MD
Architectural Recommendation & Flow
Decision: Yes — Separate the Knowledge Base ingestion/retrieval pipeline from the AI agents.
Do not make RAG a normal "tool" owned by srcsingle or srcflow. Make it a first-class platform subsystem:
graph TD
A[Agent Runtime] -->|Retrieve Context| B(KB Retrieval Service)
B --> C[(Qdrant Vector DB)]
B --> D[(MySQL Metadata)]
E[Ingestion Workers] -->|Store Vectors| C
Key Implementation Subsystems
- KB Management API: Clean REST/gRPC endpoints
- Ingestion Workers: Asynchronous chunking and embedding pipeline
- Retrieval Service: High-speed hybrid vector search
- Vector Storage: Scalable Qdrant instance
"Use logical decoupling first, then physical service extraction later."
| Component | Responsibility | Tech Stack |
|---|---|---|
| Agent Engine | Execution & Tool Calling | FastAPI / Python |
| Vector Index | High-dim Similarity Search | Qdrant |
| Metadata | Schemas & Versioning | MySQL |
156 Words1030 Characters~1 min read
Client-side Markdown & Mermaid Engine