> ## Documentation Index
> Fetch the complete documentation index at: https://docs.agntix.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Post-call analysis

> Configure intent, resolution, sentiment, and structured data extraction so each call returns rich, structured fields — not raw transcript.

<Note>
  Every endpoint on this page requires either a Clerk-issued JWT (`Authorization: Bearer <token>`) or
  an organization API key (`x-api-key: pk_…`). Anonymous calls return `401 Unauthorized`. See
  [Authentication](/authentication) for the full setup.
</Note>

Once a call ends, Agntix runs an analysis pass over the transcript. Configure what to extract on the agent's **Analytics tab** — open any agent from **Agents** in the left nav, then switch to the **Analytics** tab on the right-hand panel.

Four independent toggles are available. Enable only the ones you need.

| Toggle              | What it returns                          | Where it appears                |
| ------------------- | ---------------------------------------- | ------------------------------- |
| Intent Recognition  | Best-fit intent against your custom list | `analysis.intent.*`             |
| Resolution Analysis | Whether the call achieved its objective  | `analysis.resolution.*`         |
| Sentiment           | Score and label per call                 | `analysis.sentiment.*`          |
| Data Extraction     | Structured fields you defined            | `analysis.extractedData.<name>` |

All four surface in the `analysis` block of the [`session.ended` webhook](/webhooks/events) and in [`GET /api/v1/chat/sessions/{id}`](/api-reference/chat/sessions/get).

## Intent Recognition

Classifies the **purpose** of the conversation against a list of intents you define under **Custom Intent** (comma-separated).

Example for a delivery-confirmation use case:

```text theme={null}
Slot-Confirmed, Slot-Rescheduled, Customer-Unavailable, Wrong-Number, Refund-Request
```

The analysis returns:

<ResponseField name="analysis.intent.conversation.intent" type="string">
  Best-fit intent across the whole call.
</ResponseField>

<ResponseField name="analysis.intent.conversation.confidence" type="number">
  Confidence score, `0` – `1`.
</ResponseField>

<ResponseField name="analysis.intent.conversation.explanation" type="string">
  Short LLM-generated rationale for the classification.
</ResponseField>

<ResponseField name="analysis.intent.message.intent" type="string">
  Best-fit on the trigger message (less useful for voice).
</ResponseField>

```json theme={null}
"intent": {
  "conversation": {
    "intent": "Slot-Confirmed",
    "confidence": 0.92,
    "explanation": "Customer agreed to Wednesday afternoon delivery."
  },
  "message": { "intent": null, "confidence": 0, "explanation": "…" }
}
```

## Resolution Analysis

Classifies whether the call achieved its objective, against a **resolution definition** you provide. Click **+ Add Resolution Definition** and describe in plain language what counts as resolved.

Example resolution definition:

> Call is resolved if the customer confirmed a delivery slot for the order. If a slot was discussed but not finalised, mark `PARTIALLY_RESOLVED`. If the customer refused or the call dropped, mark `UNRESOLVED`.

The payload then carries:

<ResponseField name="analysis.resolution.classification" type="enum">
  `RESOLVED` / `PARTIALLY_RESOLVED` / `UNRESOLVED` / `NOT_APPLICABLE`.
</ResponseField>

<ResponseField name="analysis.resolution.confidenceScore" type="number">
  `0` – `1` confidence in the verdict.
</ResponseField>

<ResponseField name="analysis.resolution.satisfiedChecklistItems" type="string[]">
  The checklist points met during the call.
</ResponseField>

<ResponseField name="analysis.resolution.evidence" type="string[]">
  Quoted turns from the transcript supporting the verdict.
</ResponseField>

<ResponseField name="analysis.resolution.explanation" type="string">
  Short rationale.
</ResponseField>

```json theme={null}
"resolution": {
  "classification": "RESOLVED",
  "confidenceScore": 0.88,
  "satisfiedChecklistItems": ["delivery_slot_agreed"],
  "evidence": ["Customer: yes Wednesday afternoon works."],
  "explanation": "Customer agreed to a specific slot."
}
```

## Sentiment

Returns a score and label per call:

<ResponseField name="analysis.sentiment.conversation.sentiment" type="enum">
  `POSITIVE` / `NEUTRAL` / `NEGATIVE`.
</ResponseField>

<ResponseField name="analysis.sentiment.conversation.score" type="number">
  `-1.0` – `+1.0`.
</ResponseField>

<ResponseField name="analysis.sentiment.conversation.detectedLanguage" type="string">
  ISO-639-1 code; useful when callers may switch language mid-conversation.
</ResponseField>

```json theme={null}
"sentiment": {
  "conversation": {
    "sentiment": "POSITIVE",
    "score": 0.4,
    "detectedLanguage": "en",
    "explanation": "Customer was cooperative throughout."
  }
}
```

## Data Extraction

Define a structured schema and Agntix will pull the values out of the transcript.

Click **+ Add Data Extraction** and define one schema entry at a time. Each entry asks for:

* **Name** — the JSON key that will appear under `analysis.extractedData.<name>` in the webhook / session API.
* **Description** — a one-line hint that tells the analysis model what to look for. Treat this like a mini-prompt: the more precise, the better the extraction.
* **Type** — `String`, `Number`, `Boolean`, or `Enum` (for `Enum` you'll also be asked for the allowed values).

Repeat for each field. There is no hard cap, but **five tight fields beat fifteen fuzzy ones**.

Example schema for a delivery-confirmation call:

| Name                  | Description                                              | Type                                                            |
| --------------------- | -------------------------------------------------------- | --------------------------------------------------------------- |
| `slotConfirmed`       | True if the customer agreed to a specific delivery slot. | Boolean                                                         |
| `preferredDay`        | The day the customer chose for delivery.                 | String                                                          |
| `preferredTimeWindow` | Time window the customer preferred.                      | Enum (`Morning`, `Afternoon`, `Evening`)                        |
| `callbackRequested`   | True if the customer asked to be called back later.      | Boolean                                                         |
| `callOutcome`         | Short categorical outcome.                               | Enum (`Confirmed`, `Rescheduled`, `Unavailable`, `WrongNumber`) |

Extracted values surface under `analysis.extractedData.<name>` in the webhook and the session detail API:

```json theme={null}
"extractedData": {
  "slotConfirmed": true,
  "preferredDay": "Wednesday",
  "preferredTimeWindow": "Afternoon",
  "callbackRequested": false,
  "callOutcome": "Confirmed"
}
```

<Tip>
  Keep field descriptions specific and unambiguous. A description like *"True if the customer agreed to a specific delivery slot (e.g. 'Wednesday afternoon' counts, 'sometime next week' does not)"* is much more reliable than *"Did they confirm?"*.
</Tip>

## Combined `analysis` block

When all four toggles are on, the full `analysis` block looks like this:

```json theme={null}
"analysis": {
  "summary": "Customer confirmed delivery slot for Wednesday afternoon.",
  "intent": {
    "conversation": { "intent": "Slot-Confirmed", "confidence": 0.92, "explanation": "…" },
    "message":      { "intent": null, "confidence": 0, "explanation": "…" }
  },
  "resolution": {
    "classification": "RESOLVED",
    "confidenceScore": 0.88,
    "status": "Completed",
    "satisfiedChecklistItems": ["delivery_slot_agreed"],
    "evidence": ["Customer: yes Wednesday afternoon works."],
    "explanation": "Customer agreed to a specific slot."
  },
  "sentiment": {
    "conversation": { "sentiment": "POSITIVE", "score": 0.4, "detectedLanguage": "en", "explanation": "…" },
    "message":      { "sentiment": null, "score": 0, "detectedLanguage": "N/A", "explanation": "…" }
  },
  "extractedData": {
    "slotConfirmed": true,
    "preferredDay": "Wednesday",
    "preferredTimeWindow": "Afternoon",
    "callbackRequested": false,
    "callOutcome": "Confirmed"
  }
}
```

## Next steps

<CardGroup cols={2}>
  <Card title="Webhook events" icon="webhook" href="/webhooks/events">
    See the full `session.ended` payload that carries the `analysis` block.
  </Card>

  <Card title="Get Session" icon="code" href="/api-reference/chat/sessions/get">
    Pull the same `analysis` block via REST when you need it on demand.
  </Card>
</CardGroup>