Streamlit-Testing / src /agents.py
ajaykumar0712's picture
Update src/agents.py
fc7f322 verified
Raw History Blame Contribute Delete
9.25 kB
import os
from crewai import Agent, Task, Crew, Process, LLM
from crewai.tools import tool
from rapidfuzz import process
from data_loader import teams
from probability_engine import elo_probability, monte_carlo_group
# Create valid team set from dataset
VALID_TEAMS = set(teams["team"].str.upper())
def match_team_name(name):
if not name:
return None
match, score, _ = process.extractOne(name.upper(), VALID_TEAMS)
if score > 70:
return match
return None
# os.environ["GROQ_API_KEY"] = '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'
os.environ["GROQ_API_KEY"] = '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'
os.environ["OPENAI_API_KEY"] = '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'
os.environ["OPENAI_BASE_URL"] = "https://openrouter.ai/api/v1"
@tool
def match_probability(teamA: str, teamB: str):
"""
Simulate match probability
Returns probability of a team winning.
"""
return elo_probability(teamA, teamB)
@tool
def simulate_group_stage(group: str):
"""
Simulate a tournament group using Monte Carlo simulations.
Returns probability of teams qualifying.
"""
return monte_carlo_group(group)
llm = LLM(
# model = "openrouter/qwen/qwen-2.5-14b-instruct",
# model = "openrouter/mistralai/mistral-7b-instruct",
model = "groq/llama-3.1-8b-instant",
temperature=0.2
)
# Get the intent
def detect_intent(user_query):
prompt = f"""
You are an intent classifier for a cricket app.
Return ONLY valid JSON. No explanation. No markdown.
The JSON must have EXACTLY these fields:
{{
"intent": "match_prediction" | "group_progression" | "tournament_prediction" | "unknown",
"teamA": "country name or null",
"teamB": "country name or null",
"group": "group letter like A, B, C or null",
"venue": "city or null",
"tournament_team": "country 1" | "country 2" | "country 3" | "country 4" | "null"
}}
Examples:
- "Who wins if India take on the Aussies" β†’ intent: match_prediction, teamA: India, teamB: Australia
- "Who advances from Group C" β†’ intent: group_progression, group: C
- "Who are the favourites to win" β†’ intent: tournament_prediction
- "Give me the semi finalists" β†’ intent: tournament_prediction
User: {user_query}
"""
response = llm.call(prompt)
try:
json_start = response.find("{")
json_end = response.rfind("}") + 1
clean_json = response[json_start:json_end]
parsed = json.loads(clean_json)
except:
parsed = {
"intent": "unknown",
"teamA": None,
"teamB": None,
"group": None,
"venue": None,
"tournament_team": None,
"context": None
}
# πŸ”₯ STEP 1 β€” normalize using fuzzy match
parsed["teamA"] = match_team_name(parsed.get("teamA"))
parsed["teamB"] = match_team_name(parsed.get("teamB"))
# πŸ”₯ STEP 2 β€” fallback extraction from query if LLM failed
if parsed["intent"] == "match_prediction":
if not parsed["teamA"] or not parsed["teamB"]:
query_lower = user_query.lower()
found = []
for team in VALID_TEAMS:
if team.lower() in query_lower:
found.append(team)
if len(found) >= 2:
parsed["teamA"] = found[0]
parsed["teamB"] = found[1]
# πŸ”₯ STEP 3 β€” final safety
if parsed["intent"] == "match_prediction":
if not parsed["teamA"] or not parsed["teamB"]:
parsed["intent"] = "unknown"
return parsed
statistician = Agent(
role="Cricket Statistician",
goal="Analyze match probabilities using Elo ratings",
backstory="A veteran cricket data scientist who specializes in statistical modelling and match prediction.",
tools=[match_probability, simulate_group_stage],
llm=llm,
verbose=True
)
strategist = Agent(
role="Cricket Strategist",
goal="Evaluate player strengths and match conditions",
backstory="A former international cricket coach who studies tactical matchups and team combinations.",
llm=llm,
verbose=True
)
chaos_agent = Agent(
role="Chaos Analyst",
goal="Identify upset scenarios",
backstory="A pundit who believes cricket is unpredictable and focuses on variance and surprise outcomes.",
llm=llm,
verbose=True
)
compere = Agent(
role="Debate Host",
goal="Summarize debate and provide the verdict",
backstory="A neutral studio host moderating a panel of cricket experts discussing match predictions.",
llm=llm,
verbose=True
)
def run_agents(query_type, teamA=None, teamB=None, venue=None, group=None):
if query_type == "match":
stats_description = f"""
You are analyzing the match between {teamA} and {teamB}.
IMPORTANT:
You MUST call the tool `match_probability`
with teamA="{teamA}" and teamB="{teamB}"
Do NOT estimate probabilities yourself.
Do NOT invent Elo ratings.
Use the tool output to produce your answer.
Venue: {venue if venue else "Neutral conditions assumed"}
Output:
- Team A win probability
- Team B win probability
- Short explanation
"""
strategist_description = f"""
Consider the statistical prediction for {teamA} vs {teamB}.
Now evaluate tactical factors:
- batting strengths
- bowling strengths
- venue conditions
Explain how these factors may influence the match.
"""
chaos_description = f"""
Challenge the prediction for {teamA} vs {teamB}.
Discuss possible upset scenarios.
Consider:
- volatility of T20 cricket
- early wickets
- momentum shifts
Explain when the weaker team might win.
"""
compere_description = f"""
Summarize the panel discussion for the match {teamA} vs {teamB}.
Combine insights from:
- statistics
- strategy
- chaos analysis
Provide final verdict:
- predicted winner
- probability
- key factors
"""
elif query_type == "group":
stats_description = f"""
Simulate Group {group} using the simulate_group_stage tool.
Provide:
- probability of each team qualifying
- explanation of results
"""
strategist_description = f"""
Interpret the Group {group} simulation results.
Discuss:
- strongest teams
- potential challengers
- possible group surprises
"""
chaos_description = f"""
Highlight possible upset scenarios in Group {group}.
Explain how weaker teams could still qualify.
"""
compere_description = f"""
Summarize the analysis for Group {group}.
Provide:
- top teams expected to qualify
- overall tournament outlook
"""
elif query_type == "tournament":
stats_description = "Analyze all teams and predict the top 4 semifinalists using historical Elo data."
strategist_description = "Discuss which teams have the best tactical setups for a World Cup campaign."
chaos_description = "Which dark horses could shock the world? Who could be the Zimbabwe of this tournament?"
compere_description = "Summarize the panel's tournament predictions. Name your top 4 and a winner."
else:
raise ValueError('query_type must be "match" or "group"')
stats_task = Task(
description=stats_description,
expected_output="""
Output format:
Team A:
Win Probability:
Team B:
Win Probability:
Explanation:
""",
agent=statistician
)
strategist_task = Task(
description=strategist_description,
expected_output="Strategic match analysis",
agent=strategist
)
chaos_task = Task(
description=chaos_description,
expected_output="Upset scenarios and uncertainty discussion",
agent=chaos_agent
)
compere_task = Task(
description=compere_description,
expected_output="Final verdict combining all insights",
agent=compere
)
crew = Crew(
agents=[statistician, strategist, chaos_agent, compere],
tasks=[stats_task, strategist_task, chaos_task, compere_task],
process=Process.sequential,
verbose=True
)
result = crew.kickoff()
# EXTRACT OUTPUTS
stats_output = stats_task.output.raw
strategist_output = strategist_task.output.raw
chaos_output = chaos_task.output.raw
compere_output = compere_task.output.raw
return {
"stats": stats_output,
"strategist": strategist_output,
"chaos": chaos_output,
"compere": compere_output,
"result": result
}
# return result