chore(llmobs): send record metadata with experiment spans & make experiment spans i/o searchable

ddtrace/llmobs/_constants.py

@@ -104,13 +104,17 @@
 
 PROXY_REQUEST = "llmobs.proxy_request"
 
+# experiment span baggage keys to be propagated across boundaries
 EXPERIMENT_ID_KEY = "_ml_obs.experiment_id"
 EXPERIMENT_RUN_ID_KEY = "_ml_obs.experiment_run_id"
 EXPERIMENT_RUN_ITERATION_KEY = "_ml_obs.experiment_run_iteration"
 EXPERIMENT_PROJECT_NAME_KEY = "_ml_obs.experiment_project_name"
 EXPERIMENT_PROJECT_ID_KEY = "_ml_obs.experiment_project_id"
 EXPERIMENT_DATASET_NAME_KEY = "_ml_obs.experiment_dataset_name"
 EXPERIMENT_NAME_KEY = "_ml_obs.experiment_name"
+
+# experiment context keys
+EXPERIMENT_RECORD_METADATA = "_ml_obs.meta.metadata"
 EXPERIMENT_EXPECTED_OUTPUT = "_ml_obs.meta.input.expected_output"
 EXPERIMENTS_INPUT = "_ml_obs.meta.input"
 EXPERIMENTS_OUTPUT = "_ml_obs.meta.output"

ddtrace/llmobs/_experiment.py

@@ -24,6 +24,7 @@
 from ddtrace.internal.logger import get_logger
 from ddtrace.llmobs._constants import DD_SITES_NEEDING_APP_SUBDOMAIN
 from ddtrace.llmobs._constants import EXPERIMENT_EXPECTED_OUTPUT
+from ddtrace.llmobs._constants import EXPERIMENT_RECORD_METADATA
 from ddtrace.llmobs._utils import convert_tags_dict_to_list
 from ddtrace.llmobs._utils import safe_json
 from ddtrace.version import __version__
@@ -488,7 +489,11 @@ def _process_record(self, idx_record: Tuple[int, DatasetRecord], run: _Experimen
             except Exception:
                 span.set_exc_info(*sys.exc_info())
             self._llmobs_instance.annotate(span, input_data=input_data, output_data=output_data, tags=tags)
-            span._set_ctx_item(EXPERIMENT_EXPECTED_OUTPUT, safe_json(record["expected_output"]))
+
+            span._set_ctx_item(EXPERIMENT_EXPECTED_OUTPUT, record["expected_output"])
+            if "metadata" in record:
+                span._set_ctx_item(EXPERIMENT_RECORD_METADATA, record["metadata"])
+
             return {
                 "idx": idx,
                 "span_id": span_id,

ddtrace/llmobs/_llmobs.py

@@ -1738,9 +1738,9 @@ def _tag_freeform_io(cls, span, input_value=None, output_value=None):
         arbitrary structured or non structured IO values in its spans
         """
         if input_value is not None:
-            span._set_ctx_item(EXPERIMENTS_INPUT, safe_json(input_value))
+            span._set_ctx_item(EXPERIMENTS_INPUT, input_value)
         if output_value is not None:
-            span._set_ctx_item(EXPERIMENTS_OUTPUT, safe_json(output_value))
+            span._set_ctx_item(EXPERIMENTS_OUTPUT, output_value)
 
     @staticmethod
     def _set_dict_attribute(span: Span, key, value: Dict[str, Any]) -> None:

releasenotes/notes/llmobs-dne-experiments-fields-json-metadata-on-spans-53d679fd8e6ab202.yaml

@@ -0,0 +1,7 @@
+---
+upgrade:
+  - |
+    LLM Observability: Experiments spans now contain metadata from the dataset record.
+  - |
+    LLM Observability: Experiments spans' input, output, expected_output fields are now emitted as is so that if data in
+    any of the columns are objects, they can be searchable in Datadog. 

tests/llmobs/llmobs_cassettes/datadog/datadog_api_unstable_llm-obs_v1_datasets_0969efc9-f104-45cc-b955-25b329e91293_batch_update_post_8be41af9.yaml

@@ -0,0 +1,47 @@
+interactions:
+- request:
+    body: '{"data": {"type": "datasets", "id": "0969efc9-f104-45cc-b955-25b329e91293",
+      "attributes": {"insert_records": [{"input": {"prompt": "What is the capital
+      of France?"}, "expected_output": {"answer": "Paris"}, "metadata": {"difficulty":
+      "easy"}}], "update_records": [], "delete_records": []}}}'
+    headers:
+      Accept:
+      - '*/*'
+      ? !!python/object/apply:multidict._multidict.istr
+      - Accept-Encoding
+      : - identity
+      Connection:
+      - keep-alive
+      Content-Length:
+      - '289'
+      ? !!python/object/apply:multidict._multidict.istr
+      - Content-Type
+      : - application/json
+      User-Agent:
+      - python-requests/2.32.3
+    method: POST
+    uri: https://api.datadoghq.com/api/unstable/llm-obs/v1/datasets/0969efc9-f104-45cc-b955-25b329e91293/batch_update
+  response:
+    body:
+      string: '{"data":[]}'
+    headers:
+      content-length:
+      - '11'
+      content-security-policy:
+      - frame-ancestors 'self'; report-uri https://logs.browser-intake-datadoghq.com/api/v2/logs?dd-api-key=pube4f163c23bbf91c16b8f57f56af9fc58&dd-evp-origin=content-security-policy&ddsource=csp-report&ddtags=site%3Adatadoghq.com
+      content-type:
+      - application/vnd.api+json
+      date:
+      - Tue, 09 Dec 2025 03:28:32 GMT
+      strict-transport-security:
+      - max-age=31536000; includeSubDomains; preload
+      vary:
+      - Accept-Encoding
+      x-content-type-options:
+      - nosniff
+      x-frame-options:
+      - SAMEORIGIN
+    status:
+      code: 200
+      message: OK
+version: 1