Pipeline stages
1. Ingestion
Conversations arrive via the SDK or REST API, authenticated by API key. Each conversation contains messages, metadata, and session context.2. Translation (optional)
If auto-translation is enabled in Analysis Config, messages in non-primary languages are translated before analysis. The original text is preserved alongside the translation.3. Analysis
Each message is processed through four configurable AI prompts:| Prompt | Output |
|---|---|
| Sentiment | Numeric score and label per message |
| Intent | Classified user intent (e.g., “asking about pricing”) |
| Topics | Extracted topics (e.g., “billing”, “integration”) |
| Flags | Detected user signals (churn risk, buying signal, etc.) |
- Helpful, Yes-Man, Dodging, Over-Apologizing, Hallucinating, Redirecting, Verbose, Robotic
- Fully Resolved, Partially Resolved, Unresolved
4. Storage
Analysis results are stored as annotations linked to individual messages and aggregated at the conversation level. The Analyzed Today KPI on the dashboard reflects the daily throughput of this stage.5. Surfacing
Annotations surface across the platform:| Where | What appears |
|---|---|
| Dashboard KPIs | Average sentiment, analyzed count |
| User Insights | Sentiment trends, top frustrations, revenue signals |
| Assistant Performance | Behavior alerts, resolution outcomes, quality flags |
| Conversation detail | Per-message clickable annotations |
| Conversations table | Sentiment column with color-coded badges |
| Workflow triggers | Conditions evaluated against analysis results |
Configurable prompts
Each chatbot has its own set of analysis prompts, editable from Analysis Config > Prompts tab. This means different chatbots can be tuned for different domains — a customer support bot and a sales bot can have different sentiment calibrations. Unmodified prompts use system defaults. See Customize analysis prompts for a guide on editing them.Semantic search infrastructure
The Deep Search feature relies on a separate vector embedding pipeline:- Conversation content is embedded into vector representations
- Vectors are stored in a vector database
- Search queries are embedded and compared using cosine similarity
- Results are ranked by semantic relevance
Workflow trigger evaluation
After analysis completes, OpenBat evaluates active workflows against the new annotations. If a workflow’s conditions match (e.g., sentiment below threshold, hallucination detected), the workflow triggers and executes its configured action (webhook delivery). Workflow execution results are logged as runs with status tracking (Pending, Triggered, Success, Failed, Skipped).Next steps
Analysis configuration
Configure the analysis pipeline for your chatbot.
Data model
Entity relationships and data structures.