_example)
     async def run_example(
         self,
-        example: Dict[str, Any],
-        call_my_code: Callable[[Any, Dict[str, Any]], Union[Any, Awaitable[Any]]],
-        score_this_output: Optional[
-            Callable[[Any, Dict[str, Any], Dict[str, Any]], Awaitable[Dict[str, Any]]]
-        ] = None,
-    ) -> List[Dict[str, Any]]:
+        example: Example,
+        call_my_code: CallMyCodeType,
+        scorer_for_metric_id: Optional[Dict[str, ScoreThisInputOutputMetricType]] = None,
+    ) -> List[Result]:
         """
         Run the engine on an example with the given parameters (looping over comparison parameters),
         and score the result. Also calls scoreAndStore to store the result in the server.
 
         Args:
-            example: The example to run
+            example: The example to run. See Example.ts type
             call_my_code: Function that takes input and parameters, returns output (can be async)
-            score_this_output: Optional function that scores the output given the example and parameters
+            scorer_for_metric_id: Optional dictionary of metric IDs to functions that score the output given the example and parameters
 
         Returns:
             One set of scores for each comparison parameter set. If no comparison parameters,
@@ -261,41 +337,87 @@ class ExperimentRunner:
             )
             # Run engine anyway -- this could make sense if it's all about the parameters
 
+        # Set example.id on the root span (created by @WithTracing decorator)
+        # This ensures the root span from the trace has example=Example.id set
+        example_id = example.get("id")
+        if not example_id:
+            raise ValueError("Example must have an 'id' field")
+        set_span_attribute("example", example_id)
+        
         all_scores: List[Dict[str, Any]] = []
+        dataset_metrics = self.get_dataset().get("metrics", [])
+        specific_metrics = example.get("metrics", [])
+        metrics = [*dataset_metrics, *specific_metrics]
         # This loop should not be parallelized - it should run sequentially, one after the other
         # to avoid creating interference between the runs.
         for parameters in parameters_loop:
             parameters_here = {**parameters_fixed, **parameters}
             print(f"Running with parameters: {parameters_here}")
 
+            # Save original env var values for cleanup
+            original_env_vars: Dict[str, Optional[str]] = {}
             # Set env vars from parameters_here
             for key, value in parameters_here.items():
                 if value:
+                    original_env_vars[key] = os.environ.get(key)
                     os.environ[key] = str(value)
 
-            start = time.time() * 1000  # milliseconds
-            output = call_my_code(input_data, parameters_here)
-            # Handle async functions
-            if hasattr(output, "__await__"):
-                import asyncio
-
-                output = await output
-            end = time.time() * 1000  # milliseconds
-            duration = int(end - start)
-
-            print(f"Output: {output}")
-
-            scores: Dict[str, Any] = {}
-            if score_this_output:
-                scores = await score_this_output(output, example, parameters_here)
-
-            scores["duration"] = duration
-
-            # TODO: this call as async and wait for all to complete before returning
-            print(f"Call scoreAndStore ... for example: {example['id']} with scores: {scores}")
-            result = self.score_and_store(example, output, scores)
-            print(f"scoreAndStore returned: {result}")
-            all_scores.append(result)
+            try:
+                start = time.time() * 1000  # milliseconds
+                output = call_my_code(input_data, parameters_here)
+                # Handle async functions
+                if hasattr(output, "__await__"):
+                    output = await output
+                end = time.time() * 1000  # milliseconds
+                duration = int(end - start)
+
+                print(f"Output: {output}")
+                # Score it
+                result = Result(exampleId=example_id, scores={}, messages={}, errors={})
+                for metric in metrics:
+                    metric_id = metric.get("id")
+                    if not metric_id:
+                        print(f"Warning: Metric missing 'id' field, skipping: {metric}")
+                        continue
+                    scorer = scorer_for_metric_id.get(metric_id) if scorer_for_metric_id else None
+                    if scorer:
+                        metric_result = await scorer(input_data, output, metric)
+                    elif metric.get("type") == "llm":
+                        metric_result = await self._score_llm_metric(input_data, output, example, metric)
+                    else:
+                        metric_type = metric.get("type", "unknown")
+                        print(f"Skipping metric: {metric_id} {metric_type} - no scorer")
+                        continue
+                    
+                    # Handle None metric_result (e.g., if scoring failed)
+                    if not metric_result:
+                        print(f"Warning: Metric {metric_id} returned None result, skipping")
+                        result["errors"][metric_id] = "Scoring function returned None"
+                        continue
+                    
+                    result["scores"][metric_id] = metric_result.get("score")
+                    result["messages"][metric_id] = metric_result.get("message")
+                    result["errors"][metric_id] = metric_result.get("error")
+                # Always add duration to scores as a system metric
+                result["scores"]["duration"] = duration
+
+                # Flush spans before scoreAndStore to ensure they're indexed in ES
+                # This prevents race condition where scoreAndStore looks up spans before they're indexed
+                await flush_tracing()
+
+                print(f"Call scoreAndStore ... for example: {example_id} with scores: {result['scores']}")
+                result = await self.score_and_store(example, output, result)
+                print(f"scoreAndStore returned: {result}")
+                all_scores.append(result)
+            finally:
+                # Restore original env var values
+                for key, original_value in original_env_vars.items():
+                    if original_value is None:
+                        # Variable didn't exist before, remove it
+                        os.environ.pop(key, None)
+                    else:
+                        # Restore original value
+                        os.environ[key] = original_value
 
         return all_scores
 
@@ -306,6 +428,9 @@ class ExperimentRunner:
         Returns:
             Dictionary of metric names to summary statistics
         """
+        if not self.experiment_id:
+            raise ValueError("No experiment ID available. Create an experiment first.")
+        
         response = requests.get(
             f"{self.server_url}/experiment/{self.experiment_id}",
             headers=self._get_headers(),
@@ -317,3 +442,49 @@ class ExperimentRunner:
         experiment2 = response.json()
         return experiment2.get("summary_results", {})
 
+    async def _score_llm_metric(
+        self,
+        input_data: Any,
+        output: Any,
+        example: Example,
+        metric: Metric,
+    ) -> MetricResult:
+        """
+        Score an LLM metric by fetching model API key from server if needed.
+        
+        Args:
+            input_data: The input data to score
+            output: The output to score
+            example: The example object
+            metric: The metric definition
+            
+        Returns:
+            MetricResult object with score:[0,1], message (optional), and error (optional)
+        """
+        # If model is specified, try to fetch API key from server
+        model_id = metric.get("model")
+        api_key = None
+        provider = metric.get("provider")
+        
+        if model_id:
+            model_data = await get_model_from_server(
+                model_id, self.server_url, self._get_headers()
+            )
+            if model_data:
+                api_key = model_data.get("api_key")
+                # If provider not set in metric, try to get it from model
+                if not provider and model_data.get("provider"):
+                    provider = model_data.get("provider")
+        
+        # Create a custom llm_call_fn if we have an API key from the model
+        llm_call_fn = self.llm_call_fn
+        if api_key and not llm_call_fn:
+            async def _model_llm_call(system_prompt: str, user_message: str) -> str:
+                return await call_llm_fallback(system_prompt, user_message, api_key, provider)
+            llm_call_fn = _model_llm_call
+        
+        return await score_llm_metric_local(
+            input_data, output, example, metric, llm_call_fn
+        )
+
+