Bot Knowledge Base

Overview

The Knowledge Base is the document store that Workflow bots draw on when answering questions, using a RAG (Retrieval-Augmented Generation) approach. The data model has three layers.

Layer	Table	Meaning
Base	knowledge_base	A LINE OA can own several knowledge bases
Document	knowledge_document	A document that belongs to a base
Chunk	knowledge_chunk	A document split into pieces with embeddings for semantic search

Indexing and embedding are expensive operations, so they never run inside a request. They are handed off to the worker through the RabbitMQ queue knowledge_index.

Business Flow

Managing Bases

1. GET /api/knowledge-base lists every base belonging to the LINE OA.
2. GET /api/knowledge-base/select-options returns a dropdown-shaped payload so the workflow settings screen can pick a base before the token is bound to an OA. That is why this endpoint sits on the public group alongside JwtLoginAuth.
3. POST /api/knowledge-base creates a new base.
4. GET /api/knowledge-base/:id returns details and PUT /api/knowledge-base/:id edits them.
5. DELETE /api/knowledge-base/:id performs a two-stage soft delete: the status is set to delete first, then deleted_date is stamped, mirroring the order used in the original TypeScript implementation.

Managing Documents

6. GET /api/knowledge-base/:id/documents lists the documents in a base.
7. POST /api/knowledge-base/:id/documents adds a document via CreateDocumentDto and publishes an indexing job to the knowledge_index queue.
8. PUT /api/knowledge-base/:id/documents/:docId edits a document and triggers a re-index.
9. DELETE /api/knowledge-base/:id/documents/:docId removes a document by setting its status to delete and then soft-deleting it.

Reindexing an Entire Base

10. POST /api/knowledge-base/:id/reindex re-indexes the whole base. This is used after changing the embedding model or when chunk data has become corrupted.
11. The worker computes embeddings, writes them to knowledge_chunk, and indexes them into Meilisearch.
12. When the bot answers a question, the workflow sandbox and executor query Meilisearch together with the embedding API.

The published payload is kept byte-for-byte identical to the original, including key ordering from the source object literal, because the existing worker still consumes it. The routing key equals the queue name on the line_exchange exchange.

Key Files & Functions

The code lives in internal/modules/knowledgebase/, made up of controller.go, service.go, and dto.go.

Method	Route	Handler	Guard / Policy
GET	/api/knowledge-base/select-options	ct.getSelectOptions	public + JwtLoginAuth
GET	/api/knowledge-base	ct.findAll	authed + readAll line-oa
GET	/api/knowledge-base/:id	ct.findById	authed + read line-oa
GET	/api/knowledge-base/:id/documents	ct.listDocuments	authed + read line-oa
POST	/api/knowledge-base	ct.create	authed + create line-oa
POST	/api/knowledge-base/:id/documents	ct.addDocument	authed + create line-oa
POST	/api/knowledge-base/:id/reindex	ct.reindexAll	authed + update line-oa
PUT	/api/knowledge-base/:id	ct.update	authed + update line-oa
PUT	/api/knowledge-base/:id/documents/:docId	ct.updateDocument	authed + update line-oa
DELETE	/api/knowledge-base/:id	ct.softDelete	authed + delete line-oa
DELETE	/api/knowledge-base/:id/documents/:docId	ct.deleteDocument	authed + delete line-oa

Connections to Other Services

• Access control — The main routes sit on the authed group, which enforces JWT globally. Policy metadata references PolicyModuleLineOa but is not enforced, and there is no ModuleGate wrapper.
• Tables — knowledge_base, knowledge_document, knowledge_chunk, line_oa
• RabbitMQ — The knowledge_index queue on the line_exchange exchange
• External services — The embedding API and Meilisearch, used at query time (see Workflow for details)
• Error handling — throwInstanceofError passes HttpExceptions through unchanged; any other database error becomes an HTTP 500 with code APP_000.
• Related modules — Workflow and AI Configuration