零售RAG系统

E-commerce & Retail RAG with ADK + Gemma3:270m + Ollama + ChromaDB + MCP

A production-ready blueprint for building a low-latency, grounded RAG system for retail using:

• Google ADK (Agent Development Kit) for orchestration (agents, tools, workflows, evaluation)
• Gemma3:270m served via Ollama for generation (32K context, lightweight)
• ChromaDB for vector search with metadata filtering
• FastAPI MCP server exposing retail tools (product_search, inventory_check, policy_qa, analytics_query)

Architecture at a Glance

• ADK agent orchestrates: preprocess → retrieve (VectorSearchTool) → augment → generate (GenerateTool)
• Vector retrieval via ChromaDB; strict grounding in generation prompts
• MCP servers back tools for catalog/inventory/policy/analytics; host-enforced consent, schemas, RBAC

Quickstart

Option A: Docker Compose

docker compose up --build

Services:

• Ollama (pulls gemma3:270m on first run)
• ChromaDB (vector DB)
• FastAPI MCP server (http://localhost:8000)

Option B: Python venv

bash scripts/setup_venv.sh
# in three terminals (or use dev_up_local.sh)
ollama serve & && ollama pull gemma3:270m
chroma start --host 0.0.0.0 --port 8001 &
uvicorn src.mcp_server.main:app --host 0.0.0.0 --port 8000 --reload

Run checks:

bash scripts/checks.sh

Documentation Index

• Developer Guide: docs/developer-guide.md
• Developer Help Guide: docs/developer-help.md
• Testing Strategy: docs/testing-strategy.md
• ADK Integration: docs/adk-integration.md
• Comparison Tables: docs/comparison-tables.md
• Diagrams (Index): docs/diagrams/README.md
  • Data Flow: docs/diagrams/01_data_flow_diagram.md
  • MCP Architecture: docs/diagrams/02_mcp_architecture.md
  • ChromaDB Schema: docs/diagrams/03_chromadb_schema.md
  • Deployment: docs/diagrams/04_deployment_architecture.md
  • Gemma Optimization: docs/diagrams/05_gemma_optimization.md
  • Customer Journeys: docs/diagrams/06_customer_journeys.md

Repo Layout

configs/           # env templates, config files
scripts/           # dev scripts (venv, compose, checks)
src/
  mcp_server/      # FastAPI MCP server
  rag/             # vector store adapter
  embeddings/      # ollama client (generate, TODO: embed_text)
agents/
  tools/           # ADK tools (vector search, generate, MCP wrappers)
  eval/            # ADK evaluation cases
  rag_agent.py     # Sequential RAG agent
  workflows.py     # example workflows
docs/
  diagrams/        # mermaid diagrams

Next Steps

• Implement embed_text in src/embeddings/ollama_client.py (use Ollama embeddings if available; otherwise a sentence-embedder)
• Ingest catalogs/policies/reviews into Chroma (field-aware chunks)
• Flesh out MCP tools and schemas; enforce RBAC/rate limits
• Add reranker and evaluation harness (recall@k, groundedness)
• Wire metrics/logging/tracing, then stage and load-test