Cold-Starting a Customer‑Support Knowledge Base: What to Seed It With

Description and use

A small, carefully curated set of 20–30 entries derived from real chat logs, after‑sales policy, and product parameters enables a chatbot to answer the most frequent questions consistently from day one, reducing manual effort.

Problem

A support chatbot needs a high‑quality knowledge base to answer repetitive questions from day one, but building a large corpus before launch is unnecessary and costly.

Outcome

A small, carefully curated seed corpus enables a chatbot to handle the most frequent and dispute‑prone questions consistently from day one, reducing manual intervention and improving customer experience.

Workflow

• Export the last month of real chat logs and group them by question type.
• Identify top‑frequency clusters (e.g., pricing, shipping, returns).
• Write entries with four parts: buyer’s question, concise answer, keyword group, and boundary condition.
• Include after‑sales policy entries written as fixed policy statements.
• Add product parameter entries as structured entries with synonyms.
• Ensure the first batch covers returns, refunds, and shipping‑time questions.
• After launch, weekly review miss logs to add new entries.
• Biweekly prune expired or changed content.
• Monthly regression‑test seed entries against current rules.

Evidence

Twenty to thirty well‑chosen entries can absorb most repetitive questions on day one, and the gaps get filled from miss logs afterwards.

Source

Make sure the first batch includes returns, refunds, and shipping‑time questions.

Source

Without the boundary, the bot will answer in scenes where it shouldn't.

Source