customer-service-agent-traces

Multi-turn, tool-using customer-service agent traces for inference serving benchmarks. This is the customer-support slice of a larger agent-trace corpus: the three tau2 domains (airline, retail, telecom). Each line is one turn (a chat request) and carries the full conversation history, the tool catalog sent with it, and the recorded output length.

Source. Derived from the public Exgentic/agent-llm-traces-v2: the three tau2 domains (airline, retail, telecom) of agent runs recorded through a LiteLLM proxy, converted to aiperf mooncake_trace lines (messages and tools verbatim). Customers are tau2's synthetic users (@example.com). The upstream dataset declares no license (its v1 was CDLA-Permissive-2.0); the upstream terms apply.

Totals

sessions:  1,576
turns:     16,798
bytes:     1,126,364,854   (1.13 GB)
domains:   tau2_airline, tau2_retail, tau2_telecom

Schema

One JSON object per line (trace.tau2.jsonl):

field	type	description
session_id	str	Conversation id. Group lines by this to reconstruct a session.
messages	list[{role, content}]	Full history for this turn (system / user / assistant / tool)
tools	list / null	Tool (function) definitions sent with the request. Absent on some turns.
output_length	int	Recorded completion length (tokens) for this turn
delay	int	Client think-time before this turn (ms). Absent on the first turn of a session.

Sessions are multi-turn and the turns are stored in order.

Distributions

Counted directly from this file. Token counts use tiktoken o200k_base (the Artificial Analysis convention), cross-checked against the GLM-5.2 tokenizer with under 1% drift. ISL is input tokens per turn, meaning messages plus tools, which is what the server receives before chat-template overhead. OSL is output_length.

domain	sessions	turns	turns/sess (p50·p90·max)	OSL (p50·p90·p99)	ISL (p50·p90·p95·mean)	tool block ~tok/turn	delay (p50·p90·p99) ms
airline	370	2,689	3 · 17 · 51	167 · 647 · 1549	19.7k · 23.8k · 26.4k · 15.8k	9.5k	5937 · 30524 · 80104
retail	636	6,213	10 · 18 · 133	130 · 426 · 919	10.3k · 22.4k · 23.7k · 12.9k	7.1k	4979 · 17335 · 64581
telecom	570	7,896	14 · 29 · 50	167 · 414 · 976	13.8k · 24.1k · 24.8k · 15.6k	6.2k	5763 · 17852 · 81213
all	1,576	16,798		avg ≈ 240	13.0k · 23.8k · 24.6k · 14.6k		≈ 5.5k p50

ISL bucket share (turns, %):

domain	<4k	4–8k	8–16k	16–32k	≥32k
airline	5	22	17	54	1
retail	4	34	24	38	0
telecom	1	9	47	43	0

The tools block is most of the input (about 9.5k tokens per turn for airline) and it stays the same across a session, so it caches as a prefix. Counting messages alone gives a much lower ISL (all-turns mean about 7.7k, p95 about 17.0k), which is closer to a raw-conversation view.

The ISL panel is bimodal for airline: its large tool block pushes many turns to the 20k band, separate from the messages-driven peak lower down.

Intended use

Replay a realistic multi-turn, tool-heavy customer-support workload against a serving engine (SGLang, vLLM, TRT-LLM) to measure TTFT, ITL, throughput, and cache-hit. The constant per-session tool block plus the growing history also make this a clean, reproducible case for KV-cache and prefix-cache work. The ISL, OSL, and think-time distributions above are a first model of long-context agentic customer-service traffic for capacity planning.

Replay (aiperf)

Uses the mooncake_trace loader (messages and tools verbatim). Needs the loader patch from ai-dynamo/aiperf#1420 (branch thangquang09/agentx-mooncake-trace-loader).

aiperf profile \
  --scenario inferencex-agentx-mvp \
  --model zai-org/GLM-5.2-FP8 --url http://127.0.0.1:8888 --api-key "$API_KEY" \
  --endpoint-type chat \
  --input-file trace.tau2.jsonl --custom-dataset-type mooncake_trace \
  --no-fixed-schedule --concurrency 8 \
  --benchmark-duration 900 --request-count 16798 \   # stops at 15 min or one pool pass, whichever first
  --streaming --use-server-token-count \
  --tokenizer PhalaCloud/GLM-5.2-W4AFP8 --tokenizer-trust-remote-code \
  --num-dataset-entries 130000 --random-seed 42 \
  --no-gpu-telemetry --no-server-metrics \
  --output-artifact-dir tau2_c8 --unsafe-override

--use-server-token-count needs a server that returns token usage. Without it, ISL is counted client-side by the tokenizer and can miss the tools block, so prefer direct-to-server runs. Bench a low-to-mid concurrency ladder (e.g. 4 / 8 / 16) first.

Source

Split from Exgentic/agent-llm-traces-v2: kept only the tau2_airline, tau2_retail, and tau2_telecom sessions, byte for byte, in their original order.