mrra 0.1.1

Mobility Retrieve-and-Reflect Agent: GraphRAG + Multi-Agent Reflection for mobility trajectories

MRRA: Mobility Retrieve-and-Reflect Agent

Overview

• MRRA is a plug-and-play Python package for mobility trajectories with columns user_id, timestamp, latitude, longitude.
• It composes a GraphRAG retriever with multi-agent reflection to support tasks like next point, future time, and full-day trajectory.

Features

• Data normalization (TrajectoryBatch) and mobility graph construction (MobilityGraph).
• Graph-based retriever (GraphRAGGenerate) returning LangChain Documents.
• Multi-agent reflection with strict JSON outputs and configurable aggregator.
• MCP tools integration (adapters-first) with graceful fallback.

Install

• From source (recommended in a virtual env):
  • pip install -e .
• Optional MCP extras: pip install -e .[mcp]

Quickstart

import pandas as pd
from mrra.data.trajectory import TrajectoryBatch
from mrra.graph.pattern import PatternGenerate
from mrra.retriever.graph_rag import GraphRAGGenerate
from mrra.agents.builder import build_mrra_agent

df = pd.DataFrame({
    'user_id': ['user_1','user_1','user_1'],
    'timestamp': ['2023-01-01 09:00:00','2023-01-01 12:00:00','2023-01-01 18:00:00'],
    'latitude': [31.2304,31.2404,31.2504],
    'longitude':[121.4737,121.4837,121.4937],
})

tb = TrajectoryBatch(df)
pattern = PatternGenerate(tb)
retriever = GraphRAGGenerate(tb=tb)

agent = build_mrra_agent(
    llm={
        # OpenAI-compatible endpoint
        "provider":"openai-compatible",
        "model":"qwen-plus",
        "base_url":"https://dashscope.aliyuncs.com/compatible-mode/v1",
        "api_key":"YOUR_API_KEY"  # prefer env var in production
    },
    retriever=retriever,
    reflection={
        "max_round": 3,
        "subAgents": [
            {"name":"temporal", "prompt":"Temporal reasoning subagent.", "mcp": {"weather":{}}},
            {"name":"spatial",  "prompt":"Spatial reasoning subagent.",  "mcp": {"maps":{}}},
            {"name":"pattern",  "prompt":"Profile/pattern subagent.",  "mcp": {}},
        ],
        "aggregator": "confidence_weighted_voting"
    }
)

res = agent.invoke({"task":"next_position", "user_id":"user_1", "t":"2023-01-02 09:30:00"})
print(res)

Tasks

• next_position: next point after time t.
• future_position: point at a future time t.
• full_day_traj: full-day path for a given date.

MCP Integration (Amap example)

• Configure subagents with MCP:

reflection={
  "subAgents": [
    {"name":"spatial", "prompt":"...", "mcp": {"gmap": {"url": "https://mcp.amap.com/sse?key=YOUR_AMAP_KEY", "transport":"sse"}}}
  ]
}

• The package uses langchain-mcp-adapters first; falls back to langchain-mcp toolkit; and finally to raw MCP SSE discovery.

Data format

• Required columns: user_id, timestamp (ISO), latitude, longitude.
• Optional tz handling via TrajectoryBatch(..., tz=...).

Dev and demos

• Create env (example): conda create -n mrra-py310 python=3.10
• ISP data demo: python scripts/verify_isp.py (expects scripts/isp as input; file is git-ignored by default)

Security notes

• Do not commit API keys. Prefer environment variables or secret managers.
• Large data (like scripts/isp) is git-ignored by default.

License

• Specify your license here.

中文文档

• See README_zh.md for Chinese documentation.