Project Structure

DataMind/
├── config.py              # Config center (Pydantic Settings, reads from .env)
├── .env.example           # Environment variable template
├── server.py              # Web entry: FastAPI backend + frontend
├── main.py                # CLI entry: interactive terminal
├── benchmark/             # Concurrent inference benchmarking
│   ├── run.py             #   Benchmark runner (concurrent queries + metrics)
│   └── evaluate.py        #   Answer evaluation (EM / F1)
├── core/                  # Core layer
│   ├── bootstrap.py       #   Shared initialization logic (AppState)
│   └── session.py         #   Session isolation (SessionManager)
├── modules/               # Feature modules
│   ├── rag/               #   RAG vector retrieval
│   │   ├── indexer.py     #     Document loading + Chroma index
│   │   └── retriever.py   #     Retrieval strategies (Simple / MultiQuery)
│   ├── graphrag/          #   GraphRAG knowledge graph
│   │   └── graph_rag.py   #     Graph construction + query
│   ├── database/          #   Database NL2SQL
│   │   └── database.py    #     SQLite demo + NL2SQL engine
│   ├── skills/            #   Skills system
│   │   ├── tools.py       #     Tool skills: calculator, time, etc.
│   │   └── knowledge.py   #     Knowledge skills: Markdown doc retrieval
│   ├── memory/            #   Conversation memory
│   │   └── memory.py      #     Short-term + long-term memory
│   └── agent/             #   Agent orchestration
│       └── agent.py       #     FunctionAgent integrating all tools
├── data/                  # Data directory
│   ├── profiles/          #   Knowledge bases (isolated by DATA_PROFILE)
│   ├── bench/             #   Benchmark question sets
│   └── skills/            #   Skill documents (shared across profiles)
└── storage/               # Auto-generated: persisted indexes

Initialization Flow

core/bootstrap.py defines the AppState dataclass and the initialize() function:

initialize()
  │
  ├── Configure LlamaSettings (LLM + Embedding)
  ├── Build/load RAG index (Chroma)
  ├── Build/load GraphRAG index (NetworkX)
  ├── Init Database (SQLite + NL2SQL engine)
  ├── Build/load Skills index (Chroma)
  └── Create FunctionAgent (wire all tools)
        │
        └── AppState (holds all components)

Both server.py (Web) and main.py (CLI) call initialize() once at startup.

Agent Decision Flow

User Question
    │
    ▼
FunctionAgent receives question + tool descriptions
    │
    ▼
LLM decides which tool(s) to call
    │
    ├── knowledge_search  → RAG vector retrieval
    ├── graph_search      → GraphRAG entity/relation traversal
    ├── database_query    → NL2SQL → execute SQL
    ├── skill_search      → Knowledge skill retrieval
    ├── calculator / ...  → Tool skill execution
    └── (none)            → Direct LLM response
    │
    ▼
Combine tool results → Generate final answer

Session Isolation

core/session.py provides SessionManager for per-user memory isolation:

from core.session import SessionManager

session_mgr = SessionManager()
memory_a = session_mgr.get_memory("user_a")
memory_b = session_mgr.get_memory("user_b")

This is critical for concurrent benchmarking and multi-user Web serving.

Tech Stack

Component	Technology	Notes
Framework	LlamaIndex	Core orchestration
LLM	OpenAI-compatible API	No GPU required
Vector DB	Chroma	Local, pure Python
Graph Store	NetworkX	Local, pure Python
RDBMS	SQLite	Zero configuration
Agent	FunctionAgent	Automatic tool selection
Web Backend	FastAPI	Async, SSE streaming
Web Frontend	Pure HTML/CSS/JS	No npm, zero frontend deps