local-deep-research 0.4.4__py3-none-any.whl → 0.5.0__py3-none-any.whl

local_deep_research/advanced_search_system/candidate_exploration/progressive_explorer.py

@@ -0,0 +1,255 @@
+"""
+Progressive explorer for BrowseComp-style systematic search exploration.
+"""
+
+import concurrent.futures
+import logging
+from dataclasses import dataclass, field
+from typing import Dict, List, Set, Tuple
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class SearchProgress:
+    """Track search progress and findings."""
+
+    searched_terms: Set[str] = field(default_factory=set)
+    found_candidates: Dict[str, float] = field(
+        default_factory=dict
+    )  # name -> confidence
+    verified_facts: Dict[str, str] = field(
+        default_factory=dict
+    )  # fact -> source
+    entity_coverage: Dict[str, Set[str]] = field(
+        default_factory=dict
+    )  # entity_type -> searched_entities
+    search_depth: int = 0
+
+    def update_coverage(self, entity_type: str, entity: str):
+        """Update entity coverage tracking."""
+        if entity_type not in self.entity_coverage:
+            self.entity_coverage[entity_type] = set()
+        self.entity_coverage[entity_type].add(entity.lower())
+
+    def get_uncovered_entities(
+        self, entities: Dict[str, List[str]]
+    ) -> Dict[str, List[str]]:
+        """Get entities that haven't been searched yet."""
+        uncovered = {}
+        for entity_type, entity_list in entities.items():
+            covered = self.entity_coverage.get(entity_type, set())
+            uncovered_list = [
+                e for e in entity_list if e.lower() not in covered
+            ]
+            if uncovered_list:
+                uncovered[entity_type] = uncovered_list
+        return uncovered
+
+
+class ProgressiveExplorer:
+    """
+    Explorer that implements progressive search strategies for BrowseComp.
+
+    Key features:
+    1. Tracks search progress to avoid redundancy
+    2. Progressively combines entities
+    3. Identifies and pursues promising candidates
+    4. Maintains simple approach without over-filtering
+    """
+
+    def __init__(self, search_engine, model):
+        self.search_engine = search_engine
+        self.model = model
+        self.progress = SearchProgress()
+        self.max_results_per_search = 20  # Keep more results
+
+    def explore(
+        self,
+        queries: List[str],
+        constraints: List = None,
+        max_workers: int = 5,
+        extracted_entities: Dict[str, List[str]] = None,
+    ) -> Tuple[List, SearchProgress]:
+        """
+        Execute progressive exploration with entity tracking.
+
+        Returns both candidates and search progress for strategy use.
+        """
+        all_results = []
+        extracted_entities = extracted_entities or {}
+
+        # Execute searches in parallel (like source-based strategy)
+        search_results = self._parallel_search(queries, max_workers)
+
+        # Process results without filtering (trust the LLM later)
+        for query, results in search_results:
+            self.progress.searched_terms.add(query.lower())
+
+            # Track which entities were covered in this search
+            self._update_entity_coverage(query, extracted_entities)
+
+            # Extract any specific names/candidates from results
+            candidates = self._extract_candidates_from_results(results, query)
+            for candidate_name, confidence in candidates.items():
+                if candidate_name in self.progress.found_candidates:
+                    # Update confidence if higher
+                    self.progress.found_candidates[candidate_name] = max(
+                        self.progress.found_candidates[candidate_name],
+                        confidence,
+                    )
+                else:
+                    self.progress.found_candidates[candidate_name] = confidence
+
+            # Keep all results for final synthesis
+            all_results.extend(results)
+
+        self.progress.search_depth += 1
+
+        # Return both results and progress
+        return all_results, self.progress
+
+    def generate_verification_searches(
+        self,
+        candidates: Dict[str, float],
+        constraints: List,
+        max_searches: int = 5,
+    ) -> List[str]:
+        """Generate targeted searches to verify top candidates."""
+        if not candidates:
+            return []
+
+        # Get top candidates by confidence
+        top_candidates = sorted(
+            candidates.items(), key=lambda x: x[1], reverse=True
+        )[:3]
+
+        verification_searches = []
+        for candidate_name, confidence in top_candidates:
+            # Generate verification searches for this candidate
+            for constraint in constraints[:2]:  # Verify top constraints
+                search = f'"{candidate_name}" {constraint.description}'
+                if search.lower() not in self.progress.searched_terms:
+                    verification_searches.append(search)
+
+        return verification_searches[:max_searches]
+
+    def _extract_candidates_from_results(
+        self, results: List[Dict], query: str
+    ) -> Dict[str, float]:
+        """Extract potential answer candidates from search results."""
+        candidates = {}
+
+        # Simple extraction based on titles and snippets
+        for result in results[:10]:  # Focus on top results
+            title = result.get("title", "")
+            snippet = result.get("snippet", "")
+
+            # Look for proper nouns and specific names
+            # This is simplified - in practice, might use NER or more sophisticated extraction
+            combined_text = f"{title} {snippet}"
+
+            # Extract quoted terms as potential candidates
+            import re
+
+            quoted_terms = re.findall(r'"([^"]+)"', combined_text)
+            for term in quoted_terms:
+                if (
+                    len(term) > 2 and len(term) < 50
+                ):  # Reasonable length for an answer
+                    candidates[term] = 0.3  # Base confidence from appearance
+
+            # Boost confidence if appears in title
+            if title:
+                # Titles often contain the actual answer
+                title_words = title.split()
+                for i in range(len(title_words)):
+                    for j in range(i + 1, min(i + 4, len(title_words) + 1)):
+                        phrase = " ".join(title_words[i:j])
+                        if (
+                            len(phrase) > 3 and phrase[0].isupper()
+                        ):  # Likely proper noun
+                            candidates[phrase] = candidates.get(phrase, 0) + 0.2
+
+        return candidates
+
+    def _update_entity_coverage(
+        self, query: str, entities: Dict[str, List[str]]
+    ):
+        """Track which entities have been covered in searches."""
+        query_lower = query.lower()
+
+        for entity_type, entity_list in entities.items():
+            for entity in entity_list:
+                if entity.lower() in query_lower:
+                    self.progress.update_coverage(entity_type, entity)
+
+    def suggest_next_searches(
+        self, entities: Dict[str, List[str]], max_suggestions: int = 5
+    ) -> List[str]:
+        """Suggest next searches based on coverage and findings."""
+        suggestions = []
+
+        # 1. Check uncovered entities
+        uncovered = self.progress.get_uncovered_entities(entities)
+
+        # 2. If we have candidates, verify them with uncovered constraints
+        if self.progress.found_candidates:
+            top_candidate = max(
+                self.progress.found_candidates.items(), key=lambda x: x[1]
+            )[0]
+
+            # Combine candidate with uncovered entities
+            for entity_type, entity_list in uncovered.items():
+                for entity in entity_list[:2]:
+                    search = f'"{top_candidate}" {entity}'
+                    if search.lower() not in self.progress.searched_terms:
+                        suggestions.append(search)
+
+        # 3. Otherwise, create new combinations of uncovered entities
+        else:
+            # Focus on systematic coverage
+            if uncovered.get("temporal"):
+                # Year-by-year with key term
+                key_term = (
+                    entities.get("names", [""])[0]
+                    or entities.get("descriptors", [""])[0]
+                )
+                for year in uncovered["temporal"][:3]:
+                    search = f"{key_term} {year}".strip()
+                    if search.lower() not in self.progress.searched_terms:
+                        suggestions.append(search)
+
+            if uncovered.get("names") and uncovered.get("descriptors"):
+                # Combine names with descriptors
+                for name in uncovered["names"][:2]:
+                    for desc in uncovered["descriptors"][:2]:
+                        search = f"{name} {desc}"
+                        if search.lower() not in self.progress.searched_terms:
+                            suggestions.append(search)
+
+        return suggestions[:max_suggestions]
+
+    def _parallel_search(
+        self, queries: List[str], max_workers: int
+    ) -> List[Tuple[str, List[Dict]]]:
+        """Execute searches in parallel and return results."""
+        results = []
+
+        def search_query(query):
+            try:
+                search_results = self.search_engine.run(query)
+                return (query, search_results or [])
+            except Exception as e:
+                logger.error(f"Error searching '{query}': {str(e)}")
+                return (query, [])
+
+        # Run searches in parallel
+        with concurrent.futures.ThreadPoolExecutor(
+            max_workers=max_workers
+        ) as executor:
+            futures = [executor.submit(search_query, q) for q in queries]
+            for future in concurrent.futures.as_completed(futures):
+                results.append(future.result())
+
+        return results

local_deep_research/advanced_search_system/candidates/__init__.py

@@ -0,0 +1,5 @@
+# Candidates System Package
+
+from .base_candidate import Candidate
+
+__all__ = ["Candidate"]

local_deep_research/advanced_search_system/candidates/base_candidate.py

@@ -0,0 +1,59 @@
+"""
+Base candidate class for tracking potential answers.
+"""
+
+from dataclasses import dataclass, field
+from typing import Any, Dict, List
+
+from ..constraints.base_constraint import Constraint
+from ..evidence.base_evidence import Evidence
+
+
+@dataclass
+class Candidate:
+    """A potential answer with supporting evidence."""
+
+    name: str
+    evidence: Dict[str, Evidence] = field(default_factory=dict)
+    score: float = 0.0
+    metadata: Dict[str, Any] = field(default_factory=dict)
+
+    def add_evidence(self, constraint_id: str, evidence: Evidence):
+        """Add evidence for a constraint."""
+        self.evidence[constraint_id] = evidence
+
+    def calculate_score(self, constraints: List[Constraint]) -> float:
+        """Calculate overall score based on evidence and constraints."""
+        if not constraints:
+            return 0.0
+
+        total_score = 0.0
+        total_weight = 0.0
+
+        for constraint in constraints:
+            evidence = self.evidence.get(constraint.id)
+            if evidence:
+                score = evidence.confidence * constraint.weight
+                total_score += score
+            total_weight += constraint.weight
+
+        self.score = total_score / total_weight if total_weight > 0 else 0.0
+        return self.score
+
+    def get_unverified_constraints(
+        self, constraints: List[Constraint]
+    ) -> List[Constraint]:
+        """Get constraints that don't have evidence yet."""
+        unverified = []
+        for constraint in constraints:
+            if constraint.id not in self.evidence:
+                unverified.append(constraint)
+        return unverified
+
+    def get_weak_evidence(self, threshold: float = 0.5) -> List[str]:
+        """Get constraint IDs with weak evidence."""
+        weak = []
+        for constraint_id, evidence in self.evidence.items():
+            if evidence.confidence < threshold:
+                weak.append(constraint_id)
+        return weak

local_deep_research/advanced_search_system/constraint_checking/README.md

@@ -0,0 +1,150 @@
+# Constraint Checking System
+
+This module provides an inheritance-based constraint checking system for validating candidates against constraints in the Local Deep Research framework.
+
+## Architecture
+
+The system is built around inheritance and provides multiple implementations:
+
+### Base Class
+- **`BaseConstraintChecker`**: Abstract base class defining the interface
+
+### Concrete Implementations
+- **`DualConfidenceChecker`**: Uses positive/negative/uncertainty confidence scoring
+- **`ThresholdChecker`**: Simple threshold-based checking
+- **`StrictChecker`**: Example of very strict constraint validation
+
+### Supporting Components
+- **`EvidenceAnalyzer`**: Analyzes evidence using dual confidence scoring
+- **`RejectionEngine`**: Makes rejection decisions based on evidence
+- **`ConstraintCheckResult`**: Data class containing evaluation results
+
+## Usage Examples
+
+### Using DualConfidenceChecker
+```python
+from constraint_checking import DualConfidenceChecker
+
+checker = DualConfidenceChecker(
+    model=llm,
+    evidence_gatherer=evidence_function,
+    negative_threshold=0.25,  # Reject if negative evidence > 25%
+    positive_threshold=0.4,   # Reject if positive evidence < 40%
+)
+
+result = checker.check_candidate(candidate, constraints)
+```
+
+### Using ThresholdChecker
+```python
+from constraint_checking import ThresholdChecker
+
+checker = ThresholdChecker(
+    model=llm,
+    evidence_gatherer=evidence_function,
+    satisfaction_threshold=0.7,     # Individual constraint threshold
+    required_satisfaction_rate=0.8  # Overall satisfaction rate needed
+)
+
+result = checker.check_candidate(candidate, constraints)
+```
+
+### Using StrictChecker
+```python
+from constraint_checking import StrictChecker
+
+checker = StrictChecker(
+    model=llm,
+    evidence_gatherer=evidence_function,
+    strict_threshold=0.9,        # Very high threshold
+    name_pattern_required=True   # NAME_PATTERN constraints are mandatory
+)
+
+result = checker.check_candidate(candidate, constraints)
+```
+
+## Creating Custom Variants
+
+To create your own constraint checker variant:
+
+1. **Inherit from BaseConstraintChecker**:
+```python
+from .base_constraint_checker import BaseConstraintChecker, ConstraintCheckResult
+
+class MyCustomChecker(BaseConstraintChecker):
+    def __init__(self, *args, my_param=0.5, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.my_param = my_param
+```
+
+2. **Implement required methods**:
+```python
+    def check_candidate(self, candidate, constraints):
+        # Your implementation here
+        return ConstraintCheckResult(...)
+
+    def should_reject_candidate(self, candidate, constraint, evidence_data):
+        # Your rejection logic here
+        return should_reject, reason
+```
+
+3. **Add custom logic**:
+```python
+    def _my_custom_evaluation(self, candidate, constraint):
+        # Your custom evaluation logic
+        pass
+```
+
+## Integration with Strategies
+
+Use in your strategy by initializing the checker:
+
+```python
+class MyStrategy(BaseStrategy):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # Choose your constraint checker
+        self.constraint_checker = DualConfidenceChecker(
+            model=self.model,
+            evidence_gatherer=self._gather_evidence_for_constraint,
+            # ... other parameters
+        )
+
+    def _evaluate_candidate(self, candidate):
+        result = self.constraint_checker.check_candidate(candidate, self.constraints)
+
+        # Process result
+        candidate.evaluation_results = result.detailed_results
+        candidate.score = result.total_score
+
+        return result.total_score
+```
+
+## Available Checkers
+
+### DualConfidenceChecker
+- **Best for**: Nuanced evaluation with detailed confidence scoring
+- **Parameters**: `negative_threshold`, `positive_threshold`, `uncertainty_penalty`, `negative_weight`
+- **Output**: Detailed positive/negative/uncertainty scores per constraint
+
+### ThresholdChecker
+- **Best for**: Fast, simple constraint checking
+- **Parameters**: `satisfaction_threshold`, `required_satisfaction_rate`
+- **Output**: Simple satisfied/not satisfied per constraint
+
+### StrictChecker
+- **Best for**: Cases requiring very high confidence
+- **Parameters**: `strict_threshold`, `name_pattern_required`
+- **Output**: Binary pass/fail with strict requirements
+
+## Extending the System
+
+The inheritance-based design makes it easy to:
+
+1. **Create specialized checkers** for specific domains
+2. **Mix and match components** (e.g., use DualConfidence evidence analysis with custom rejection logic)
+3. **Add new constraint types** with custom handling
+4. **Implement domain-specific optimizations**
+
+See `strict_checker.py` for an example of creating a custom variant.

local_deep_research/advanced_search_system/constraint_checking/__init__.py

@@ -0,0 +1,35 @@
+"""
+Constraint checking and candidate assessment system.
+
+This module provides inheritance-based components for checking candidates
+against constraints, with different implementations available.
+"""
+
+from .base_constraint_checker import (
+    BaseConstraintChecker,
+    ConstraintCheckResult,
+)
+
+# Legacy imports for backward compatibility
+from .constraint_checker import ConstraintChecker
+from .dual_confidence_checker import DualConfidenceChecker
+from .evidence_analyzer import ConstraintEvidence, EvidenceAnalyzer
+from .rejection_engine import RejectionEngine
+from .strict_checker import StrictChecker
+from .threshold_checker import ThresholdChecker
+
+__all__ = [
+    # Base classes
+    "BaseConstraintChecker",
+    "ConstraintCheckResult",
+    # Concrete implementations
+    "DualConfidenceChecker",
+    "ThresholdChecker",
+    "StrictChecker",
+    # Supporting components
+    "EvidenceAnalyzer",
+    "ConstraintEvidence",
+    "RejectionEngine",
+    # Legacy
+    "ConstraintChecker",
+]

local_deep_research/advanced_search_system/constraint_checking/base_constraint_checker.py

@@ -0,0 +1,122 @@
+"""
+Base constraint checker for inheritance-based constraint checking system.
+
+This module provides the base interface and common functionality for
+constraint checking implementations.
+"""
+
+from abc import ABC, abstractmethod
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple
+
+from langchain_core.language_models import BaseChatModel
+from loguru import logger
+
+from ..candidates.base_candidate import Candidate
+from ..constraints.base_constraint import Constraint
+
+
+@dataclass
+class ConstraintCheckResult:
+    """Result of checking a candidate against all constraints."""
+
+    candidate: Candidate
+    total_score: float
+    constraint_scores: Dict[str, Dict]
+    should_reject: bool
+    rejection_reason: Optional[str]
+    detailed_results: List[Dict]
+
+
+class BaseConstraintChecker(ABC):
+    """
+    Base class for constraint checking implementations.
+
+    This provides the common interface and shared functionality that
+    all constraint checkers should implement.
+    """
+
+    def __init__(
+        self,
+        model: BaseChatModel,
+        evidence_gatherer=None,  # Will be passed in from strategy
+        **kwargs,
+    ):
+        """
+        Initialize the base constraint checker.
+
+        Args:
+            model: Language model for evidence analysis
+            evidence_gatherer: Function to gather evidence (from strategy)
+            **kwargs: Additional parameters for specific implementations
+        """
+        self.model = model
+        self.evidence_gatherer = evidence_gatherer
+
+    @abstractmethod
+    def check_candidate(
+        self, candidate: Candidate, constraints: List[Constraint]
+    ) -> ConstraintCheckResult:
+        """
+        Check a candidate against all constraints.
+
+        Args:
+            candidate: The candidate to check
+            constraints: List of constraints to check against
+
+        Returns:
+            ConstraintCheckResult: Complete evaluation result
+        """
+        pass
+
+    @abstractmethod
+    def should_reject_candidate(
+        self, candidate: Candidate, constraint: Constraint, evidence_data: any
+    ) -> Tuple[bool, str]:
+        """
+        Determine if a candidate should be rejected for a specific constraint.
+
+        Args:
+            candidate: The candidate being evaluated
+            constraint: The constraint being checked
+            evidence_data: Evidence data (format depends on implementation)
+
+        Returns:
+            Tuple[bool, str]: (should_reject, reason)
+        """
+        pass
+
+    def _gather_evidence_for_constraint(
+        self, candidate: Candidate, constraint: Constraint
+    ) -> List[Dict]:
+        """Gather evidence for a constraint using the provided evidence gatherer."""
+        if self.evidence_gatherer:
+            return self.evidence_gatherer(candidate, constraint)
+        else:
+            logger.warning(
+                "No evidence gatherer provided - cannot gather evidence"
+            )
+            return []
+
+    def _log_constraint_result(
+        self,
+        candidate: Candidate,
+        constraint: Constraint,
+        score: float,
+        details: Dict,
+    ):
+        """Log constraint evaluation result."""
+        symbol = "✓" if score >= 0.8 else "○" if score >= 0.5 else "✗"
+        logger.info(
+            f"{symbol} {candidate.name} | {constraint.value}: {int(score * 100)}%"
+        )
+
+    def _calculate_weighted_score(
+        self, constraint_scores: List[float], weights: List[float]
+    ) -> float:
+        """Calculate weighted average score."""
+        if not constraint_scores or not weights:
+            return 0.0
+        return sum(s * w for s, w in zip(constraint_scores, weights)) / sum(
+            weights
+        )