RAG Knowledge Base: Train Your Assistant on Custom Data

Primary relaunch CTA: Explore Pricing
Explore Flows: <a href="/zh/features/%e5%8f%af%e8%a7%86%e5%8c%96%e6%b5%81%e7%a8%8b%e6%9e%84%e5%bb%ba%e5%99%a8/">可视化流程生成器

The Problem: Generic AI Hallucinations and Untrained Answers

Standard AI models do not know your shipping rules, return policies, or service pricing. If you connect a raw GPT or Claude API to your chat, it will make up answers, giving customers wrong information. Businesses need a secure way to ground responses in actual company documentation.

MessengerBot Capability: Retrieval-Augmented Generation (RAG)

Messenger Bot allows you to train your draft assistant using your own business records through our integrated Retrieval-Augmented Generation (RAG) system:

Approved Source Types: Upload your custom database context using four distinct source types: manual_qa (questions and answers), plain 文本 entries, public url references, and uploaded document files.
Smart Chunking: Text is split into manageable blocks ranging from 80 to 2,000 characters (defaulting to 500) to support more focused context retrieval.
On-Server Retrieval Database: Queries utilize local MySQL fulltext search indexing and term-count ranking rather than expensive external vector databases, saving database overhead.
Source-Backed Limits:
- 入门级： 0 knowledge sources.
- 优点: up to 5 knowledge sources (Beta).
- Agency: up to 50 knowledge sources (Beta).

Safety Boundary: Local Validation and No File-Size Marketing

BYOK and RAG are labeled as advanced beta configurations during our launch window. File-size upload constraints must be verified in the member dashboard before publishing public limits. All source write actions are guarded via server-side check flags.

The Outcome: Accurate, Grounded Inbox Suggestions

Support agents receive draft suggestions grounded in company facts. Less time is spent searching documents, and operators have better context before deciding what to send.