Preview classification
Preview classification runs the AI against a small subset of responses (typically 20–50 rows) so you can check whether the topics are being applied as expected before spending tokens on the full dataset.Confirm your topics
In the Topics editor, click Confirm topics (or the equivalent confirm button) once your code plan is ready.
Run a preview
Click Preview classification. The AI classifies the sample and shows results immediately in the codes view.
Review results
Scan the classified responses. If topics are being applied incorrectly — for example, too broadly or missing obvious matches — go back to the Topics editor and refine your labels or guidelines.
Classification configuration
Before launching a full classification, you can adjust the following parameters in the configuration panel.Depth
| Setting | Behavior |
|---|---|
| L1 only | Assigns only top-level topics to each response. Faster and cheaper. |
| L1 + L2 | Assigns parent topics and their sub-topics. Use when you need granular sub-codes. |
Tolerance
Tolerance controls how strictly the AI applies topic assignments, on a scale of 1–5.- Low tolerance (1–2): The AI only assigns a topic when it is confident. Fewer topics per response, lower false-positive rate.
- High tolerance (4–5): The AI assigns topics more freely. More topics per response, but may include weaker matches.
Max codes per response
Set a hard limit on how many topics can be assigned to a single response. Set to 0 for no limit. For L1+L2 classification, you can set separate limits for L1 topics, L2 sub-topics per response, and L2 sub-topics per L1 parent.Consensus settings
Pulse Classifier runs multiple AI passes per response and uses a voting mechanism to decide which topics to keep. This improves accuracy at the cost of additional token usage.| Setting | Description | Default |
|---|---|---|
| Consensus N | Number of classification passes per response batch. | 20 |
| Consensus threshold | Minimum votes required to assign a topic. | 17 out of 20 |
| L1 min votes | Minimum votes to keep a parent L1 topic. | 2 |
Higher consensus settings improve accuracy but increase token consumption roughly proportionally to the number of passes. For large datasets, consider reducing consensus N if token cost is a concern.
Running the full classification
Click Classify to launch the full classification job. Pulse Classifier submits the job to the background queue and shows a progress indicator in the classification drawer. The drawer displays:- Status — Pending, processing, or complete.
- Progress bar — Percentage of responses classified.
- Estimated time remaining — Updates as the job progresses.
Token consumption
Token usage depends on the number of responses, the classification depth, and the consensus N setting. As a rough guide:- L1-only classification with default consensus uses approximately 1× the response count in API calls.
- L1+L2 classification doubles the number of calls.
- Increasing consensus N multiplies token usage proportionally.
Manual corrections
After classification completes, you can correct individual topic assignments directly in the results view.Add or remove a topic
Use the topic dropdown on the response row to add a topic, or click the × on an existing topic badge to remove it.