Glossary

Terminology used across rendfly.

Terms used throughout the rendfly docs and dashboard, in alphabetical order.

Agent — the conversational AI that rendfly monitors. Typically a WhatsApp support bot, a Slack help assistant, a Telegram agent, or a web chat widget powered by an LLM. rendfly is the layer that watches whether the agent is behaving correctly in production.

Alert — a notification fired when a score drops below a configured threshold or when drift exceeds the baseline by more than the configured delta. Delivered to email, Slack, or a custom webhook depending on your plan and project settings.

Baseline — a rolling 7-day average of conversation scores for a project. Drift detection compares recent scores against this baseline to identify when behavior is shifting. The baseline needs at least 7 days of data before drift alerts activate.

Conversation — one complete back-and-forth between a user and the agent: all turns from the first user message to the last agent reply. The unit of judgment. A single WhatsApp session, one Slack thread, one web chat session — all count as one conversation.

Drift — gradual deviation from baseline behavior. Not a single bad reply, but a sustained downward trend in the rolling score average. See Drift detection for the full explanation of how rendfly measures and alerts on it.

Judge — the LLM call that evaluates a conversation against the project’s extracted rules and produces a per-rule verdict and an aggregate score. rendfly uses a sandwich pattern to reduce prompt injection risk. See How rendfly judges conversations.

Project — one agent. Each project has one system message, one LLM provider connection, one set of extracted rules, one alert configuration, and its own score history. A workspace can have multiple projects — one per agent — on Agency and Enterprise tiers.

Provider — the LLM API your agent uses: OpenAI, Anthropic, Google, Groq, Mistral, or any compatible provider. rendfly connects to your provider via proxy mode (change base_url), API mode (read-only admin key for OpenAI Stored Completions), or SDK wrapper (one-line wrap in your application code).

Rule — one extracted constraint from the system message. The atomic unit that the judge evaluates. Examples: “do not quote prices,” “always reply in English,” “escalate refund requests to a human.” Rules are extracted automatically but can be edited, disabled, or supplemented with custom ones in the project dashboard.

Score — a 0–100 aggregate per conversation. Each rule that fails subtracts from the maximum. The aggregate score is the primary input to both per-conversation verdicts and the drift detection baseline. Configurable alert threshold; default fires when the rolling 24-hour average drops below 80.

System message — the prompt that defines your agent’s identity, role, constraints, and behavior. In rendfly’s model, this is the contract: rendfly treats it as the binding specification that every production conversation should honor. When the system message changes, rule extraction reruns automatically.

Workspace — the top-level billing and access boundary. Contains projects, users, and plan settings. On the Indie plan, one workspace with one project. On Agency, one workspace with up to 10 projects and 5 users. On Enterprise, unlimited.