chat/summarizer.py

# chat/summarizer.py
"""

Advanced multi-document summarization with cross-paper synthesis

Identifies consensus, contradictions, and trends across multiple papers

"""

from typing import List, Dict, Any, Tuple
from llm.llm_provider import GrokLLM
from llm.prompt_templates import MedicalResearchPrompts, ResponseFormatter
import re


class MultiDocumentSummarizer:
    """

    Advanced summarizer that synthesizes information across multiple research papers

    Goes beyond simple aggregation to identify patterns and insights

    """

    def __init__(self, llm=None):
        self.llm = llm or GrokLLM(model="model")  # Use shared LLM
        self.prompts = MedicalResearchPrompts()
        self.formatter = ResponseFormatter()

    def summarize_research(self, papers: List[Dict], query: str, domain: str) -> Dict[str, Any]:
        """

        Generate comprehensive multi-document summary

        Returns consistent dict format that other components expect

        """
        print(f"📚 Summarizing {len(papers)} papers for query: '{query}'")

        # Filter relevant papers if we have too many
        if len(papers) > 15:
            print(f"🔄 Filtering {len(papers)} papers to top 15 most relevant")
            papers = papers[:15]

        try:
            # Generate comprehensive summary
            summary_prompt = self.prompts.multi_document_summary(papers, query, domain)

            response = self.llm.generate(
                summary_prompt,
                system_message=self.prompts.SYSTEM_MESSAGES["research_analyst"],
                temperature=0.1,
                max_tokens=4000
            )

            # Extract key insights programmatically
            key_insights = self._extract_key_insights(response, papers)
            consensus_analysis = self._analyze_consensus(papers)

            # Return consistent format that RAG engine expects
            return {
                "summary": response,  # Main summary field
                "comprehensive_summary": response,  # Alternative field
                "key_insights": key_insights,
                "consensus_analysis": consensus_analysis,
                "papers_analyzed": len(papers),
                "papers_cited": [self.formatter.format_citation(paper, i + 1) for i, paper in enumerate(papers)],
                "query": query,
                "domain": domain
            }

        except Exception as e:
            print(f"❌ Summarization error: {e}")
            return self._create_fallback_summary(papers, query, domain)

    def _extract_key_insights(self, summary: str, papers: List[Dict]) -> List[Dict[str, str]]:
        """Extract structured key insights from summary text"""
        insights = []

        # Look for key findings patterns in the summary
        patterns = [
            r"\d+\.\s*\*\*([^*]+)\*\*[^\n]*\n([^*]+?(?=\n\d+\.|\n\*\*|$))",
            r"-?\s*\*\*([^*]+)\*\*[^\n]*\n([^*]+?(?=\n-|\n\*\*|$))",
            r"([A-Z][^.:!?]+:[^.:!?]+)[.:!?]\s*([^.:!?]+[.:!?])"
        ]

        for pattern in patterns:
            matches = re.findall(pattern, summary, re.IGNORECASE | re.DOTALL)
            for match in matches:
                if len(match) == 2:
                    insight_type, insight_text = match
                    insights.append({
                        "type": insight_type.strip(),
                        "description": insight_text.strip(),
                        "confidence": "high" if any(keyword in insight_type.lower() for keyword in
                                                    ["consensus", "proven", "established"]) else "medium"
                    })

        # If no structured insights found, create from summary sentences
        if not insights:
            sentences = re.split(r'[.!?]+', summary)
            for sentence in sentences[:5]:  # Top 5 sentences
                sentence = sentence.strip()
                if len(sentence) > 50 and any(
                        keyword in sentence.lower() for keyword in ["find", "show", "demonstrate", "conclude"]):
                    insights.append({
                        "type": "Key Finding",
                        "description": sentence,
                        "confidence": "medium"
                    })

        return insights[:10]  # Limit to top 10 insights

    def _analyze_consensus(self, papers: List[Dict]) -> Dict[str, Any]:
        """Analyze consensus and contradictions across papers"""
        if len(papers) < 2:
            return {"consensus_level": "insufficient_data", "contradictions": []}

        # Simple consensus analysis based on abstract content
        consensus_keywords = {}
        all_keywords = set()

        for paper in papers:
            abstract = paper.get('abstract', '').lower()
            # Extract meaningful keywords (simplified)
            words = re.findall(r'\b[a-z]{5,15}\b', abstract)
            for word in words:
                if word not in ['which', 'their', 'about', 'using', 'method', 'study']:
                    consensus_keywords[word] = consensus_keywords.get(word, 0) + 1
                    all_keywords.add(word)

        # Find common keywords (appearing in multiple papers)
        common_keywords = {word: count for word, count in consensus_keywords.items()
                           if count >= max(2, len(papers) * 0.3)}  # At least 30% of papers

        # Analyze potential contradictions (simplified)
        contradictions = []
        if len(papers) >= 3:
            # Look for papers with opposing conclusions (simplified heuristic)
            positive_terms = ['improve', 'better', 'superior', 'effective', 'accurate']
            negative_terms = ['limit', 'challenge', 'worse', 'ineffective', 'inaccurate']

            for i, paper1 in enumerate(papers):
                abstract1 = paper1.get('abstract', '').lower()
                paper1_pos = sum(1 for term in positive_terms if term in abstract1)
                paper1_neg = sum(1 for term in negative_terms if term in abstract1)

                for j, paper2 in enumerate(papers[i + 1:], i + 1):
                    abstract2 = paper2.get('abstract', '').lower()
                    paper2_pos = sum(1 for term in positive_terms if term in abstract2)
                    paper2_neg = sum(1 for term in negative_terms if term in abstract2)

                    # Simple contradiction detection
                    if (paper1_pos > paper2_pos + 2 and paper2_neg > paper1_neg + 2) or \
                            (paper2_pos > paper1_pos + 2 and paper1_neg > paper2_neg + 2):
                        contradictions.append({
                            "paper1": paper1.get('title', f'Paper {i + 1}'),
                            "paper2": paper2.get('title', f'Paper {j + 1}'),
                            "nature": "differing_conclusions"
                        })

        return {
            "consensus_level": "high" if len(common_keywords) > 10 else "medium" if len(common_keywords) > 5 else "low",
            "common_themes": list(common_keywords.keys())[:10],
            "contradictions": contradictions[:5],
            "papers_analyzed": len(papers)
        }

    def _create_fallback_summary(self, papers: List[Dict], query: str, domain: str) -> Dict[str, Any]:
        """Create a basic fallback summary when LLM fails"""
        print("🔄 Using fallback summary method")

        # Extract basic information from papers
        titles = [paper.get('title', 'Unknown') for paper in papers]
        sources = list(set(paper.get('source', 'Unknown') for paper in papers))
        domains = list(set(paper.get('domain', 'Unknown') for paper in papers))

        basic_summary = f"""

        Research Summary for: {query}

        Domain: {domain}



        Analyzed {len(papers)} papers from sources including: {', '.join(sources[:3])}



        Key Papers:

        {chr(10).join(f'- {title}' for title in titles[:5])}



        Note: Detailed analysis unavailable due to technical limitations.

        Please try refining your search query or using fewer papers.

        """

        return {
            "summary": basic_summary,  # Main field
            "comprehensive_summary": basic_summary,  # Alternative field
            "key_insights": [{"type": "Technical Note", "description": "Basic summary only - LLM analysis unavailable",
                              "confidence": "low"}],
            "consensus_analysis": {"consensus_level": "unknown", "common_themes": [], "contradictions": []},
            "papers_analyzed": len(papers),
            "papers_cited": [self.formatter.format_citation(paper, i + 1) for i, paper in enumerate(papers)],
            "query": query,
            "domain": domain,
            "fallback_used": True
        }

    def generate_executive_summary(self, summary_data) -> str:
        """

        Generate a concise executive summary from detailed analysis

        Handles both string and dict inputs for compatibility

        """
        try:
            # Handle both string and dict inputs for backward compatibility
            if isinstance(summary_data, str):
                # If we get a string, create a basic executive summary
                prompt = f"Create a concise executive summary (2-3 paragraphs) from this research summary:\n\n{summary_data}"
                return self.llm.generate(
                    prompt,
                    system_message="Create professional executive summaries for research papers",
                    max_tokens=400
                )

            elif isinstance(summary_data, dict):
                # If we get a dict (preferred), use the structured data
                insights = summary_data.get('key_insights', [])
                consensus = summary_data.get('consensus_analysis', {})
                papers_analyzed = summary_data.get('papers_analyzed', 0)
                domain = summary_data.get('domain', 'unknown domain')

                # Build executive summary from structured data
                executive_summary = f"**Executive Summary**\n\n"
                executive_summary += f"Based on analysis of {papers_analyzed} papers in {domain}:\n\n"

                # Add top insights
                if insights:
                    executive_summary += "**Key Findings:**\n"
                    for insight in insights[:3]:
                        executive_summary += f"• {insight.get('description', 'No description')}\n"

                # Add consensus level
                if consensus.get('consensus_level') != 'unknown':
                    executive_summary += f"\n**Consensus Level:** {consensus['consensus_level'].title()}\n"

                # Add common themes if available
                if consensus.get('common_themes'):
                    executive_summary += f"\n**Common Themes:** {', '.join(consensus['common_themes'][:5])}\n"

                return executive_summary
            else:
                return "Executive summary unavailable - invalid input format"

        except Exception as e:
            print(f"❌ Executive summary generation failed: {e}")
            return "Executive summary unavailable due to technical error"


# Quick test
def test_summarizer():
    """Test the multi-document summarizer"""
    print("🧪 Testing Multi-Document Summarizer")
    print("=" * 50)

    test_papers = [
        {
            'title': 'Deep Learning for Medical Image Analysis',
            'authors': ['Smith J', 'Johnson A', 'Brown K'],
            'abstract': 'We demonstrate that convolutional neural networks significantly improve accuracy in medical image segmentation tasks. Our method achieves 95% accuracy on the BRATS dataset, outperforming traditional methods by 15%.',
            'source': 'Nature Medicine',
            'domain': 'medical_imaging',
            'publication_date': '2024-01-15'
        },
        {
            'title': 'Transformers in Radiology',
            'authors': ['Lee K', 'Chen R', 'Wang L'],
            'abstract': 'This study shows that transformer architectures provide better context understanding in radiology images compared to CNNs. However, they require more computational resources and larger datasets for training.',
            'source': 'Radiology AI Journal',
            'domain': 'medical_imaging',
            'publication_date': '2024-02-20'
        }
    ]

    summarizer = MultiDocumentSummarizer()

    try:
        summary = summarizer.summarize_research(
            test_papers,
            "Latest advances in AI for medical imaging",
            "medical_imaging"
        )

        print(f"✅ Summary generated successfully")
        print(f"📊 Papers analyzed: {summary['papers_analyzed']}")
        print(f"💡 Key insights: {len(summary['key_insights'])}")
        print(f"🤝 Consensus level: {summary['consensus_analysis']['consensus_level']}")

        # Test executive summary with dict input
        executive = summarizer.generate_executive_summary(summary)
        print(f"\n📋 Executive Summary:\n{executive}")

        # Test executive summary with string input (backward compatibility)
        executive_str = summarizer.generate_executive_summary(summary['summary'])
        print(f"\n📋 Executive Summary (from string):\n{executive_str}")

    except Exception as e:
        print(f"❌ Summarization test failed: {e}")


if __name__ == "__main__":
    test_summarizer()