Datasets

Datasets are where your interaction data lives — tickets, calls, chats, CSAT responses, CRM metadata. They support the full data lifecycle: ingestion, enrichment, search, and discovery.

Data Structure

Each record in a dataset represents a unit of customer interaction — an individual utterance, a full conversation, or an entire errand. The text column holds the content your models classify. Everything else is metadata: who handled it, when, which queue, what product, and any custom fields your team tracks.

Record Example

A single dataset record with text and metadata fields. When models classify this record, the predictions are stored alongside the original data — making everything queryable in dashboards and exports.

{
  "text": "I've been waiting 3 weeks for a replacement router and no one has called me back",
  "agent_id": "agent_4821",
  "customer_id": "cust_90312",
  "queue": "technical_support",
  "channel": "phone",
  "product": "fiber_100",
  "csat_score": 2,
  "handle_time_sec": 487,
  "created_at": "2025-11-14T09:22:00Z"
}

Data Sources

Labelf ingests data from every channel where customers interact with your team. Tickets, calls, chats, emails, CSAT/NPS surveys, and CRM metadata — all flow into datasets with full metadata preserved.

Call transcripts can be imported from your existing transcription provider or transcribed directly through Labelf's built-in speech-to-text pipeline.

Ingestion Methods

Four ways to get data in, from real-time streaming to bulk historical imports.

Search Within Datasets

Every dataset is fully searchable. Combine semantic search (find conversations similar to a query, regardless of exact wording), keyword search (exact term matching), and metadata filters — date ranges, agent, team, product, channel, CSAT score — to zero in on exactly the interactions you need.

This is how quality teams audit thousands of calls efficiently: search for "mentioned competitor" across all calls this quarter, filtered to the retention queue, sorted by churn risk score.

Auto-Clustering & Discovery

Not sure what categories exist in your data? Labelf's auto-clustering analyzes your dataset and automatically discovers groupings — no predefined labels required. It finds patterns humans miss: a telecom might discover that 12% of "billing" calls are actually about a confusing invoice redesign, or that a specific product SKU drives 3x the complaint volume.

Clusters become the starting point for your classification models. Instead of guessing what labels you need, you let the data tell you.

Data Lifecycle

The typical flow:

1. Create a dataset with a schema — define the text column and any metadata fields
2. Connect a data source or upload historical data (CSV, JSON, API push)
3. Discover — run auto-clustering to understand what topics exist in your data
4. Apply models for classification — every record gets scored automatically
5. Query results via the API, dashboards, or exports to your data warehouse
6. Refill — scheduled syncs keep the dataset current without manual uploads