Skip to content
product

Howzer v2.5.2-DE: full analysis envelope, faster analysis, long-text severity

The v1 analyze response now returns the full analyzer breadth — purely additive, no breaking changes. Independent analyzers run in parallel for a 17% lower median latency, the severity model was retrained for long texts, and churn alerts were re-tuned on real data.

By Howzer Team, Product

What's in v2.5.2-DE

v2.5.2-DE builds on the v2.5.1 maintenance work and focuses on the API surface and analysis quality: the v1 analyze response now exposes the full breadth of what the 12 analyzers compute, independent analyzers run in parallel for faster responses, the severity model was retrained to handle long feedback texts, and churn alert thresholds were re-tuned against measured data.

−17%
Median analysis latency
0
Breaking changes
0
New dependencies

Full analysis envelope in the v1 API

Until now, the analyze response returned only a subset of what the pipeline actually computes. The v1 envelope now carries the full analyzer breadth — purely additive, so existing consumers keep working unchanged.

  • Sentiment distribution and analyzer metadata, plus emotion and root-cause metadata.
  • Severity confidence and the four severity dimension scores.
  • Risk rules, and the escalation queue with SLA, reasons, and recommended measures.
  • Competitor switching risk, geo coordinates, and new language and text_metadata sections.
  • PII boundary: reference numbers appear only as type and count — raw values never leave the pipeline.

Faster analysis

  • Independent analyzers now run in parallel instead of sequentially — median (p50) latency drops by 17%, with field-for-field equivalent results.
  • Systematics write paths are batched, cutting index scans by roughly 58×, complemented by index hygiene applied via database migration.

Severity ML v2.2: long texts

The severity model's input window grows from 192 to 512 tokens, and the model was retrained with long-text golden labels. Triage accuracy on feedback longer than 1,500 characters improves from around 0.50 to 0.74, while short texts stay strong (0.98 at up to 500 characters).

Churn alerts re-tuned on data

The churn alert threshold moves from 0.50 to 0.40, validated on 10,000 feedbacks: recall improves from 0.72 to 0.80 at practically unchanged precision — more at-risk customers are caught without flooding the queue.

Upgrade notes

The severity model is published on the model hub as ml-v2.2.0-de. Fresh v2.5.1 installations are not compatible with the current model state — existing installations keep running, but should upgrade to v2.5.2. A database migration is required (index-only changes).
This release requires no new dependencies. The health endpoint reports the canonical version 2.5.2.