Datavision / backend /core /deep_agents.py
DataVision CI/CD Bot
release: clean production build for HuggingFace Space
09801ca
Raw
History Blame Contribute Delete
21.3 kB
"""
🤖 DEEP AGENTS - Advanced Agent Patterns
========================================
Silicon Valley-grade autonomous agents with multiple reasoning strategies.
Agent Types:
- ReActAgent: Reason + Act loop with tool calling
- PlanAndExecuteAgent: Plan first, then execute
- ReflexionAgent: Self-reflection and improvement
- HybridAgent: Combines multiple strategies
Features:
- Tool registry and execution
- Thought chains and reasoning traces
- Self-correction and improvement
- Memory and context management
"""
import logging
from enum import Enum
from typing import Dict, Any, List, Optional, Callable, Tuple
from dataclasses import dataclass, field
from datetime import datetime
import json
import re
logger = logging.getLogger(__name__)
# LLM for reasoning
try:
from core.llm import chat as llm_chat
LLM_AVAILABLE = True
except ImportError:
LLM_AVAILABLE = False
def llm_chat(*args, **kwargs):
return "LLM not available"
# =============================================================================
# AGENT TYPES
# =============================================================================
class AgentType(Enum):
"""Types of deep agents."""
REACT = "react" # Reason + Act loop
PLAN_EXECUTE = "plan_execute" # Plan then execute
REFLEXION = "reflexion" # Self-reflection
HYBRID = "hybrid" # Combined strategies
@dataclass
class AgentTool:
"""Definition of a tool the agent can use."""
name: str
description: str
function: Callable
params: List[str] = field(default_factory=list)
def to_prompt_text(self) -> str:
"""Format tool for LLM prompt."""
params_str = ", ".join(self.params) if self.params else "none"
return f"- {self.name}: {self.description} (params: {params_str})"
@dataclass
class ThoughtStep:
"""A step in the agent's reasoning chain."""
thought: str
action: Optional[str] = None
action_input: Optional[str] = None
observation: Optional[str] = None
@dataclass
class AgentResult:
"""Result from agent execution."""
answer: str
thought_chain: List[ThoughtStep]
tools_used: List[str]
agent_type: AgentType
success: bool = True
reflection: Optional[str] = None
# =============================================================================
# BASE AGENT
# =============================================================================
class BaseAgent:
"""Base class for all deep agents."""
def __init__(self, user_id: str):
self.user_id = user_id
self.tools: Dict[str, AgentTool] = {}
self.thought_chain: List[ThoughtStep] = []
self._register_default_tools()
def _register_default_tools(self):
"""Register default tools available to all agents."""
self.register_tool(AgentTool(
name="search_data",
description="Search user's data for information",
function=self._tool_search_data,
params=["query"]
))
self.register_tool(AgentTool(
name="calculate",
description="Perform mathematical calculations",
function=self._tool_calculate,
params=["expression"]
))
self.register_tool(AgentTool(
name="get_statistics",
description="Get statistics for a column (mean, sum, count, etc.)",
function=self._tool_get_statistics,
params=["column", "operation"]
))
self.register_tool(AgentTool(
name="get_unique_values",
description="Get unique values count for a categorical column",
function=self._tool_get_unique,
params=["column"]
))
self.register_tool(AgentTool(
name="ask_ai",
description="Ask AI for general knowledge (not from user data)",
function=self._tool_ask_ai,
params=["question"]
))
def register_tool(self, tool: AgentTool):
"""Register a tool."""
self.tools[tool.name] = tool
def _tool_search_data(self, query: str) -> str:
"""Search user data."""
try:
from core.rag import rag_search
context, sources = rag_search(self.user_id, query, k=3)
return context if context else "No relevant data found."
except Exception as e:
return f"Search error: {str(e)[:100]}"
def _tool_calculate(self, expression: str) -> str:
"""Calculate expression."""
try:
# Safe eval
result = eval(expression, {"__builtins__": {}}, {
"abs": abs, "round": round, "min": min, "max": max, "sum": sum,
"len": len, "float": float, "int": int
})
return str(result)
except Exception as e:
return f"Calculation error: {str(e)[:50]}"
def _tool_get_statistics(self, column: str, operation: str = "mean") -> str:
"""Get column statistics."""
try:
from api.v1.endpoints.charts import get_user_data
df = get_user_data(self.user_id)
if df is None or column not in df.columns:
return f"Column '{column}' not found"
ops = {
"mean": df[column].mean,
"sum": df[column].sum,
"count": df[column].count,
"min": df[column].min,
"max": df[column].max,
"std": df[column].std
}
if operation in ops:
result = ops[operation]()
return f"{operation}({column}) = {result:,.2f}" if isinstance(result, float) else str(result)
return f"Unknown operation: {operation}"
except Exception as e:
return f"Statistics error: {str(e)[:100]}"
def _tool_get_unique(self, column: str) -> str:
"""Get unique values count."""
try:
from api.v1.endpoints.charts import get_user_data
df = get_user_data(self.user_id)
if df is None or column not in df.columns:
return f"Column '{column}' not found"
unique_count = df[column].nunique()
top_values = df[column].value_counts().head(5).to_dict()
return f"Unique {column}: {unique_count}. Top values: {top_values}"
except Exception as e:
return f"Error: {str(e)[:100]}"
def _tool_ask_ai(self, question: str) -> str:
"""Ask AI general knowledge."""
prompt = f"""Answer this general knowledge question briefly:
{question}
Be concise (2-3 sentences max). Indicate this is general AI knowledge, not from user data."""
return llm_chat(prompt, temperature=0.3, max_tokens=150)
def execute_tool(self, tool_name: str, tool_input: str) -> str:
"""Execute a tool by name."""
if tool_name not in self.tools:
return f"Unknown tool: {tool_name}"
tool = self.tools[tool_name]
try:
# Parse input as dict or single value
if tool_input.startswith("{"):
try:
kwargs = json.loads(tool_input)
return tool.function(**kwargs)
except:
pass
return tool.function(tool_input)
except Exception as e:
return f"Tool error: {str(e)[:100]}"
def get_tools_prompt(self) -> str:
"""Get formatted tools list for prompt."""
return "\n".join([t.to_prompt_text() for t in self.tools.values()])
# =============================================================================
# REACT AGENT (Reason + Act)
# =============================================================================
class ReActAgent(BaseAgent):
"""
ReAct Agent: Reason + Act loop.
Pattern: Thought → Action → Observation → Thought → ...
Inspired by the ReAct paper (Yao et al., 2022)
"""
def process(self, query: str, max_steps: int = 5) -> AgentResult:
"""Run ReAct loop."""
self.thought_chain = []
tools_used = []
system_prompt = f"""You are a ReAct agent. Solve problems by alternating between:
- THOUGHT: Reason about what to do next
- ACTION: Use a tool to get information
- OBSERVATION: Process the tool result
AVAILABLE TOOLS:
{self.get_tools_prompt()}
FORMAT YOUR RESPONSE EXACTLY AS:
THOUGHT: [your reasoning]
ACTION: [tool_name]
INPUT: [tool input]
When you have the final answer, respond:
THOUGHT: I now have enough information
ANSWER: [your final answer]
USER QUESTION: {query}"""
conversation = [{"role": "system", "content": system_prompt}]
for step in range(max_steps):
# Get agent's next thought/action
response = llm_chat(
messages=conversation,
temperature=0.2,
max_tokens=300
)
# Parse response
thought_match = re.search(r'THOUGHT:\s*(.+?)(?=ACTION:|ANSWER:|$)', response, re.DOTALL)
action_match = re.search(r'ACTION:\s*(\w+)', response)
input_match = re.search(r'INPUT:\s*(.+?)(?=THOUGHT:|$)', response, re.DOTALL)
answer_match = re.search(r'ANSWER:\s*(.+)', response, re.DOTALL)
thought = thought_match.group(1).strip() if thought_match else ""
# Check for final answer
if answer_match:
final_answer = answer_match.group(1).strip()
self.thought_chain.append(ThoughtStep(
thought=thought,
action="FINAL",
observation=final_answer
))
return AgentResult(
answer=final_answer,
thought_chain=self.thought_chain,
tools_used=tools_used,
agent_type=AgentType.REACT,
success=True
)
# Execute action
if action_match:
action = action_match.group(1).strip()
action_input = input_match.group(1).strip() if input_match else ""
observation = self.execute_tool(action, action_input)
tools_used.append(action)
self.thought_chain.append(ThoughtStep(
thought=thought,
action=action,
action_input=action_input,
observation=observation
))
# Add observation to conversation
conversation.append({"role": "assistant", "content": response})
conversation.append({"role": "user", "content": f"OBSERVATION: {observation}"})
else:
# No action parsed, try to extract answer
self.thought_chain.append(ThoughtStep(thought=thought or response))
break
# Max steps reached - generate final answer
final_prompt = f"""Based on the observations, provide a final answer to: {query}
Be direct and concise."""
final_answer = llm_chat(final_prompt, temperature=0.3, max_tokens=300)
return AgentResult(
answer=final_answer,
thought_chain=self.thought_chain,
tools_used=tools_used,
agent_type=AgentType.REACT,
success=True
)
# =============================================================================
# PLAN AND EXECUTE AGENT
# =============================================================================
class PlanAndExecuteAgent(BaseAgent):
"""
Plan-and-Execute Agent: Creates plan first, then executes.
Pattern: Plan → Execute Step 1 → ... → Execute Step N → Synthesize
"""
def create_plan(self, query: str) -> List[Dict[str, str]]:
"""Create execution plan."""
prompt = f"""Create a step-by-step plan to answer this question:
QUESTION: {query}
AVAILABLE TOOLS:
{self.get_tools_prompt()}
Create a plan with 2-4 steps. Format:
STEP 1: [action description] | TOOL: [tool_name] | INPUT: [input]
STEP 2: [action description] | TOOL: [tool_name] | INPUT: [input]
...
Keep it simple and focused."""
response = llm_chat(prompt, temperature=0.2, max_tokens=300)
# Parse plan
plan = []
for line in response.split('\n'):
if 'STEP' in line.upper() and 'TOOL:' in line:
try:
parts = line.split('|')
step_desc = parts[0].split(':')[1].strip() if ':' in parts[0] else parts[0].strip()
tool = parts[1].split(':')[1].strip().lower() if len(parts) > 1 else ""
tool_input = parts[2].split(':')[1].strip() if len(parts) > 2 else ""
if tool in self.tools:
plan.append({
"description": step_desc,
"tool": tool,
"input": tool_input
})
except:
pass
# Default plan if parsing fails
if not plan:
plan = [{"description": "Search data", "tool": "search_data", "input": query}]
return plan
def process(self, query: str) -> AgentResult:
"""Run plan-and-execute."""
self.thought_chain = []
tools_used = []
# Create plan
plan = self.create_plan(query)
self.thought_chain.append(ThoughtStep(
thought=f"Created plan with {len(plan)} steps"
))
# Execute plan
observations = []
for i, step in enumerate(plan):
result = self.execute_tool(step["tool"], step["input"])
observations.append(f"Step {i+1} ({step['tool']}): {result}")
tools_used.append(step["tool"])
self.thought_chain.append(ThoughtStep(
thought=step["description"],
action=step["tool"],
action_input=step["input"],
observation=result
))
# Synthesize final answer
synthesis_prompt = f"""Based on these observations, answer the question.
QUESTION: {query}
OBSERVATIONS:
{chr(10).join(observations)}
Provide a clear, direct answer synthesizing all the information."""
final_answer = llm_chat(synthesis_prompt, temperature=0.3, max_tokens=400)
return AgentResult(
answer=final_answer,
thought_chain=self.thought_chain,
tools_used=tools_used,
agent_type=AgentType.PLAN_EXECUTE,
success=True
)
# =============================================================================
# REFLEXION AGENT
# =============================================================================
class ReflexionAgent(BaseAgent):
"""
Reflexion Agent: Self-reflection and improvement.
Pattern: Act → Evaluate → Reflect → Improve → Act again (if needed)
Inspired by Reflexion paper (Shinn et al., 2023)
"""
def evaluate_answer(self, query: str, answer: str, context: str = "") -> Tuple[bool, str]:
"""Evaluate if answer is good enough."""
prompt = f"""Evaluate this answer for quality and accuracy.
QUESTION: {query}
CONTEXT: {context[:300] if context else 'No context'}
ANSWER: {answer}
Evaluation criteria:
1. Does it directly answer the question?
2. Is it based on actual data/evidence?
3. Is it complete and clear?
Respond with:
VERDICT: PASS or FAIL
FEEDBACK: [one line of constructive feedback]"""
response = llm_chat(prompt, temperature=0.2, max_tokens=100)
is_pass = "PASS" in response.upper() and "FAIL" not in response.upper()
return is_pass, response
def reflect(self, query: str, previous_answer: str, feedback: str) -> str:
"""Generate reflection for improvement."""
prompt = f"""Reflect on the previous attempt and how to improve.
QUESTION: {query}
PREVIOUS ANSWER: {previous_answer}
FEEDBACK: {feedback}
What should be done differently? Provide a brief reflection (2-3 sentences)."""
return llm_chat(prompt, temperature=0.3, max_tokens=150)
def process(self, query: str, max_attempts: int = 2) -> AgentResult:
"""Run reflexion loop."""
self.thought_chain = []
tools_used = []
# Initial attempt using ReAct
react_agent = ReActAgent(self.user_id)
react_agent.tools = self.tools
for attempt in range(max_attempts):
# Execute
result = react_agent.process(query)
tools_used.extend(result.tools_used)
self.thought_chain.append(ThoughtStep(
thought=f"Attempt {attempt + 1}",
observation=result.answer[:200]
))
# Evaluate
is_good, feedback = self.evaluate_answer(query, result.answer)
if is_good:
return AgentResult(
answer=result.answer,
thought_chain=self.thought_chain,
tools_used=tools_used,
agent_type=AgentType.REFLEXION,
success=True,
reflection=f"Passed after {attempt + 1} attempt(s)"
)
# Reflect and improve
if attempt < max_attempts - 1:
reflection = self.reflect(query, result.answer, feedback)
self.thought_chain.append(ThoughtStep(
thought=f"Reflection: {reflection}"
))
# Return best attempt
return AgentResult(
answer=result.answer,
thought_chain=self.thought_chain,
tools_used=tools_used,
agent_type=AgentType.REFLEXION,
success=True,
reflection="Max attempts reached"
)
# =============================================================================
# HYBRID AGENT (Auto-Select Best Strategy)
# =============================================================================
class HybridAgent(BaseAgent):
"""
Hybrid Agent: Selects best agent strategy based on query.
"""
def __init__(self, user_id: str):
super().__init__(user_id)
self.agents = {
AgentType.REACT: ReActAgent(user_id),
AgentType.PLAN_EXECUTE: PlanAndExecuteAgent(user_id),
AgentType.REFLEXION: ReflexionAgent(user_id)
}
def select_strategy(self, query: str) -> AgentType:
"""Select best agent strategy."""
q_lower = query.lower()
# Reflexion for high-stakes queries
if any(kw in q_lower for kw in ['important', 'accurate', 'critical', 'exact', 'precise']):
return AgentType.REFLEXION
# Plan-and-Execute for multi-step queries
if any(kw in q_lower for kw in ['and then', 'first', 'next', 'step by step', 'multiple']):
return AgentType.PLAN_EXECUTE
# Default to ReAct
return AgentType.REACT
def process(self, query: str) -> AgentResult:
"""Run with auto-selected strategy."""
strategy = self.select_strategy(query)
logger.info(f"🤖 Hybrid Agent selected: {strategy.value}")
agent = self.agents.get(strategy, self.agents[AgentType.REACT])
result = agent.process(query)
# Mark as hybrid
result.thought_chain.insert(0, ThoughtStep(
thought=f"Selected strategy: {strategy.value}"
))
return result
# =============================================================================
# CONVENIENCE FUNCTIONS
# =============================================================================
def deep_agent_query(user_id: str, query: str, agent_type: str = "hybrid") -> Dict[str, Any]:
"""Quick function for deep agent query."""
agents = {
"react": ReActAgent,
"plan": PlanAndExecuteAgent,
"reflexion": ReflexionAgent,
"hybrid": HybridAgent
}
agent_class = agents.get(agent_type.lower(), HybridAgent)
agent = agent_class(user_id)
result = agent.process(query)
return {
"answer": result.answer,
"tools_used": result.tools_used,
"agent_type": result.agent_type.value,
"thought_chain": [
{"thought": t.thought, "action": t.action, "observation": t.observation}
for t in result.thought_chain
],
"reflection": result.reflection
}
# Module exports
__all__ = [
'AgentType',
'AgentTool',
'AgentResult',
'BaseAgent',
'ReActAgent',
'PlanAndExecuteAgent',
'ReflexionAgent',
'HybridAgent',
'deep_agent_query'
]