local-deep-research 0.4.4__py3-none-any.whl → 0.5.2__py3-none-any.whl

local_deep_research/__init__.py

@@ -5,12 +5,17 @@ Local Deep Research - A tool for conducting deep research using AI.
 __author__ = "Your Name"
 __description__ = "A tool for conducting deep research using AI"
 
+from loguru import logger
+
 from .__version__ import __version__
 from .config.llm_config import get_llm
 from .config.search_config import get_search
 from .report_generator import get_report_generator
 from .web.app import main
 
+# Disable logging by default to not interfere with user setup.
+logger.disable("local_deep_research")
+
 
 def get_advanced_search_system(strategy_name: str = "iterdrag"):
     """
@@ -32,4 +37,6 @@ __all__ = [
     "get_search",
     "get_report_generator",
     "get_advanced_search_system",
+    "main",
+    "__version__",
 ]

local_deep_research/__version__.py

@@ -1 +1 @@
-__version__ = "0.4.4"
+__version__ = "0.5.2"

local_deep_research/advanced_search_system/answer_decoding/__init__.py

@@ -0,0 +1,5 @@
+"""Answer decoding module for BrowseComp."""
+
+from .browsecomp_answer_decoder import BrowseCompAnswerDecoder
+
+__all__ = ["BrowseCompAnswerDecoder"]

local_deep_research/advanced_search_system/answer_decoding/browsecomp_answer_decoder.py

@@ -0,0 +1,421 @@
+"""
+BrowseComp Answer Decoding Pipeline
+
+This module handles encoded answers found in BrowseComp datasets.
+Some BrowseComp answers appear to be encoded (e.g., "Y00Qh+ep") and need
+decoding to extract the actual answer.
+
+Based on BROWSECOMP_IMPROVEMENT_STRATEGY.md recommendations.
+"""
+
+import base64
+import logging
+import re
+import urllib.parse
+from typing import Optional, Tuple
+
+logger = logging.getLogger(__name__)
+
+
+class BrowseCompAnswerDecoder:
+    """
+    Handle encoded BrowseComp answers with multiple decoding schemes.
+
+    Features:
+    1. Automatic encoding detection
+    2. Multiple decoding scheme support
+    3. Answer validation
+    4. Fallback to original if decoding fails
+    """
+
+    def __init__(self):
+        self.encoding_schemes = [
+            "base64",
+            "hex",
+            "url_encoding",
+            "rot13",
+            "caesar_cipher",
+        ]
+
+        # Patterns that suggest encoded content
+        self.encoded_patterns = [
+            r"^[A-Za-z0-9+/]+=*$",  # Base64 pattern
+            r"^[0-9A-Fa-f]+$",  # Hex pattern
+            r"%[0-9A-Fa-f]{2}",  # URL encoded
+            r"^[A-Za-z0-9]{8,}$",  # Random string pattern
+        ]
+
+    def decode_answer(self, raw_answer: str) -> Tuple[str, Optional[str]]:
+        """
+        Attempt to decode a potentially encoded answer.
+
+        Args:
+            raw_answer: The raw answer string that may be encoded
+
+        Returns:
+            Tuple of (decoded_answer, encoding_scheme_used)
+            If no decoding works, returns (original_answer, None)
+        """
+        if not raw_answer or len(raw_answer.strip()) == 0:
+            return raw_answer, None
+
+        # Clean the input
+        clean_answer = raw_answer.strip()
+
+        # Check if answer looks like plaintext first
+        if self.is_likely_direct_answer(clean_answer):
+            logger.debug(f"Answer appears to be plaintext: {clean_answer}")
+            return clean_answer, None
+
+        logger.info(
+            f"Attempting to decode potentially encoded answer: {clean_answer}"
+        )
+
+        # Try each encoding scheme
+        for scheme in self.encoding_schemes:
+            try:
+                decoded = self.apply_decoding_scheme(clean_answer, scheme)
+                if decoded and self.validate_decoded_answer(decoded):
+                    logger.info(
+                        f"Successfully decoded using {scheme}: {clean_answer} -> {decoded}"
+                    )
+                    return decoded, scheme
+
+            except Exception as e:
+                logger.debug(f"Failed to decode with {scheme}: {e}")
+                continue
+
+        # No successful decoding
+        logger.warning(
+            f"Could not decode answer, returning original: {clean_answer}"
+        )
+        return clean_answer, None
+
+    def is_likely_direct_answer(self, answer: str) -> bool:
+        """
+        Check if answer looks like plaintext rather than encoded.
+
+        Args:
+            answer: The answer string to check
+
+        Returns:
+            True if answer appears to be plaintext
+        """
+        # Very short answers are likely plaintext
+        if len(answer) < 4:
+            return True
+
+        # Check for common English words
+        english_indicators = [
+            "the",
+            "and",
+            "or",
+            "of",
+            "in",
+            "to",
+            "a",
+            "an",
+            "company",
+            "group",
+            "inc",
+            "ltd",
+            "corp",
+            "corporation",
+            "person",
+            "people",
+            "event",
+            "year",
+            "years",
+            "million",
+            "billion",
+            "thousand",
+        ]
+
+        answer_lower = answer.lower()
+        if any(word in answer_lower for word in english_indicators):
+            return True
+
+        # Check for sentence-like structure
+        if " " in answer and len(answer.split()) > 1:
+            # Has spaces and multiple words - likely plaintext
+            return True
+
+        # Check if it matches common answer patterns
+        common_patterns = [
+            r"^\d{4}$",  # Year
+            r"^\$?\d+\.?\d*[KMB]?$",  # Number/money
+            r"^[A-Z][a-z]+ [A-Z][a-z]+$",  # Name format
+            r"^\d+%$",  # Percentage
+        ]
+
+        for pattern in common_patterns:
+            if re.match(pattern, answer):
+                return True
+
+        # Check character distribution - encoded text often has unusual distribution
+        char_diversity = (
+            len(set(answer)) / len(answer) if len(answer) > 0 else 0
+        )
+        if char_diversity < 0.3:  # Low diversity suggests repetitive/encoded
+            return False
+
+        # If none of the encoded patterns match, probably plaintext
+        is_encoded = any(
+            re.search(pattern, answer) for pattern in self.encoded_patterns
+        )
+        return not is_encoded
+
+    def apply_decoding_scheme(self, text: str, scheme: str) -> Optional[str]:
+        """
+        Apply a specific decoding scheme to text.
+
+        Args:
+            text: Text to decode
+            scheme: Decoding scheme to use
+
+        Returns:
+            Decoded text or None if decoding fails
+        """
+        try:
+            if scheme == "base64":
+                return self._decode_base64(text)
+            elif scheme == "hex":
+                return self._decode_hex(text)
+            elif scheme == "url_encoding":
+                return self._decode_url(text)
+            elif scheme == "rot13":
+                return self._decode_rot13(text)
+            elif scheme == "caesar_cipher":
+                return self._decode_caesar(text)
+            else:
+                logger.warning(f"Unknown decoding scheme: {scheme}")
+                return None
+
+        except Exception as e:
+            logger.debug(f"Failed to apply {scheme} decoding: {e}")
+            return None
+
+    def _decode_base64(self, text: str) -> Optional[str]:
+        """Decode base64 encoded text."""
+        try:
+            # Add padding if needed
+            missing_padding = len(text) % 4
+            if missing_padding:
+                text += "=" * (4 - missing_padding)
+
+            decoded_bytes = base64.b64decode(text)
+            return decoded_bytes.decode("utf-8")
+
+        except Exception:
+            return None
+
+    def _decode_hex(self, text: str) -> Optional[str]:
+        """Decode hexadecimal encoded text."""
+        try:
+            # Remove any whitespace or non-hex characters
+            clean_hex = re.sub(r"[^0-9A-Fa-f]", "", text)
+
+            # Must have even length
+            if len(clean_hex) % 2 != 0:
+                return None
+
+            decoded_bytes = bytes.fromhex(clean_hex)
+            return decoded_bytes.decode("utf-8")
+
+        except Exception:
+            return None
+
+    def _decode_url(self, text: str) -> Optional[str]:
+        """Decode URL encoded text."""
+        try:
+            return urllib.parse.unquote(text)
+        except Exception:
+            return None
+
+    def _decode_rot13(self, text: str) -> Optional[str]:
+        """Decode ROT13 encoded text."""
+        try:
+            import codecs
+
+            return codecs.decode(text, "rot13")
+        except Exception:
+            return None
+
+    def _decode_caesar(self, text: str) -> Optional[str]:
+        """
+        Try different Caesar cipher shifts.
+        Returns the most English-like result.
+        """
+        best_result = None
+        best_score = 0
+
+        # Try shifts 1-25
+        for shift in range(1, 26):
+            try:
+                decoded = self._caesar_shift(text, shift)
+                score = self._english_score(decoded)
+
+                if score > best_score:
+                    best_score = score
+                    best_result = decoded
+
+            except Exception:
+                continue
+
+        # Only return if it looks reasonably English-like
+        return best_result if best_score > 0.3 else None
+
+    def _caesar_shift(self, text: str, shift: int) -> str:
+        """Apply Caesar cipher shift."""
+        result = []
+
+        for char in text:
+            if char.isalpha():
+                # Determine if uppercase or lowercase
+                start = ord("A") if char.isupper() else ord("a")
+                # Apply shift with wraparound
+                shifted = (ord(char) - start + shift) % 26 + start
+                result.append(chr(shifted))
+            else:
+                result.append(char)
+
+        return "".join(result)
+
+    def _english_score(self, text: str) -> float:
+        """
+        Score how English-like a text appears.
+        Simple heuristic based on common letters and words.
+        """
+        if not text:
+            return 0.0
+
+        text_lower = text.lower()
+
+        # Common English letter frequencies (approximate)
+        common_letters = "etaoinshrdlcumwfgypbvkjxqz"
+        letter_score = 0
+        letter_count = 0
+
+        for char in text_lower:
+            if char.isalpha():
+                letter_count += 1
+                # More common letters get higher scores
+                if char in common_letters[:10]:  # Top 10 most common
+                    letter_score += 2
+                elif char in common_letters[:20]:  # Top 20
+                    letter_score += 1
+
+        if letter_count == 0:
+            return 0.0
+
+        base_score = letter_score / letter_count
+
+        # Bonus for common English words
+        common_words = [
+            "the",
+            "and",
+            "of",
+            "to",
+            "a",
+            "in",
+            "is",
+            "it",
+            "you",
+            "that",
+        ]
+        word_bonus = sum(1 for word in common_words if word in text_lower)
+
+        return min(1.0, base_score + word_bonus * 0.1)
+
+    def validate_decoded_answer(self, decoded: str) -> bool:
+        """
+        Validate that decoded text looks like a reasonable answer.
+
+        Args:
+            decoded: The decoded text to validate
+
+        Returns:
+            True if decoded text appears valid
+        """
+        if not decoded or len(decoded.strip()) == 0:
+            return False
+
+        # Remove leading/trailing whitespace
+        decoded = decoded.strip()
+
+        # Check length - should be reasonable
+        if len(decoded) < 1 or len(decoded) > 1000:
+            return False
+
+        # Check for readable characters
+        printable_count = sum(1 for c in decoded if c.isprintable())
+        if printable_count / len(decoded) < 0.8:  # At least 80% printable
+            return False
+
+        # Check for control characters (bad sign)
+        if any(ord(c) < 32 and c not in "\t\n\r" for c in decoded):
+            return False
+
+        # Check character distribution
+        char_types = {
+            "alpha": sum(1 for c in decoded if c.isalpha()),
+            "digit": sum(1 for c in decoded if c.isdigit()),
+            "space": sum(1 for c in decoded if c.isspace()),
+            "punct": sum(
+                1 for c in decoded if not c.isalnum() and not c.isspace()
+            ),
+        }
+
+        total_chars = len(decoded)
+
+        # Should have some letters
+        if char_types["alpha"] / total_chars < 0.3:
+            return False
+
+        # Shouldn't be mostly punctuation
+        if char_types["punct"] / total_chars > 0.5:
+            return False
+
+        return True
+
+    def analyze_answer_encoding(self, answer: str) -> dict:
+        """
+        Analyze an answer to determine likely encoding type.
+
+        Returns analysis results for debugging/logging.
+        """
+        analysis = {
+            "original": answer,
+            "length": len(answer),
+            "likely_plaintext": self.is_likely_direct_answer(answer),
+            "pattern_matches": [],
+            "attempted_decodings": {},
+        }
+
+        # Check which patterns match
+        for i, pattern in enumerate(self.encoded_patterns):
+            if re.search(pattern, answer):
+                analysis["pattern_matches"].append(
+                    {
+                        "pattern": pattern,
+                        "type": ["base64", "hex", "url", "random"][i],
+                    }
+                )
+
+        # Try each decoding scheme
+        for scheme in self.encoding_schemes:
+            try:
+                decoded = self.apply_decoding_scheme(answer, scheme)
+                is_valid = (
+                    self.validate_decoded_answer(decoded) if decoded else False
+                )
+
+                analysis["attempted_decodings"][scheme] = {
+                    "decoded": decoded,
+                    "valid": is_valid,
+                    "length": len(decoded) if decoded else 0,
+                }
+            except Exception as e:
+                analysis["attempted_decodings"][scheme] = {"error": str(e)}
+
+        return analysis

local_deep_research/advanced_search_system/candidate_exploration/README.md

@@ -0,0 +1,219 @@
+# Candidate Exploration System
+
+This module provides an inheritance-based candidate exploration system for discovering and collecting candidates in the Local Deep Research framework.
+
+## Architecture
+
+The system is built around inheritance and provides multiple exploration strategies:
+
+### Base Class
+- **`BaseCandidateExplorer`**: Abstract base class defining the exploration interface
+
+### Concrete Implementations
+- **`ParallelExplorer`**: Runs multiple searches in parallel for speed and breadth
+- **`AdaptiveExplorer`**: Learns which search strategies work best and adapts
+- **`ConstraintGuidedExplorer`**: Uses constraints to guide the exploration process
+- **`DiversityExplorer`**: Prioritizes finding diverse candidates across categories
+
+### Supporting Components
+- **`ExplorationResult`**: Data class containing exploration results and metadata
+- **`ExplorationStrategy`**: Enum defining different exploration approaches
+
+## Usage Examples
+
+### Using ParallelExplorer
+```python
+from candidate_exploration import ParallelExplorer
+
+explorer = ParallelExplorer(
+    model=llm,
+    search_engine=search,
+    max_workers=5,           # Parallel search threads
+    queries_per_round=8,     # Queries generated per round
+    max_rounds=3             # Maximum exploration rounds
+)
+
+result = explorer.explore(
+    initial_query="hiking locations",
+    constraints=constraints,
+    entity_type="location"
+)
+```
+
+### Using AdaptiveExplorer
+```python
+from candidate_exploration import AdaptiveExplorer
+
+explorer = AdaptiveExplorer(
+    model=llm,
+    search_engine=search,
+    initial_strategies=["direct_search", "synonym_expansion", "category_exploration"],
+    adaptation_threshold=5   # Adapt after 5 searches
+)
+
+result = explorer.explore("scenic viewpoints", constraints, "viewpoint")
+```
+
+### Using ConstraintGuidedExplorer
+```python
+from candidate_exploration import ConstraintGuidedExplorer
+
+explorer = ConstraintGuidedExplorer(
+    model=llm,
+    search_engine=search,
+    constraint_weight_threshold=0.7,  # Focus on high-weight constraints
+    early_validation=True              # Validate during exploration
+)
+
+result = explorer.explore("mountain peaks", constraints, "mountain")
+```
+
+### Using DiversityExplorer
+```python
+from candidate_exploration import DiversityExplorer
+
+explorer = DiversityExplorer(
+    model=llm,
+    search_engine=search,
+    diversity_threshold=0.7,    # Minimum diversity score
+    category_limit=10,          # Max per category
+    similarity_threshold=0.8    # Similarity threshold
+)
+
+result = explorer.explore("natural landmarks", constraints, "landmark")
+```
+
+## Creating Custom Variants
+
+To create your own exploration strategy:
+
+1. **Inherit from BaseCandidateExplorer**:
+```python
+from .base_explorer import BaseCandidateExplorer, ExplorationResult
+
+class MyCustomExplorer(BaseCandidateExplorer):
+    def __init__(self, *args, my_param=0.5, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.my_param = my_param
+```
+
+2. **Implement required methods**:
+```python
+    def explore(self, initial_query, constraints=None, entity_type=None):
+        # Your exploration implementation
+        return ExplorationResult(...)
+
+    def generate_exploration_queries(self, base_query, found_candidates, constraints=None):
+        # Your query generation logic
+        return ["query1", "query2", "query3"]
+```
+
+3. **Add custom exploration logic**:
+```python
+    def _my_custom_search_strategy(self, query, context):
+        # Your custom search approach
+        pass
+```
+
+## Integration with Strategies
+
+Use in your strategy by initializing the explorer:
+
+```python
+class MyStrategy(BaseStrategy):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # Choose your explorer
+        self.explorer = AdaptiveExplorer(
+            model=self.model,
+            search_engine=self.search,
+            max_candidates=50,
+            max_search_time=120.0
+        )
+
+    def find_candidates(self, query, constraints):
+        result = self.explorer.explore(
+            initial_query=query,
+            constraints=constraints,
+            entity_type=self._detect_entity_type(query)
+        )
+
+        return result.candidates
+```
+
+## Available Explorers
+
+### ParallelExplorer
+- **Best for**: Fast, broad candidate discovery
+- **Strategy**: Breadth-first parallel search
+- **Parameters**: `max_workers`, `queries_per_round`, `max_rounds`
+- **Output**: Many candidates found quickly
+
+### AdaptiveExplorer
+- **Best for**: Learning optimal search approaches
+- **Strategy**: Adapts based on search success
+- **Parameters**: `initial_strategies`, `adaptation_threshold`
+- **Output**: Candidates found using best-performing strategies
+
+### ConstraintGuidedExplorer
+- **Best for**: Constraint-driven discovery
+- **Strategy**: Constraint-guided search prioritization
+- **Parameters**: `constraint_weight_threshold`, `early_validation`
+- **Output**: Candidates likely to satisfy constraints
+
+### DiversityExplorer
+- **Best for**: Diverse candidate sets
+- **Strategy**: Diversity-focused exploration
+- **Parameters**: `diversity_threshold`, `category_limit`, `similarity_threshold`
+- **Output**: Diverse candidates across categories
+
+## ExplorationResult Structure
+
+```python
+@dataclass
+class ExplorationResult:
+    candidates: List[Candidate]           # Found candidates
+    total_searched: int                   # Number of searches performed
+    unique_candidates: int                # Number of unique candidates
+    exploration_paths: List[str]          # Search path descriptions
+    metadata: Dict                        # Strategy-specific metadata
+    elapsed_time: float                   # Time taken for exploration
+    strategy_used: ExplorationStrategy    # Strategy that was used
+```
+
+## Performance Considerations
+
+### Speed vs. Quality Trade-offs
+- **ParallelExplorer**: Fastest, good breadth
+- **AdaptiveExplorer**: Medium speed, learns over time
+- **ConstraintGuidedExplorer**: Medium speed, higher constraint satisfaction
+- **DiversityExplorer**: Slower, but most diverse results
+
+### Memory Usage
+- All explorers track found candidates to avoid duplicates
+- Large candidate sets may use significant memory
+- Consider using `max_candidates` parameter to limit memory usage
+
+### Search Engine Load
+- Parallel explorers generate more concurrent search requests
+- Consider rate limiting or using fewer `max_workers`
+- Monitor search engine response times
+
+## Extending the System
+
+The inheritance-based design makes it easy to:
+
+1. **Create domain-specific explorers** (e.g., GeoExplorer, PersonExplorer)
+2. **Combine exploration strategies** (e.g., parallel + adaptive)
+3. **Add new search patterns** and query generation methods
+4. **Implement caching strategies** for discovered candidates
+5. **Add quality scoring** for candidate ranking
+
+## Best Practices
+
+1. **Choose the right explorer** for your use case
+2. **Set appropriate limits** (`max_candidates`, `max_search_time`)
+3. **Provide good constraints** when using ConstraintGuidedExplorer
+4. **Monitor diversity scores** when using DiversityExplorer
+5. **Let AdaptiveExplorer learn** over multiple runs for best results

local_deep_research/advanced_search_system/candidate_exploration/__init__.py

@@ -0,0 +1,25 @@
+"""
+Candidate exploration system for discovering and refining candidates.
+
+This module provides inheritance-based components for exploring and discovering
+candidates through different search strategies and approaches.
+"""
+
+from .adaptive_explorer import AdaptiveExplorer
+from .base_explorer import BaseCandidateExplorer, ExplorationResult
+from .constraint_guided_explorer import ConstraintGuidedExplorer
+from .diversity_explorer import DiversityExplorer
+from .parallel_explorer import ParallelExplorer
+from .progressive_explorer import ProgressiveExplorer
+
+__all__ = [
+    # Base classes
+    "BaseCandidateExplorer",
+    "ExplorationResult",
+    # Concrete implementations
+    "ParallelExplorer",
+    "AdaptiveExplorer",
+    "ConstraintGuidedExplorer",
+    "DiversityExplorer",
+    "ProgressiveExplorer",
+]