Agents
System instructions, model parameters, knowledge scope, escalation rules, and testing practices that keep assistants accurate and reviewable.
Configuration is where product, support, and legal align: what the assistant may say, how it cites knowledge, and when it must hand off. FlexyAgents exposes these controls in the dashboard without requiring engineers for every copy tweak.
Treat instructions as living documents: version them in tickets, review after major product launches, and sync with structured Q&A when wording must be exact.
Write instructions as you would onboard a human agent: role, audience, taboo topics, escalation triggers, and formatting preferences. Avoid contradictory lines (“be brief” vs “always include five paragraphs”).
Pair tone guidance with examples of good and bad replies for your brand—models follow patterns better than abstract adjectives alone.
Match model capability to task complexity; smaller models may suffice for narrow FAQs while technical doc assistants benefit from larger context windows.
Temperature and related controls trade creativity for determinism; customer support in regulated industries often biases toward lower variance.
Attach only the bases an agent should see. Mixing internal HR docs into a public widget base is a scope error retrieval cannot fully fix.
If your plan includes automations or tools, document side effects clearly in instructions—when a webhook may fire, what data it sends, and failure behavior.
When bases contain screenshots, video tutorials, or podcasts, verify that ingestion completed (processed status) and that instructions tell the model how to refer to visual content (“see the screenshot in …”) versus inventing details.
Maintain a golden set of user questions with expected behaviors (answer, cite, refuse, escalate). Re-run after any instruction or model change.
Include edge cases: empty retrieval, conflicting docs, and profanity or prompt-injection attempts your policy expects you to block politely.
继续探索
平台功能 · HTTP API 与 Webhook · 快速开始 · 联系