1# Conversation Detail — Reconstruct Full Span Tree
2 
3Reconstruct the complete span tree for a single conversation to see exactly what happened: every LLM call, tool execution, and agent invocation with timing, tokens, and errors.
4 
5## Step 1 — Fetch All Spans for a Conversation
6 
7Use `operation_Id` (trace ID) to get all spans in a single request:
8 
9```kql
10dependencies
11| where operation_Id == "<operation_id>"
12| project timestamp, name, duration, resultCode, success,
13    spanId = id,
14    parentSpanId = operation_ParentId,
15    operation = tostring(customDimensions["gen_ai.operation.name"]),
16    model = tostring(customDimensions["gen_ai.request.model"]),
17    responseModel = tostring(customDimensions["gen_ai.response.model"]),
18    inputTokens = toint(customDimensions["gen_ai.usage.input_tokens"]),
19    outputTokens = toint(customDimensions["gen_ai.usage.output_tokens"]),
20    responseId = tostring(customDimensions["gen_ai.response.id"]),
21    finishReason = tostring(customDimensions["gen_ai.response.finish_reasons"]),
22    errorType = tostring(customDimensions["error.type"]),
23    toolName = tostring(customDimensions["gen_ai.tool.name"]),
24    toolCallId = tostring(customDimensions["gen_ai.tool.call.id"])
25| order by timestamp asc
26```
27 
28Also fetch the parent request:
29 
30```kql
31requests
32| where operation_Id == "<operation_id>"
33| project timestamp, name, duration, resultCode, success, id, operation_ParentId
34```
35 
36## Step 2 — Build Span Tree
37 
38Use `spanId` and `parentSpanId` to reconstruct the hierarchy:
39 
40```
41invoke_agent (root) ─── 4200ms
42├── chat (LLM call #1) ─── 1800ms, gpt-4o, 450→120 tokens
43│   └── [output: "Let me check the weather..."]
44├── execute_tool (get_weather) [tool: remote_functions.weather_api] ─── 200ms
45│   └── [result: "rainy, 57°F"]
46├── chat (LLM call #2) ─── 1500ms, gpt-4o, 620→85 tokens
47│   └── [output: "The weather in Paris is rainy, 57°F"]
48└── [total: 450+620=1070 input, 120+85=205 output tokens]
49```
50 
51Present as an indented tree with:
52- **Operation type** and name
53- **Duration** (highlight if > P95 for that operation type)
54- **Model** and token counts (for chat operations)
55- **Error type** and result code (if failed, highlight in red)
56- **Finish reason** (stop, length, content_filter, tool_calls)
57 
58## Step 3 — Extract Conversation Content from invoke_agent Spans
59 
60The full input/output content lives on `invoke_agent` dependency spans in `gen_ai.input.messages` and `gen_ai.output.messages`. These JSON arrays contain the complete conversation (system prompt, user query, assistant response):
61 
62```kql
63dependencies
64| where operation_Id == "<operation_id>"
65| where customDimensions["gen_ai.operation.name"] == "invoke_agent"
66| project timestamp,
67    inputMessages = tostring(customDimensions["gen_ai.input.messages"]),
68    outputMessages = tostring(customDimensions["gen_ai.output.messages"])
69| order by timestamp asc
70```
71 
72Message structure: `[{"role": "user", "parts": [{"type": "text", "content": "..."}]}]`
73 
74Also check the `traces` table for additional GenAI log events:
75 
76```kql
77traces
78| where operation_Id == "<operation_id>"
79| where message contains "gen_ai"
80| project timestamp, message, customDimensions
81| order by timestamp asc
82```
83 
84## Step 4 — Check for Exceptions
85 
86```kql
87exceptions
88| where operation_Id == "<operation_id>"
89| project timestamp, type, message, outerMessage,
90    details = parse_json(details)
91| order by timestamp asc
92```
93 
94Present exceptions inline in the span tree at their position in the timeline.
95 
96## Step 5 — Fetch Evaluation Results
97 
98See [Eval Correlation](eval-correlation.md) for the full workflow to look up evaluation scores by response ID or conversation ID. Use `gen_ai.response.id` values from Step 1 spans to correlate.
99