feat(llm-observability): spans

frontend/src/lib/taxonomy.tsx

@@ -184,6 +184,11 @@ export const CORE_FILTER_DEFINITIONS_BY_GROUP = {
             description:
                 'A generative AI trace. Usually a trace tracks a single user interaction and contains one or more AI generation calls',
         },
+        $ai_span: {
+            label: 'AI Span',
+            description:
+                'A generative AI span. Usually a span tracks a unit of work for a trace of generative AI models (LLMs)',
+        },
         $ai_metric: {
             label: 'AI Metric',
             description: 'An evaluation metric for a trace of generative AI models (LLMs)',
@@ -1454,11 +1459,6 @@ export const CORE_FILTER_DEFINITIONS_BY_GROUP = {
             label: 'AI Output State (LLM)',
             description: 'Output state of the LLM agent',
         },
-        $ai_trace_name: {
-            label: 'AI Trace Name (LLM)',
-            description: 'The name given to this trace of LLM API calls',
-            examples: ['LangGraph'],
-        },
         $ai_provider: {
             label: 'AI Provider (LLM)',
             description: 'The provider of the AI model used to generate the output from the LLM API',
@@ -1490,6 +1490,21 @@ export const CORE_FILTER_DEFINITIONS_BY_GROUP = {
             description: 'The text provided by the user for feedback on the LLM trace',
             examples: ['"The response was helpful, but it did not use the provided context."'],
         },
+        $ai_parent_id: {
+            label: 'AI Parent ID (LLM)',
+            description: 'The parent span ID of a span or generation, used to group a trace into a tree view',
+            examples: ['bdf42359-9364-4db7-8958-c001f28c9255'],
+        },
+        $ai_span_id: {
+            label: 'AI Span ID (LLM)',
+            description: 'The unique identifier for a LLM trace, generation, or span.',
+            examples: ['bdf42359-9364-4db7-8958-c001f28c9255'],
+        },
+        $ai_span_name: {
+            label: 'AI Span Name (LLM)',
+            description: 'The name given to this LLM trace, generation, or span.',
+            examples: ['summarize_text'],
+        },
     },
     numerical_event_properties: {}, // Same as event properties, see assignment below
     person_properties: {}, // Currently person properties are the same as event properties, see assignment below

jest.config.ts

@@ -11,7 +11,7 @@ process.env.TZ = process.env.TZ || 'UTC'
 const esmModules = ['query-selector-shadow-dom', 'react-syntax-highlighter', '@react-hook', '@medv', 'monaco-editor']
 const eeFolderExists = fs.existsSync('ee/frontend/exports.ts')
 function rootDirectories() {
-    const rootDirectories = ['<rootDir>/frontend/src']
+    const rootDirectories = ['<rootDir>/frontend/src', '<rootDir>/products']
     if (eeFolderExists) {
         rootDirectories.push('<rootDir>/ee/frontend')
     }

posthog/demo/matrix/models.py

@@ -1,15 +1,14 @@
-from contextlib import contextmanager
 import datetime as dt
 from abc import ABC, abstractmethod
 from collections import defaultdict
+from collections.abc import Callable, Generator, Iterable
+from contextlib import contextmanager
 from copy import deepcopy
 from dataclasses import dataclass
 from enum import Enum, auto
 from itertools import chain
 from typing import TYPE_CHECKING, Any, Generic, Literal, Optional, TypeVar
-from collections.abc import Generator
-from collections.abc import Callable, Iterable
-from urllib.parse import urlparse, parse_qs
+from urllib.parse import parse_qs, urlparse
 from uuid import UUID, uuid4
 
 import tiktoken
@@ -228,7 +227,7 @@ def set_trace_output(output: Any):
                 {
                     "$ai_input_state": input_state,
                     "$ai_output_state": output_state,
-                    "$ai_trace_name": "SpikeChain",
+                    "$ai_span_name": "SpikeChain",
                     "$ai_trace_id": trace_id,
                 },
                 distinct_id=distinct_id,

posthog/hogql_queries/ai/test/test_traces_query_runner.py

@@ -95,7 +95,7 @@ def _create_ai_generation_event(
 def _create_ai_trace_event(
     *,
     trace_id: str,
-    trace_name: str,
+    trace_name: str | None,
     input_state: Any,
     output_state: Any,
     team: Team | None = None,
@@ -106,7 +106,7 @@ def _create_ai_trace_event(
 ):
     props = {
         "$ai_trace_id": trace_id,
-        "$ai_trace_name": trace_name,
+        "$ai_span_name": trace_name,
         "$ai_input_state": input_state,
         "$ai_output_state": output_state,
     }
@@ -123,6 +123,41 @@ def _create_ai_trace_event(
     )
 
 
+def _create_ai_span_event(
+    *,
+    trace_id: str,
+    input_state: Any,
+    output_state: Any,
+    span_id: str | None = None,
+    parent_id: str | None = None,
+    span_name: str | None = None,
+    team: Team | None = None,
+    distinct_id: str | None = None,
+    properties: dict[str, Any] | None = None,
+    timestamp: datetime | None = None,
+    event_uuid: str | UUID | None = None,
+):
+    props = {
+        "$ai_trace_id": trace_id,
+        "$ai_span_name": span_name,
+        "$ai_input_state": input_state,
+        "$ai_output_state": output_state,
+        "$ai_span_id": span_id or str(uuid.uuid4()),
+        "$ai_parent_id": parent_id or trace_id,
+    }
+    if properties:
+        props.update(properties)
+
+    _create_event(
+        event="$ai_span",
+        distinct_id=distinct_id,
+        properties=props,
+        team=team,
+        timestamp=timestamp,
+        event_uuid=str(event_uuid) if event_uuid else None,
+    )
+
+
 class TestTracesQueryRunner(ClickhouseTestMixin, BaseTest):
     def setUp(self):
         super().setUp()
@@ -556,11 +591,19 @@ def test_model_parameters(self):
 
     def test_full_trace(self):
         _create_person(distinct_ids=["person1"], team=self.team, properties={"foo": "bar"})
+        _create_ai_span_event(
+            trace_id="trace1",
+            span_name="runnable",
+            input_state={"messages": [{"role": "user", "content": "Foo"}]},
+            output_state={"messages": [{"role": "user", "content": "Foo"}, {"role": "assistant", "content": "Bar"}]},
+            team=self.team,
+            timestamp=datetime(2024, 12, 1, 0, 9),
+        )
         _create_ai_generation_event(
             distinct_id="person1",
             trace_id="trace1",
             team=self.team,
-            timestamp=datetime(2024, 12, 1, 0, 0),
+            timestamp=datetime(2024, 12, 1, 0, 9, 30),
         )
         _create_ai_generation_event(
             distinct_id="person1",
@@ -591,10 +634,18 @@ def test_full_trace(self):
             response.results[0].outputState,
             {"messages": [{"role": "user", "content": "Foo"}, {"role": "assistant", "content": "Bar"}]},
         )
-        self.assertEqual(len(response.results[0].events), 2)
+        self.assertEqual(len(response.results[0].events), 3)
+
+        self.assertEqual(response.results[0].events[0].event, "$ai_span")
         self.assertEqual(response.results[0].events[0].properties["$ai_trace_id"], "trace1")
+        self.assertEqual(response.results[0].events[0].properties["$ai_span_name"], "runnable")
+
+        self.assertEqual(response.results[0].events[1].event, "$ai_generation")
         self.assertEqual(response.results[0].events[1].properties["$ai_trace_id"], "trace1")
 
+        self.assertEqual(response.results[0].events[2].event, "$ai_generation")
+        self.assertEqual(response.results[0].events[2].properties["$ai_trace_id"], "trace1")
+
     @snapshot_clickhouse_queries
     def test_properties_filter_with_multiple_events_in_group(self):
         _create_person(distinct_ids=["person1"], team=self.team)
@@ -671,7 +722,7 @@ def test_trace_property_filter_for_event_group(self):
             team=self.team,
             query=TracesQuery(
                 properties=[
-                    EventPropertyFilter(key="$ai_trace_name", value="runnable", operator=PropertyOperator.EXACT)
+                    EventPropertyFilter(key="$ai_span_name", value="runnable", operator=PropertyOperator.EXACT)
                 ],
                 dateRange=DateRange(date_from="2024-12-01T00:00:00Z", date_to="2024-12-01T00:10:00Z"),
             ),
@@ -694,10 +745,65 @@ def test_trace_property_filter_for_event_group(self):
             team=self.team,
             query=TracesQuery(
                 properties=[
-                    EventPropertyFilter(key="$ai_trace_name", value="runnable", operator=PropertyOperator.EXACT),
+                    EventPropertyFilter(key="$ai_span_name", value="runnable", operator=PropertyOperator.EXACT),
                     EventPropertyFilter(key="foo", value="bar", operator=PropertyOperator.EXACT),
                 ],
                 dateRange=DateRange(date_from="2024-12-01T00:00:00Z", date_to="2024-12-01T00:10:00Z"),
             ),
         ).calculate()
         self.assertEqual(len(response.results), 0)
+
+    def test_trace_name_fallback(self):
+        """
+        $ai_trace_name is a deprecated property, but we still want to support it for backwards compatibility.
+        """
+        _create_person(distinct_ids=["person1"], team=self.team)
+        _create_ai_generation_event(
+            distinct_id="person1",
+            trace_id="trace1",
+            team=self.team,
+            timestamp=datetime(2024, 12, 1, 0, 0),
+        )
+        _create_ai_trace_event(
+            trace_id="trace1",
+            trace_name="runnable",
+            input_state={"messages": [{"role": "user", "content": "Foo"}]},
+            output_state={"messages": [{"role": "user", "content": "Foo"}, {"role": "assistant", "content": "Bar"}]},
+            team=self.team,
+            timestamp=datetime(2024, 12, 1, 0, 5),
+        )
+
+        response = TracesQueryRunner(
+            team=self.team,
+            query=TracesQuery(
+                traceId="trace1",
+                dateRange=DateRange(date_from="2024-12-01T00:00:00Z", date_to="2024-12-01T00:10:00Z"),
+            ),
+        ).calculate()
+        self.assertEqual(len(response.results), 1)
+        self.assertEqual(response.results[0].traceName, "runnable")
+
+        _create_ai_generation_event(
+            distinct_id="person1",
+            trace_id="trace2",
+            team=self.team,
+            timestamp=datetime(2024, 12, 1, 0, 0),
+        )
+        _create_ai_trace_event(
+            trace_id="trace2",
+            trace_name=None,
+            input_state={"messages": [{"role": "user", "content": "Foo"}]},
+            output_state={"messages": [{"role": "user", "content": "Foo"}, {"role": "assistant", "content": "Bar"}]},
+            team=self.team,
+            timestamp=datetime(2024, 12, 1, 0, 5),
+            properties={"$ai_trace_name": "bar"},
+        )
+        response = TracesQueryRunner(
+            team=self.team,
+            query=TracesQuery(
+                traceId="trace2",
+                dateRange=DateRange(date_from="2024-12-01T00:00:00Z", date_to="2024-12-01T00:10:00Z"),
+            ),
+        ).calculate()
+        self.assertEqual(len(response.results), 1)
+        self.assertEqual(response.results[0].traceName, "bar")

posthog/hogql_queries/ai/traces_query_runner.py

@@ -222,15 +222,15 @@ def _get_event_query(self) -> ast.SelectQuery:
                     arraySort(x -> x.3, groupArray(tuple(uuid, event, timestamp, properties))) as events,
                     {filter_conditions}
                 FROM events
-                WHERE event IN ('$ai_generation', '$ai_metric', '$ai_feedback') AND {common_conditions}
+                WHERE event IN ('$ai_span', '$ai_generation', '$ai_metric', '$ai_feedback') AND {common_conditions}
                 GROUP BY id
             ) AS generations
             LEFT JOIN (
                 SELECT
                     properties.$ai_trace_id as id,
                     argMin(properties.$ai_input_state, timestamp) as input_state,
                     argMin(properties.$ai_output_state, timestamp) as output_state,
-                    argMin(properties.$ai_trace_name, timestamp) as trace_name,
+                    argMin(ifNull(properties.$ai_span_name, properties.$ai_trace_name), timestamp) as trace_name,
                     {filter_conditions}
                 FROM events
                 WHERE event = '$ai_trace' AND {common_conditions}

posthog/taxonomy/taxonomy.py

@@ -196,6 +196,14 @@ class CoreFilterDefinition(TypedDict):
             "label": "AI Feedback (LLM)",
             "description": "User-provided feedback for a trace of a generative AI model (LLM).",
         },
+        "$ai_trace": {
+            "label": "AI Trace (LLM)",
+            "description": "A generative AI trace. Usually a trace tracks a single user interaction and contains one or more AI generation calls",
+        },
+        "$ai_span": {
+            "label": "AI Span (LLM)",
+            "description": "A generative AI span. Usually a span tracks a unit of work for a trace of generative AI models (LLMs)",
+        },
         "Application Opened": {
             "label": "Application Opened",
             "description": "When a user opens the mobile app either for the first time or from the foreground.",
@@ -1377,6 +1385,16 @@ class CoreFilterDefinition(TypedDict):
             "description": "The text provided by the user for feedback on the LLM trace",
             "examples": ['"The response was helpful, but it did not use the provided context."'],
         },
+        "$ai_parent_id": {
+            "label": "AI Parent ID (LLM)",
+            "description": "The parent span ID of a span or generation, used to group a trace into a tree view",
+            "examples": ["bdf42359-9364-4db7-8958-c001f28c9255"],
+        },
+        "$ai_span_id": {
+            "label": "AI Span ID (LLM)",
+            "description": "The unique identifier for a LLM trace, generation, or span.",
+            "examples": ["bdf42359-9364-4db7-8958-c001f28c9255"],
+        },
     },
     "numerical_event_properties": {},
     "person_properties": {},

products/llm_observability/frontend/ConversationDisplay/ConversationDisplay.tsx

@@ -17,8 +17,9 @@ export function ConversationDisplay({ eventProperties }: { eventProperties: Even
             </header>
             <ConversationMessagesDisplay
                 input={eventProperties.$ai_input}
-                output={eventProperties.$ai_output_choices}
+                output={eventProperties.$ai_output_choices ?? eventProperties.$ai_output ?? eventProperties.$ai_error}
                 httpStatus={eventProperties.$ai_http_status}
+                raisedError={eventProperties.$ai_is_error}
                 bordered
             />
         </>