See where friction lives, by subject
SentiTrack summarises every scored message to a stable topic — so "damaged" and "defective" land under one heading. Now you can read sentiment by subject, pricing, onboarding or returns, instead of only by person.
Topic clustering is the automatic grouping of messages by what they are about, collapsing paraphrases into one stable subject. SentiTrack summarises every scored message to a topic and reuses the existing topic when the subject matter matches — so \"damaged\" and \"defective\" stay together — letting you read sentiment by theme, not just by person.
Person-level tone hides the why
Knowing a relationship is souring tells you who, never what. The same complaint shows up as "damaged", "defective" and "broken on arrival", so it scatters across messages and never adds up. By the time a pattern is obvious, the subject driving the friction has been live for months.
One stable topic across every paraphrase
SentiTrack summarises each message to a topic as it is scored, and reuses an existing topic whenever the subject matter matches. Semantically similar messages — different words, same meaning — collapse into a single heading, so sentiment rolls up cleanly by subject and the real source of friction surfaces on its own.
- Topics generated automatically — no taxonomy or keyword lists to maintain
- Paraphrases merge: "damaged" and "defective" become one topic
- Read average sentiment per topic alongside who and when
- Works across email, voice, chat and the /ingest API, in 7 languages
How topics are clustered
Email, calls, chat and any /ingest text are scored 1–10 — the same engine behind every SentiTrack view — then the body is discarded.
As it scores, the model condenses each message to a short subject and reuses an existing topic whenever the meaning matches, so paraphrases stay grouped under one stable heading.
Scores aggregate per topic so you can sort by most-used, browse A–Z, filter a graph by topic, or split a time series by the themes driving the trend.
See whether friction clusters around pricing, onboarding or returns — and act on the cause, not the symptom.
"Damaged", "defective" and "broken" merge into one topic, so a recurring issue shows its true volume.
Topics are generated automatically from your own traffic — nothing to configure, no keyword lists to keep current.
Filter the Net Graph or split a Time Graph by topic to isolate exactly which subject is moving sentiment.
Messages in any of 7 scored languages cluster to coherent topics with no per-language setup.
Topics are short summaries plus the 1–10 score — the original body is never stored.
Topics are summaries, never stored bodies
Topic clustering reads tone and subject without keeping content. SentiTrack scores each message, derives a short topic, and discards the body — only the score, participants, timestamp, direction, a subject hash and the topic summary are retained, with configurable retention and a hard-delete endpoint.
- 0 message bodies, subject text or attachments stored
- Only score + metadata + subject hash + topic summary persisted
- Configurable retention (default 365 days) and hard-delete on request
- Self-host with SentiTrack Edge so content never leaves your network
Topic clustering — FAQ
How does topic clustering work in SentiTrack?
As each message is scored 1–10, the model summarises it to a short topic and reuses an existing topic whenever the subject matter matches. Paraphrases like "damaged" and "defective" collapse into one stable heading, so sentiment rolls up cleanly by subject across email, voice, chat and the /ingest API.
Does topic clustering store my message content?
No. SentiTrack stores only the 1–10 sentiment score, participant metadata, a timestamp, direction and a short topic summary. Message bodies, subject text and attachments are scored, then discarded — so you get subject-level insight without retaining what anyone actually wrote.
Do I have to set up a list of topics or keywords first?
No. Topics are generated automatically from your own communication traffic — there is no taxonomy, tagging or keyword list to build or maintain. The model proposes a topic per message and reuses existing ones as matching subjects recur, so the topic set grows with your data.
How are paraphrases of the same subject kept together?
The model summarises by meaning, not exact wording, so semantically similar messages stay under one topic. For example, "damaged", "defective" and "broken on arrival" about the same product all map to a single heading, giving each subject its true volume instead of scattering across near-duplicate labels.
Can I combine topics with the graphs and reports?
Yes. Topic is a filter and split dimension across SentiTrack: filter the Net Graph by topic, split a Time Graph series by topic to see which subjects move a trend, and surface topic breakdowns in scheduled PDF reports — all built from scores and metadata, never bodies.
Does topic clustering work in other languages?
Yes. SentiTrack scores and clusters natively across 7 interface languages — English, German, French, Italian, Spanish, Russian and Turkish, with Arabic also scored — and no per-language configuration is needed. Messages in different languages about the same subject group into coherent topics automatically.
See your friction sorted by subject
Open the live demo with sample data — no signup — and browse auto-generated topics, or connect your channels and watch your own subjects cluster as messages score.