Improvement levers at a glance — data-driven.
The analysis shows you what really happens inside your bot. Which questions come up often, where zep is unsure, which sources stay empty. Instead of ten dashboards you get three views: conversations, gaps and suggestions — and you can trigger fixes right from there.
Three steps, one clear workflow.
No plugin, no wizard with required fields. You work at your own pace and jump between the steps — everything syncs back to the dashboard automatically.
- 1. Conversations
See what users really ask.
All chats are anonymized and grouped by cluster. You see the most common question types, where the bot takes a long time and which conversations get abandoned — that is your user reality.
- 2. Gaps
Where the bot answers with uncertainty.
Confidence scores per answer, sorted by frequency. Sources that are never hit are often outdated. Sources that always end without a match are missing entirely — you get both as a list.
- 3. Suggestions
Concrete tasks, no gut feeling.
From the gaps, zep generates mini-tasks: "Express shipping is asked often but not explained in your sources — add a new FAQ entry?" Click -> opens Training with the matching slot.
Three pipelines, one insights picture.
While your bot answers, a mini telemetry layer runs in the background — all tenant-isolated, anonymized and EU-hosted:
- Conversation aggregation. For each chat, question embeddings are built and clustered. This lets you see topic frequency without us storing original questions longer than necessary. PII filters run before aggregation, not after display.
- Confidence scoring. Every answer gets a multi-part score: how close the matched snippets were, how much the LLM took from those snippets, how clear the re-ranking was. That becomes the visible confidence — you spot weaknesses before your users do.
- Suggestion engine. A small rule engine compares frequent questions against the knowledge base. When a cluster is asked often but scores low on confidence, an improvement suggestion is generated — with a preview text and a direct jump into Training.
Typical use cases we see every day.
Service bot tuning
You see that shipping questions make up 40% of all requests but only have 70% confidence. Click the suggestion -> add two FAQ entries -> confidence rises to 92% next week.
Course instructor reporting
Which topics block learners most often? The cluster overview shows it. You build targeted teaching material instead of producing PDFs all over the place.
Stakeholder report without Excel
PNG export of the key KPIs for the monthly meeting. Conversation count, top topics, confidence trend — straight from the tool, no BI setup.
Multi-bot comparison
With several bots you see in parallel how volume and confidence develop. That tells you which bot needs care and which one runs on its own.
Coding bot analysis
Which patterns do developers ask about most often? Which resources go unexplained? From that you can add your own snippets in a targeted way.
Private bot audit
For team-internal bots: who asks what, without naming individuals. You see the knowledge needs in your team and plan onboarding material more precisely.
FAQ.
Do you store my bot conversations permanently?
No. We keep aggregated clusters for 90 days for trend analysis. Original conversations live on your tenant in the EU and are viewable for 30 days — after that they are anonymized or deleted, depending on your plan setting. You can configure retention per bot.
How exactly does the confidence score work?
Three components: (1) vector distance between the question and the best snippets, (2) how clearly the top sources stand out from the rest (re-ranking margin), (3) the answering LLM's own assessment of its certainty. From this we compute a score between 0 and 100 that is shown to you and optionally to the end user.
Do I get to see PII data?
By default, no. We filter email addresses, phone numbers and IBAN patterns before storage. On the Business plan you can enable PII display for owner roles — it is then audit-logged and visible only to you.
Can I export analysis data?
Yes. CSV export of the clusters, JSON export of individual conversations, PNG export of the charts. The Pro plan additionally includes a webhook that reports important events (e.g. a sharp confidence drop) to your Slack/Discord integration.
Know instead of guess — improve your bot with data.
Even on the Free plan you see top questions and confidence per answer. The Pro plan unlocks cluster suggestions and webhook alerts.
