Performance

Performance records are read-only QA data the system writes after an agent handles a call: pipeline latencies, token consumption, an automated quality assessment, and the full transcript including the agent’s tool calls. This is the same data the Dashboard shows under Performance Analysis. These records are separate from Calls. Calls are the customer-facing contact history (who called, what they wanted, the summary staff works with). Performance records are the technical and qualitative analysis of the same conversation — useful for monitoring agent quality, debugging slow responses, or feeding a QA dashboard.

Access

Permission: performance:read Issue an API key with this permission scoped to a single organization when a partner or internal tool needs read access to that organization’s QA data — that is narrower than granting Dashboard access, which would expose far more. With Dashboard JWT authentication the endpoints are additionally restricted to the super_admin, dev_admin, and dev_employee roles, mirroring the Dashboard’s own guard on the Performance Analysis page. client_admin and client_employee receive 403 FORBIDDEN.

List vs detail view

Field	List	Detail
Identifiers, phone numbers, duration, status	yes	yes
Latency percentiles (end-to-end, STT, LLM, TTS)	yes	yes
Token stats	yes	yes
Pipeline configuration (providers, model, mode)	yes	yes
`final_result` and redflags	yes	yes
`pre_call_variables`, `extracted_variables`	yes	yes
`transcript`	—	yes

The transcript is detail-only on purpose: it is by far the largest field, and a list request would multiply it by the page size. Internal debug logs are never returned by the API.

Data model

Field	Type	Description
`id`	uuid	Performance record identifier — the `statisticId` path parameter
`call_id`	string	LiveKit call identifier; use it to correlate with the Calls resource
`caller_phone`	string	Caller’s phone number
`agent_phone_number`	string	Number the caller dialled
`agent_id`	uuid	Agent that handled the call
`duration_seconds`	integer	Call duration in seconds
`disconnect_reason`	string	How the session ended
`session_status`	string	Final status of the agent session
`e2e_latency_p90`	number	End-to-end response latency, 90th percentile (seconds)
`e2e_latency_median`	number	End-to-end response latency, median
`e2e_latency_min`	number	End-to-end response latency, fastest turn
`transcription_latency_p90`	number	Speech-to-text latency, 90th percentile
`transcription_latency_median`	number	Speech-to-text latency, median
`transcription_latency_min`	number	Speech-to-text latency, minimum
`llm_latency_p90`	number	Language model latency, 90th percentile
`llm_latency_median`	number	Language model latency, median
`llm_latency_min`	number	Language model latency, minimum
`tts_latency_p90`	number	Text-to-speech latency, 90th percentile
`tts_latency_median`	number	Text-to-speech latency, median
`tts_latency_min`	number	Text-to-speech latency, minimum
`llm_total_tokens`	integer	Total tokens consumed by the call
`llm_prompt_tokens`	integer	Prompt tokens
`llm_completion_tokens`	integer	Completion tokens
`pre_call_variables`	object	Variables handed to the agent before the call started
`extracted_variables`	object	Variables the agent extracted during the call
`llm_provider`	string	Language model provider
`llm_model`	string	Language model used
`tts_provider`	string	Text-to-speech provider
`stt`	string	Speech-to-text provider
`pipeline_mode`	string	Voice pipeline mode
`created_at`	datetime	When the record was written (call start)
`final_result`	object	Overall evaluation, see below
`prompt_fingerprint`	string	SHA-256 of the prompt template — groups calls that ran on the same prompt version

Redflags. Seven fields carry an automated assessment of a specific failure pattern: assistant_misunderstood_multiple_times, repeated_same_sentence, abrupt_end_without_goodbye, connection_issue_repeated, assistant_policy_violation, technical_error, early_hangup_by_caller. Each is an object, not a boolean — check detected, do not test the field for truthiness:

Field	Type	Description
`detected`	boolean	Whether the pattern occurred
`reasoning`	string	Short explanation of the assessment
`score_impact`	number	How strongly this lowered the final score

final_result:

Field	Type	Description
`outcome`	boolean	Whether the call achieved its goal
`outcome_label`	string	`success`, `failed`, or `neutral`
`final_score`	number	Overall quality score
`summary_reasoning`	string	Explanation of the score

Detail-only field:

Field	Type	Description
`transcript`	array	Speech turns and tool calls, see Transcript format

Endpoints

List performance records

GET /v1/agents/{agentId}/performance

Permission: performance:read | Pagination: yes (ordered by created_at descending) Filters:

Parameter	Type	Description
`from`	datetime	Records created after this date (ISO 8601)
`to`	datetime	Records created before this date (ISO 8601)

# Performance records from the last 7 days
curl -H "X-API-Key: $TP_KEY" \
  "$TP_BASE/agents/{agentId}/performance?from=2026-03-15T00:00:00Z&limit=50"

Get performance record details

GET /v1/agents/{agentId}/performance/{statisticId}

Permission: performance:read Returns one record including the full transcript.

curl -H "X-API-Key: $TP_KEY" \
  "$TP_BASE/agents/{agentId}/performance/9f1c2d3e-4a5b-6c7d-8e9f-0a1b2c3d4e5f"

{
  "id": "9f1c2d3e-4a5b-6c7d-8e9f-0a1b2c3d4e5f",
  "call_id": "lk_call_abc123",
  "caller_phone": "+491701234567",
  "agent_id": "550e8400-e29b-41d4-a716-446655440000",
  "duration_seconds": 184,
  "session_status": "completed",
  "e2e_latency_p90": 1.42,
  "e2e_latency_median": 0.98,
  "llm_total_tokens": 4820,
  "llm_provider": "openai",
  "llm_model": "gpt-4o-mini",
  "created_at": "2026-03-19T10:30:00Z",
  "final_result": {
    "outcome": true,
    "outcome_label": "success",
    "final_score": 8,
    "summary_reasoning": "Termin erfolgreich vereinbart, keine Rueckfragen offen."
  },
  "technical_error": {
    "detected": false,
    "reasoning": "Keine technischen Fehler.",
    "score_impact": 0
  },
  "transcript": [
    { "role": "assistant", "text": "Guten Tag, Praxis Mustermann, was kann ich fuer Sie tun?", "time_sec": 2.1 },
    { "role": "user", "text": "Ich moechte einen Termin am zweiten April.", "time_sec": 9.4 }
  ]
}

Transcript format

The transcript is an array of objects — not a JSON string, and not the simple {role, text} array used by Calls. Do not reuse the parsing example from there. All entries live in one list sorted chronologically by time_sec (seconds since call start). Three kinds of entry are distinguished by role: Speech turn — role is user or assistant, the text is in text:

{ "role": "assistant", "text": "Guten Tag, was kann ich fuer Sie tun?", "timestamp": "2026-03-19T10:30:02Z", "time_sec": 2.1 }

Tool call — the agent invoking a tool. arguments is a JSON string, not an object:

{ "role": "tool_call_invocation", "tool_call_id": "call_7Hd2", "name": "check_availability", "arguments": "{\"date\":\"2026-04-02\"}", "time_sec": 10.8 }

Tool result — the tool’s answer. The payload is in content, not in text:

{ "role": "tool_call_result", "tool_call_id": "call_7Hd2", "successful": true, "content": "{\"slots\":[\"09:00\",\"14:30\"]}", "time_sec": 11.6 }

Match a result to its invocation via tool_call_id. Parse accordingly:

for (const entry of record.transcript) {
  switch (entry.role) {
    case 'user':
    case 'assistant':
      console.log(`${entry.role}: ${entry.text}`);
      break;
    case 'tool_call_invocation':
      console.log(`-> ${entry.name}(${JSON.parse(entry.arguments)})`);
      break;
    case 'tool_call_result':
      console.log(`<- ${entry.successful ? 'ok' : 'failed'}: ${entry.content}`);
      break;
  }
}

transcript is null for calls recorded before transcript capture was enabled.

Personal data

Transcripts are partially redacted before they are stored: phone numbers, e-mail addresses, IBANs, and digit runs of seven or more are removed. Names and addresses remain in clear text. Treat the data accordingly, and scope keys to the organization that actually owns the calls.

Common patterns

Watch response latency over time

const res = await talkpilot(
  `/agents/${agentId}/performance?from=2026-03-01T00:00:00Z&limit=100`
);
const slow = res.data.filter(r => r.e2e_latency_p90 > 2);
console.log(`${slow.length} of ${res.data.length} calls above 2s p90`);

Find calls with a detected failure pattern

const REDFLAGS = [
  'assistant_misunderstood_multiple_times', 'repeated_same_sentence',
  'abrupt_end_without_goodbye', 'connection_issue_repeated',
  'assistant_policy_violation', 'technical_error', 'early_hangup_by_caller',
];

const flagged = res.data.filter(r => REDFLAGS.some(f => r[f]?.detected));

Export a date range with transcripts

The list view omits transcripts, so fetch the detail per record:

const list = await talkpilot(
  `/agents/${agentId}/performance?from=2026-03-01T00:00:00Z&to=2026-03-31T23:59:59Z&limit=100`
);

for (const record of list.data) {
  const full = await talkpilot(`/agents/${agentId}/performance/${record.id}`);
  // full.transcript is now available
}

Agents — Parent resource
Calls — Customer-facing call history and summaries
Performance Analysis — Dashboard UI guide

API Resources

Endpoint Reference

Guides

Code Examples

Access

List vs detail view

Data model

Endpoints

List performance records

Get performance record details

Transcript format

Personal data

Common patterns

Watch response latency over time

Find calls with a detected failure pattern

Export a date range with transcripts

​Access

​List vs detail view

​Data model

​Endpoints

​List performance records

​Get performance record details

​Transcript format

​Personal data

​Common patterns

​Watch response latency over time

​Find calls with a detected failure pattern

​Export a date range with transcripts

​Related resources

Access

List vs detail view

Data model

Endpoints

List performance records

Get performance record details

Transcript format

Personal data

Common patterns

Watch response latency over time

Find calls with a detected failure pattern

Export a date range with transcripts

Related resources