Guide🇹🇷 Türkçe

Knowledge Base

The bot doesn't make answers up: it generates them from the documents in the Knowledge Center (RAG). This guide covers document management, source citation, and good knowledge-base practices.

Uploading Documents

  1. Open Admin Panel → Knowledge Center and pick a dataset on the left.
  2. Choose your file with the Upload button.
  3. After upload, the document indexes automatically; processing status shows in the list. Once indexing finishes, the bot starts answering from the new content.

Supported types: .md, .txt, .pdf, .docx, .csv, .json, .html — up to 15 MB per file.

Uploading and deleting require the administrator role and are written to the audit log. A deleted document leaves the bot's knowledge; historical conversations are unaffected.

Source Citation

When the bot answers from the knowledge base, the conversation records which documents the answer was grounded in (up to 3 distinct documents):

  • Agents see the sources in Chatwoot and you see them in the panel — "where did the bot get that?" is always answerable.
  • Only the document name goes to the customer side; internal details (dataset, match score) stay in system logs.

Good Knowledge-Base Practices

  • Short, focused documents beat long mixed ones: let each document cover a single topic ("Refund policy", "Password reset").
  • Write in the customer's language — matching improves when the words customers actually ask with appear in the document.
  • Keep it current: delete or replace outdated documents when features change; the bot cannot tell old information from new.
  • Feed it from knowledge gaps: the knowledge-gaps section of the weekly report lists questions the bot could not answer — every gap is the topic of a new document to upload. This loop grows your knowledge base from real customer questions.

Technical Details

  • Storage and indexing run on Dify (economy indexing, automatic chunking). Advanced settings (chunk size, embedding model) are reachable from the ai.<your-domain> console if ever needed; day-to-day use doesn't require it.
  • Documents uploaded from the Knowledge Center and via MCP (the upload_document tool) land in the same dataset.

Troubleshooting

Symptom Check
"This action requires an administrator" Your account has the viewer role; ask an administrator for access.
Upload fails with the 15 MB error Split the file or use a compressed PDF.
Document uploaded but the bot ignores its content Wait for indexing to complete; make sure your question overlaps with the document's wording.