client_agent.py

# client_agent.py
import os
import asyncio
from dataclasses import dataclass
from dotenv import load_dotenv
load_dotenv()

from pydantic_ai import Agent, RunContext
from pydantic_ai.mcp import MCPServerStdio
from pydantic_ai.providers.groq import GroqProvider
from pydantic_ai.models.groq import GroqModel
from pydantic_graph import End

from tools.searching import SearchingTools

import pandas as pd
from typing import List, Dict, Any, Optional
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import io
from PIL import Image, ImageDraw, ImageFont

SYSTEM_PROMPT = """
# You are an Expert Football Analyst with access to Barcelona graphdatabase from five seasons - from 2016/2017 to 2020/2021.
Use the tools available to respond the user query. You are connected to neo4j by a MCP server created by Frederico Caixeta.
Base your analysis **ONLY** by the query results. If the database can't provide what the user is 
asking for, report that in a professional way. Limit your answer in **1500** characters.
When the user asks for visualizations, graphs, or charts, you MUST use the create_chart tool 
to generate the appropriate visualization. The chart will be displayed in the Visualization tab.
# Below you can find some cypher queries as example, so you can understand which artifacts and metadatas are available in the database:
// General Overview: players, connections, goals per temporada
MATCH (p:Player {team: "Barcelona"})
OPTIONAL MATCH (p)-[r:PASSED_TO]->()
OPTIONAL MATCH (g:GoalSequence {team: "Barcelona"})
RETURN 
    count(DISTINCT p) as TotalPlayers,
    count(DISTINCT r) as TotalPassConnections,
    sum(r.weight) as TotalPasses,
    count(DISTINCT g) as TotalGoalSequences,
    collect(DISTINCT p.season_date) as Seasons
// List all seasons available in neo4j
MATCH (p:Player)
RETURN DISTINCT p.season_date as Season, p.season_id as SeasonID
ORDER BY p.season_date
// Top 5 connections per season
MATCH (p1:Player)-[r:PASSED_TO]->(p2:Player)
WITH p1.season_date as Season, p1.name as P1, p2.name as P2, r.weight as Weight
ORDER BY Weight DESC
WITH Season, collect({passer: P1, receiver: P2, passes: Weight})[0..5] as TopConnections
RETURN Season, TopConnections
ORDER BY Season
// Connections between different zones in field
MATCH (p1:Player)-[r:PASSED_TO]->(p2:Player)
WHERE p1.season_id = 90
WITH p1, p2, r,
     CASE WHEN p1.avg_x < 40 THEN 'Def' WHEN p1.avg_x < 80 THEN 'Mid' ELSE 'Att' END as Zone1,
     CASE WHEN p2.avg_x < 40 THEN 'Def' WHEN p2.avg_x < 80 THEN 'Mid' ELSE 'Att' END as Zone2
WHERE Zone1 <> Zone2
RETURN Zone1 + ' -> ' + Zone2 as Transition, sum(r.weight) as TotalPasses
ORDER BY TotalPasses DESC
// Total number of sequences of goals that Rakitić (a player) was involved
MATCH (p:Player {name: "Ivan Rakitić"})-[:INVOLVED_IN]->(g:GoalSequence)
RETURN count(g) as TotalGoalSequences
# The Property Keys available are: avg_x, avg_y, data, end_x, end_y, id, match_id, name, nodes, num_passes, order, possession, relationships, season_date, season_id, sequence_id, style, team, visualisation, weight, x, y.
The Nodes are: Player, GoalSequence.
The Relationships are: INVOLVED_IN, PASSED_IN_SEQUENCE, PASSED_TO.
The seasons ids and their dates: [{90: '2020/2021'}, {42: '2019/2020'}, {4: '2018/2019'}, {1: '2017/2018'}, {2: '2016/2017'}]
# All players played in all seasons are:
Abel Ruiz Ortega
Aleix Vidal Parreu
André Filipe Tavares Gomes
Andrés Iniesta Luján
Anssumane Fati
Antoine Griezmann
Arda Turan
Arthur Henrique Ramos de Oliveira Melo
Arturo Erasmo Vidal Pardo
Carles Aleña Castillo
Carles Pérez Sayol
Claudio Andrés Bravo Muñoz
Clément Lenglet
Denis Suárez Fernández
Francisco Alcácer García
Francisco António Machado Mota de Castro Trincão
Frenkie de Jong
Gerard Deulofeu Lázaro
Gerard Piqué Bernabéu
Héctor Junior Firpo Adames
Ivan Rakitić
Jasper Cillessen
Javier Alejandro Mascherano
Jean-Clair Todibo
Jordi Alba Ramos
José Manuel Arnáiz Díaz
José Paulo Bezzera Maciel Júnior
Jérémy Mathieu
Kevin-Prince Boateng
Lionel Andrés Messi Cuccittini
Lucas Digne
Luis Alberto Suárez Díaz
Malcom Filipe Silva de Oliveira
Marc-André ter Stegen
Marlon Santos da Silva Barbosa
Martin Braithwaite Christensen
Miralem Pjanić
Moriba Kourouma Kourouma
Moussa Wagué
Munir El Haddadi Mohamed
Neymar da Silva Santos Junior
Norberto Murara Neto
Nélson Cabral Semedo
Ousmane Dembélé
Pedro González López
Philippe Coutinho Correia
Rafael Alcântara do Nascimento
Ricard Puig Martí
Ronald Federico Araújo da Silva
Samuel Yves Umtiti
Sergi Roberto Carnicer
Sergino Dest
Sergio Busquets i Burgos
Thomas Vermaelen
Yerry Fernando Mina González
Álex Collado Gutiérrez
Óscar Mingueza García
"""

api_key = os.getenv("GROQ_DEV_API_KEY")
groq_model = GroqModel(
    "moonshotai/kimi-k2-instruct-0905",
    provider=GroqProvider(api_key=api_key)
)

# ✅ MCP via stdio (subprocess) — ideal no Hugging Face
neo4j_server = MCPServerStdio(
    "python",
    args=["-u", "mcp_server.py"],
    env=dict(os.environ),
    tool_prefix="neo4j",
    timeout=80,
)

@dataclass
class SearchAgentDeps:
    tools: SearchingTools

agent = Agent(
    model=groq_model,
    toolsets=[neo4j_server],     # ✅ server MCP anexado
    system_prompt=SYSTEM_PROMPT,
    deps_type=SearchAgentDeps,
)

tools_instance = SearchingTools()
deps = SearchAgentDeps(tools=tools_instance)

@agent.tool(name="web_search", retries=3)
async def procura_web(ctx: RunContext[SearchAgentDeps], search_query: str) -> str:
    """Pesquisa na web"""
    return ctx.deps.tools.search_web(search_query=search_query, max_results=15)

last_chart_image = None

def create_placeholder_image():
    bg_color = (248, 250, 252)
    text_color = (100, 116, 139)
    img = Image.new("RGB", (800, 600), color=bg_color)
    draw = ImageDraw.Draw(img)
    try:
        font = ImageFont.truetype("/usr/share/fonts/truetype/dejavu/DejaVuSans.ttf", 48)
    except Exception:
        font = ImageFont.load_default()
    text = "Waiting for plot..."
    bbox = draw.textbbox((0, 0), text, font=font)
    x = (800 - (bbox[2] - bbox[0])) // 2
    y = (600 - (bbox[3] - bbox[1])) // 2
    draw.text((x, y), text, fill=text_color, font=font)
    return img

PLACEHOLDER_IMAGE = create_placeholder_image()

@agent.tool(name="create_chart", retries=2, timeout=30.0)
async def create_chart(
    ctx: RunContext[SearchAgentDeps],
    data: List[Dict[str, Any]],
    x_column: str,
    y_column: str,
    chart_type: str = "bar",
    title: Optional[str] = None,
    x_title: Optional[str] = None,
    y_title: Optional[str] = None,
) -> str:
    global last_chart_image
    try:
        df = pd.DataFrame(data)
        if x_column not in df.columns or y_column not in df.columns:
            return f"❌ Erro: Colunas não encontradas. Disponíveis: {list(df.columns)}"

        fig, ax = plt.subplots(figsize=(10, 6))

        if chart_type == "bar":
            ax.bar(df[x_column], df[y_column])
        elif chart_type == "horizontal_bar":
            ax.barh(df[x_column], df[y_column])
        elif chart_type == "line":
            ax.plot(df[x_column], df[y_column], marker="o")
        elif chart_type == "scatter":
            ax.scatter(df[x_column], df[y_column])
        else:
            return f"❌ Tipo '{chart_type}' não suportado"

        ax.set_title(title or f"{y_column} por {x_column}")
        ax.set_xlabel(x_title or x_column)
        ax.set_ylabel(y_title or y_column)
        plt.tight_layout()

        buf = io.BytesIO()
        plt.savefig(buf, format="png", dpi=150, bbox_inches="tight", facecolor="white")
        buf.seek(0)
        last_chart_image = Image.open(buf).copy()
        plt.close()

        return f"✅ Gráfico '{chart_type}' criado ({len(df)} registros)."
    except Exception as e:
        return f"❌ Erro ao criar gráfico: {str(e)}"

def get_current_chart():
    global last_chart_image
    return last_chart_image if last_chart_image is not None else PLACEHOLDER_IMAGE

async def stream_agent_response_safe(user_query: str) -> str:
    # ✅ manter iter() que já funciona
    async with agent.iter(user_query, deps=deps) as agent_run:
        async for node in agent_run:
            if isinstance(node, End) and agent_run.result:
                return str(agent_run.result.output)
    return "Erro na execução do agente"

__all__ = [
    "agent",
    "deps",
    "PLACEHOLDER_IMAGE",
    "get_current_chart",
    "stream_agent_response_safe",
]