339 lines
10 KiB
Python
339 lines
10 KiB
Python
|
"""Research Agent for Opus Orchestrator.
|
||
|
|
|
||
|
|
Enhanced nonfiction agent with live research capabilities.
|
||
|
|
"""
|
||
|
|
|
||
|
|
import os
|
||
|
|
from typing import Any, Optional
|
||
|
|
|
||
|
|
from dotenv import load_dotenv
|
||
|
|
|
||
|
|
|
||
|
|
from opus_orchestrator.agents.base import BaseAgent, AgentResponse
|
||
|
|
from opus_orchestrator.utils.research import (
|
||
|
|
ResearchOrchestrator,
|
||
|
|
create_research_orchestrator,
|
||
|
|
SearchTool,
|
||
|
|
WikipediaTool,
|
||
|
|
AcademicSearchTool,
|
||
|
|
)
|
||
|
|
|
||
|
|
|
||
|
|
# System prompt for research agent
|
||
|
|
RESEARCH_AGENT_SYSTEM_PROMPT = """## Role: Research Agent with Live Web Access
|
||
|
|
|
||
|
|
You are The Researcher — an AI agent with live access to the internet, academic databases, and research tools.
|
||
|
|
|
||
|
|
## Your Capabilities
|
||
|
|
|
||
|
|
1. **Web Search** - Search the current web for latest information
|
||
|
|
2. **Wikipedia** - Access encyclopedic knowledge
|
||
|
|
3. **Academic Search** - Find peer-reviewed papers (CrossRef, Semantic Scholar)
|
||
|
|
4. **Innovation Detection** - Identify gaps and new ideas beyond training data
|
||
|
|
|
||
|
|
## Your Mission
|
||
|
|
|
||
|
|
NOT just verify facts — **DISCOVER new information, trends, and innovations**.
|
||
|
|
|
||
|
|
- Find what's NEW since your training cutoff
|
||
|
|
- Identify research gaps and opportunities
|
||
|
|
- Connect disparate ideas into novel insights
|
||
|
|
- Go beyond what you "know" to what you can FIND
|
||
|
|
|
||
|
|
## Research Process
|
||
|
|
|
||
|
|
1. **Explore** - Broad search on topic
|
||
|
|
2. **Deep Dive** - Specific searches on subtopics
|
||
|
|
3. **Cross-Reference** - Find connections between sources
|
||
|
|
4. **Innovate** - Generate original insights beyond training data
|
||
|
|
|
||
|
|
## Output Format
|
||
|
|
|
||
|
|
Provide your research in this structure:
|
||
|
|
|
||
|
|
```
|
||
|
|
## Findings (What you discovered)
|
||
|
|
- [New information 1]
|
||
|
|
- [New information 2]
|
||
|
|
- [Latest developments]
|
||
|
|
|
||
|
|
## Sources (Where you found it)
|
||
|
|
- [URL 1]: [Title]
|
||
|
|
- [URL 2]: [Title]
|
||
|
|
|
||
|
|
## Innovations (Original insights beyond training data)
|
||
|
|
- [Novel connection 1]
|
||
|
|
- [Novel connection 2]
|
||
|
|
|
||
|
|
## Research Gaps (What's not well-covered)
|
||
|
|
- [Gap 1]
|
||
|
|
- [Gap 2]
|
||
|
|
```
|
||
|
|
|
||
|
|
## Remember
|
||
|
|
|
||
|
|
You're not just fact-checking — you're RESEARCHING. Actively seek new information,
|
||
|
|
challenge assumptions, and generate original ideas. This keeps the content fresh
|
||
|
|
and prevents "AI slop" from repetitive training data patterns.
|
||
|
|
"""
|
||
|
|
|
||
|
|
|
||
|
|
class ResearchAgent(BaseAgent):
|
||
|
|
"""Enhanced research agent with live web access and innovation detection."""
|
||
|
|
|
||
|
|
def __init__(
|
||
|
|
self,
|
||
|
|
config=None,
|
||
|
|
search_provider: str = "tavily",
|
||
|
|
use_wikipedia: bool = True,
|
||
|
|
use_academic: bool = True,
|
||
|
|
):
|
||
|
|
"""Initialize research agent with tools.
|
||
|
|
|
||
|
|
Args:
|
||
|
|
config: Agent configuration
|
||
|
|
search_provider: Search provider (tavily, serper, brave, duckduckgo)
|
||
|
|
use_wikipedia: Include Wikipedia search
|
||
|
|
use_academic: Include academic search
|
||
|
|
"""
|
||
|
|
# Initialize research tools
|
||
|
|
self.research = create_research_orchestrator(
|
||
|
|
search_provider=search_provider,
|
||
|
|
use_wikipedia=use_wikipedia,
|
||
|
|
use_academic=use_academic,
|
||
|
|
)
|
||
|
|
|
||
|
|
self.search_tool = SearchTool(provider=search_provider)
|
||
|
|
self.wikipedia = WikipediaTool() if use_wikipedia else None
|
||
|
|
self.academic = AcademicSearchTool() if use_academic else None
|
||
|
|
|
||
|
|
super().__init__(
|
||
|
|
role="Research Agent",
|
||
|
|
description="Live web research with innovation detection",
|
||
|
|
system_prompt=RESEARCH_AGENT_SYSTEM_PROMPT,
|
||
|
|
config=config,
|
||
|
|
)
|
||
|
|
|
||
|
|
async def execute(self, input_data: Any, context: dict[str, Any]) -> AgentResponse:
|
||
|
|
"""Execute research task with live tools.
|
||
|
|
|
||
|
|
Args:
|
||
|
|
input_data: Research query and parameters
|
||
|
|
context: Additional context
|
||
|
|
|
||
|
|
Returns:
|
||
|
|
Research findings with sources and innovations
|
||
|
|
"""
|
||
|
|
# Extract query
|
||
|
|
if isinstance(input_data, dict):
|
||
|
|
query = input_data.get("query", "")
|
||
|
|
subtopics = input_data.get("subtopics", [])
|
||
|
|
deep = input_data.get("deep_research", False)
|
||
|
|
else:
|
||
|
|
query = str(input_data)
|
||
|
|
subtopics = []
|
||
|
|
deep = False
|
||
|
|
|
||
|
|
if not query:
|
||
|
|
return AgentResponse(
|
||
|
|
success=False,
|
||
|
|
output=None,
|
||
|
|
error="No research query provided",
|
||
|
|
metadata={"role": "Research Agent"},
|
||
|
|
)
|
||
|
|
|
||
|
|
try:
|
||
|
|
# Perform research
|
||
|
|
if deep or subtopics:
|
||
|
|
# Deep research with subtopics
|
||
|
|
results = self.research.deep_research(query, subtopics)
|
||
|
|
else:
|
||
|
|
# Quick comprehensive search
|
||
|
|
results = self.research.comprehensive_search(query)
|
||
|
|
|
||
|
|
# Format results for LLM
|
||
|
|
research_summary = self._format_research_for_llm(results)
|
||
|
|
|
||
|
|
# Use LLM to synthesize and provide analysis
|
||
|
|
synthesis = await self.call_llm(
|
||
|
|
system_prompt=self.build_system_prompt(context),
|
||
|
|
user_prompt=f"""Based on this research data, provide analysis and insights:
|
||
|
|
|
||
|
|
{research_summary}
|
||
|
|
|
||
|
|
Task: {query}
|
||
|
|
|
||
|
|
Provide:
|
||
|
|
1. Key findings synthesized
|
||
|
|
2. Most important innovations/discoveries
|
||
|
|
3. How this goes beyond typical training data
|
||
|
|
4. Recommendations for the manuscript""",
|
||
|
|
)
|
||
|
|
|
||
|
|
return AgentResponse(
|
||
|
|
success=True,
|
||
|
|
output={
|
||
|
|
"raw_results": results,
|
||
|
|
"synthesis": synthesis,
|
||
|
|
"query": query,
|
||
|
|
},
|
||
|
|
metadata={
|
||
|
|
"role": "Research Agent",
|
||
|
|
"search_provider": self.research.search.provider,
|
||
|
|
},
|
||
|
|
)
|
||
|
|
|
||
|
|
except Exception as e:
|
||
|
|
return AgentResponse(
|
||
|
|
success=False,
|
||
|
|
output=None,
|
||
|
|
error=f"Research failed: {str(e)}",
|
||
|
|
metadata={"role": "Research Agent"},
|
||
|
|
)
|
||
|
|
|
||
|
|
def _format_research_for_llm(self, results: dict) -> str:
|
||
|
|
"""Format research results for LLM consumption."""
|
||
|
|
output = []
|
||
|
|
|
||
|
|
# Query
|
||
|
|
output.append(f"# Research Query: {results.get('query', '')}")
|
||
|
|
output.append(f"Timestamp: {results.get('timestamp', '')}")
|
||
|
|
output.append("")
|
||
|
|
|
||
|
|
# Web results
|
||
|
|
web = results.get("web", [])
|
||
|
|
if web:
|
||
|
|
output.append("## Web Search Results")
|
||
|
|
for i, r in enumerate(web[:5], 1):
|
||
|
|
output.append(f"{i}. **{r.get('title', '')}**")
|
||
|
|
output.append(f" URL: {r.get('url', '')}")
|
||
|
|
output.append(f" {r.get('content', '')[:200]}...")
|
||
|
|
output.append("")
|
||
|
|
|
||
|
|
# Wikipedia
|
||
|
|
wiki = results.get("wikipedia", [])
|
||
|
|
if wiki:
|
||
|
|
output.append("## Wikipedia Results")
|
||
|
|
for r in wiki[:3]:
|
||
|
|
output.append(f"- {r.get('title', '')}: {r.get('summary', '')[:200]}...")
|
||
|
|
output.append("")
|
||
|
|
|
||
|
|
# Academic
|
||
|
|
academic = results.get("academic", [])
|
||
|
|
if academic:
|
||
|
|
output.append("## Academic Papers")
|
||
|
|
for r in academic[:5]:
|
||
|
|
output.append(f"- {r.get('title', '')} ({r.get('year', 'N/A')})")
|
||
|
|
output.append(f" {r.get('journal', '')}")
|
||
|
|
output.append("")
|
||
|
|
|
||
|
|
# Innovations
|
||
|
|
innovations = results.get("innovations", [])
|
||
|
|
if innovations:
|
||
|
|
output.append("## Innovations & New Ideas")
|
||
|
|
for i in innovations:
|
||
|
|
output.append(f"- {i}")
|
||
|
|
output.append("")
|
||
|
|
|
||
|
|
return "\n".join(output)
|
||
|
|
|
||
|
|
|
||
|
|
# Fact-checking with live verification
|
||
|
|
class VerifiedFactChecker:
|
||
|
|
"""Fact checker with live source verification."""
|
||
|
|
|
||
|
|
def __init__(self, search_provider: str = "tavily"):
|
||
|
|
"""Initialize verified fact checker."""
|
||
|
|
self.search = SearchTool(provider=search_provider)
|
||
|
|
self.wikipedia = WikipediaTool()
|
||
|
|
|
||
|
|
async def verify_claim(
|
||
|
|
self,
|
||
|
|
claim: str,
|
||
|
|
context: str = "",
|
||
|
|
) -> dict:
|
||
|
|
"""Verify a factual claim against live sources.
|
||
|
|
|
||
|
|
Args:
|
||
|
|
claim: The claim to verify
|
||
|
|
context: Additional context
|
||
|
|
|
||
|
|
Returns:
|
||
|
|
Verification result with confidence and sources
|
||
|
|
"""
|
||
|
|
# Search for the claim
|
||
|
|
results = self.search.search(claim, num_results=5)
|
||
|
|
|
||
|
|
# Check Wikipedia
|
||
|
|
wiki_results = self.wikipedia.search(claim, num_results=2)
|
||
|
|
|
||
|
|
# Analyze
|
||
|
|
supporting = []
|
||
|
|
contradicting = []
|
||
|
|
neutral = []
|
||
|
|
|
||
|
|
for r in results:
|
||
|
|
content = r.get("content", "").lower()
|
||
|
|
claim_lower = claim.lower()
|
||
|
|
|
||
|
|
# Simple keyword matching
|
||
|
|
claim_words = set(claim_lower.split())
|
||
|
|
content_words = set(content.split())
|
||
|
|
overlap = claim_words & content_words
|
||
|
|
|
||
|
|
if len(overlap) > len(claim_words) * 0.7:
|
||
|
|
supporting.append(r)
|
||
|
|
elif "not" in content or "false" in content or "incorrect" in content:
|
||
|
|
contradicting.append(r)
|
||
|
|
else:
|
||
|
|
neutral.append(r)
|
||
|
|
|
||
|
|
# Calculate confidence
|
||
|
|
total = len(supporting) + len(contradicting) + len(neutral)
|
||
|
|
if total == 0:
|
||
|
|
confidence = 0.0
|
||
|
|
else:
|
||
|
|
confidence = len(supporting) / total
|
||
|
|
|
||
|
|
return {
|
||
|
|
"claim": claim,
|
||
|
|
"verified": len(supporting) > 0,
|
||
|
|
"confidence": confidence,
|
||
|
|
"supporting_sources": supporting,
|
||
|
|
"contradicting_sources": contradicting,
|
||
|
|
"neutral_sources": neutral,
|
||
|
|
"needs_citation": confidence < 0.8,
|
||
|
|
}
|
||
|
|
|
||
|
|
async def verify_batch(
|
||
|
|
self,
|
||
|
|
claims: list[str],
|
||
|
|
) -> list[dict]:
|
||
|
|
"""Verify multiple claims.
|
||
|
|
|
||
|
|
Args:
|
||
|
|
claims: List of claims to verify
|
||
|
|
|
||
|
|
Returns:
|
||
|
|
List of verification results
|
||
|
|
"""
|
||
|
|
results = []
|
||
|
|
for claim in claims:
|
||
|
|
result = await self.verify_claim(claim)
|
||
|
|
results.append(result)
|
||
|
|
return results
|
||
|
|
|
||
|
|
|
||
|
|
def create_research_agent(
|
||
|
|
search_provider: str = "tavily",
|
||
|
|
) -> ResearchAgent:
|
||
|
|
"""Factory to create a research agent.
|
||
|
|
|
||
|
|
Args:
|
||
|
|
search_provider: Search provider
|
||
|
|
|
||
|
|
Returns:
|
||
|
|
Configured ResearchAgent
|
||
|
|
"""
|
||
|
|
return ResearchAgent(search_provider=search_provider)
|