loom-agent 0.0.1__py3-none-any.whl → 0.0.3__py3-none-any.whl

loom/core/compression_manager.py

@@ -151,8 +151,8 @@ Now compress the conversation above following this exact structure:"""
             return messages, CompressionMetadata(
                 original_message_count=0,
                 compressed_message_count=0,
-                original_tokens=0,
-                compressed_tokens=0,
+                original_token_count=0,
+                compressed_token_count=0,
                 compression_ratio=0.0,
                 key_topics=[],
             )
@@ -163,11 +163,12 @@ Now compress the conversation above following this exact structure:"""
 
         if not compressible:
             # No messages to compress, return as-is
+            token_count = count_messages_tokens(messages)
             return messages, CompressionMetadata(
                 original_message_count=len(messages),
                 compressed_message_count=len(messages),
-                original_tokens=count_messages_tokens(messages),
-                compressed_tokens=count_messages_tokens(messages),
+                original_token_count=token_count,
+                compressed_token_count=token_count,
                 compression_ratio=1.0,
                 key_topics=[],
             )
@@ -201,13 +202,16 @@ Now compress the conversation above following this exact structure:"""
             windowed_messages = self.sliding_window_fallback(compressible, self.sliding_window_size)
             final_messages = system_messages + windowed_messages
             compressed_tokens = count_messages_tokens(windowed_messages)
+            ratio = compressed_tokens / original_tokens if original_tokens > 0 else 0.0
+            # Ensure ratio is clamped to [0.0, 1.0]
+            ratio = min(max(ratio, 0.0), 1.0)
 
             metadata = CompressionMetadata(
                 original_message_count=len(compressible),
                 compressed_message_count=len(windowed_messages),
-                original_tokens=original_tokens,
-                compressed_tokens=compressed_tokens,
-                compression_ratio=compressed_tokens / original_tokens if original_tokens > 0 else 0.0,
+                original_token_count=original_tokens,
+                compressed_token_count=compressed_tokens,
+                compression_ratio=ratio,
                 key_topics=["fallback"],
             )
             return final_messages, metadata
@@ -223,13 +227,16 @@ Now compress the conversation above following this exact structure:"""
 
         # Combine system messages + compressed summary
         final_messages = system_messages + [compressed_message]
+        ratio = compressed_tokens / original_tokens if original_tokens > 0 else 0.0
+        # Ensure ratio is clamped to [0.0, 1.0]
+        ratio = min(max(ratio, 0.0), 1.0)
 
         metadata = CompressionMetadata(
             original_message_count=len(compressible),
             compressed_message_count=1,
-            original_tokens=original_tokens,
-            compressed_tokens=compressed_tokens,
-            compression_ratio=compressed_tokens / original_tokens if original_tokens > 0 else 0.0,
+            original_token_count=original_tokens,
+            compressed_token_count=compressed_tokens,
+            compression_ratio=ratio,
             key_topics=key_topics,
         )
 

loom/core/context_assembly.py

@@ -0,0 +1,437 @@
+"""
+Context Assembly Module
+
+Provides intelligent context assembly with priority-based component management
+and token budget constraints.
+
+This module fixes the RAG Context Bug where retrieved documents were being
+overwritten by system prompts.
+"""
+
+from __future__ import annotations
+
+import hashlib
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Callable, Any
+from enum import IntEnum
+
+
+class ComponentPriority(IntEnum):
+    """
+    Component priority levels for context assembly.
+
+    Higher values = higher priority = less likely to be truncated.
+    """
+    CRITICAL = 100     # Base instructions (must be included)
+    HIGH = 90          # RAG context, important configurations
+    MEDIUM = 70        # Tool definitions
+    LOW = 50           # Examples, additional hints
+    OPTIONAL = 30      # Optional content
+
+
+@dataclass
+class ContextComponent:
+    """
+    A single component of the context.
+
+    Attributes:
+        name: Component identifier (e.g., "base_instructions", "retrieved_docs")
+        content: The actual content text
+        priority: Priority level (0-100)
+        token_count: Estimated number of tokens
+        truncatable: Whether this component can be truncated
+    """
+    name: str
+    content: str
+    priority: int
+    token_count: int
+    truncatable: bool = True
+
+
+class ContextAssembler:
+    """
+    Intelligent context assembler with priority-based management.
+
+    Features:
+    - Priority-based component ordering
+    - Token budget management
+    - Smart truncation of low-priority components
+    - Guarantee high-priority component integrity
+    - Component caching for performance
+    - Dynamic priority adjustment
+    - Context reuse optimization
+
+    Example:
+        ```python
+        assembler = ContextAssembler(max_tokens=4000)
+
+        # Add components with priorities
+        assembler.add_component(
+            "base_instructions",
+            "You are a helpful assistant.",
+            priority=ComponentPriority.CRITICAL,
+            truncatable=False
+        )
+
+        assembler.add_component(
+            "retrieved_docs",
+            doc_context,
+            priority=ComponentPriority.HIGH,
+            truncatable=True
+        )
+
+        # Assemble final context
+        final_prompt = assembler.assemble()
+        ```
+    """
+
+    def __init__(
+        self,
+        max_tokens: int = 16000,
+        token_counter: Optional[Callable[[str], int]] = None,
+        token_buffer: float = 0.9,  # Use 90% of budget for safety
+        enable_caching: bool = True,
+        cache_size: int = 100
+    ):
+        """
+        Initialize the context assembler.
+
+        Args:
+            max_tokens: Maximum token budget
+            token_counter: Custom token counting function (defaults to simple estimation)
+            token_buffer: Safety buffer ratio (0.9 = use 90% of max_tokens)
+            enable_caching: Enable component caching for performance
+            cache_size: Maximum number of cached components
+        """
+        self.max_tokens = int(max_tokens * token_buffer)
+        self.token_counter = token_counter or self._estimate_tokens
+        self.components: List[ContextComponent] = []
+        
+        # Performance optimizations
+        self.enable_caching = enable_caching
+        self._component_cache: Dict[str, ContextComponent] = {}
+        self._cache_size = cache_size
+        self._assembly_cache: Optional[str] = None
+        self._last_components_hash: Optional[str] = None
+
+    def add_component(
+        self,
+        name: str,
+        content: str,
+        priority: int,
+        truncatable: bool = True
+    ) -> None:
+        """
+        Add a context component.
+
+        Args:
+            name: Component identifier (e.g., "base_instructions", "retrieved_docs")
+            content: Component content
+            priority: Priority level (0-100, higher = more important)
+            truncatable: Whether this component can be truncated
+        """
+        if not content or not content.strip():
+            return
+
+        token_count = self.token_counter(content)
+        component = ContextComponent(
+            name=name,
+            content=content.strip(),
+            priority=priority,
+            token_count=token_count,
+            truncatable=truncatable
+        )
+        self.components.append(component)
+        
+        # Update cache if enabled
+        if self.enable_caching:
+            self._component_cache[name] = component
+            # Maintain cache size limit
+            if len(self._component_cache) > self._cache_size:
+                # Remove oldest entries (simple LRU)
+                oldest_key = next(iter(self._component_cache))
+                del self._component_cache[oldest_key]
+
+    def assemble(self) -> str:
+        """
+        Assemble the final context from all components.
+
+        Strategy:
+        1. Check cache for identical component configuration
+        2. Sort components by priority (descending)
+        3. Add components until budget is reached
+        4. Truncate low-priority components if needed
+        5. Merge all components into final string
+
+        Returns:
+            Assembled context string
+        """
+        if not self.components:
+            return ""
+
+        # Check cache if enabled
+        if self.enable_caching:
+            current_hash = self._get_components_hash()
+            if (self._assembly_cache is not None and 
+                self._last_components_hash == current_hash):
+                return self._assembly_cache
+
+        # Sort by priority (highest first)
+        sorted_components = sorted(
+            self.components,
+            key=lambda c: c.priority,
+            reverse=True
+        )
+
+        # Calculate total tokens
+        total_tokens = sum(c.token_count for c in sorted_components)
+
+        # Truncate if over budget
+        if total_tokens > self.max_tokens:
+            sorted_components = self._truncate_components(sorted_components)
+
+        # Merge components
+        sections = []
+        for component in sorted_components:
+            # Add section header and content
+            header = f"# {component.name.replace('_', ' ').upper()}"
+            sections.append(f"{header}\n{component.content}")
+
+        result = "\n\n".join(sections)
+        
+        # Update cache if enabled
+        if self.enable_caching:
+            self._assembly_cache = result
+            self._last_components_hash = self._get_components_hash()
+
+        return result
+
+    def _get_components_hash(self) -> str:
+        """
+        Generate hash for current component configuration
+
+        优化版本：
+        - 使用 blake2b 替代 MD5（更快）
+        - 直接update字节而非拼接字符串
+        - 移除不必要的排序
+        """
+        # 使用 blake2b，比 MD5 更快且安全
+        hasher = hashlib.blake2b(digest_size=16)
+
+        # 直接更新hasher，避免字符串拼接
+        for comp in self.components:
+            hasher.update(comp.name.encode())
+            hasher.update(str(comp.priority).encode())
+            hasher.update(str(comp.token_count).encode())
+            hasher.update(b'1' if comp.truncatable else b'0')
+
+        return hasher.hexdigest()
+
+    def adjust_priority(self, component_name: str, new_priority: int) -> bool:
+        """
+        Dynamically adjust component priority.
+        
+        Args:
+            component_name: Name of the component to adjust
+            new_priority: New priority value
+            
+        Returns:
+            True if component was found and adjusted, False otherwise
+        """
+        for component in self.components:
+            if component.name == component_name:
+                component.priority = new_priority
+                # Clear cache since configuration changed
+                if self.enable_caching:
+                    self._assembly_cache = None
+                    self._last_components_hash = None
+                return True
+        return False
+
+    def get_component_stats(self) -> Dict[str, Any]:
+        """Get statistics about current components"""
+        if not self.components:
+            return {"total_components": 0, "total_tokens": 0}
+        
+        total_tokens = sum(c.token_count for c in self.components)
+        priority_distribution = {}
+        
+        for comp in self.components:
+            priority_distribution[comp.priority] = priority_distribution.get(comp.priority, 0) + 1
+        
+        return {
+            "total_components": len(self.components),
+            "total_tokens": total_tokens,
+            "budget_utilization": total_tokens / self.max_tokens if self.max_tokens > 0 else 0,
+            "priority_distribution": priority_distribution,
+            "cache_enabled": self.enable_caching,
+            "cache_size": len(self._component_cache) if self.enable_caching else 0
+        }
+
+    def clear_cache(self) -> None:
+        """Clear all caches"""
+        if self.enable_caching:
+            self._component_cache.clear()
+            self._assembly_cache = None
+            self._last_components_hash = None
+
+    def _truncate_components(
+        self,
+        components: List[ContextComponent]
+    ) -> List[ContextComponent]:
+        """
+        Intelligently truncate components to fit token budget.
+
+        Strategy:
+        1. Always include non-truncatable components
+        2. Add truncatable components by priority
+        3. Truncate lower-priority components if needed
+
+        Args:
+            components: Sorted list of components (by priority, descending)
+
+        Returns:
+            List of components that fit within budget
+        """
+        budget_remaining = self.max_tokens
+        result = []
+
+        # Phase 1: Add all non-truncatable components
+        for comp in components:
+            if not comp.truncatable:
+                if comp.token_count <= budget_remaining:
+                    result.append(comp)
+                    budget_remaining -= comp.token_count
+                else:
+                    # Non-truncatable component is too large
+                    print(
+                        f"Warning: Non-truncatable component '{comp.name}' "
+                        f"({comp.token_count} tokens) exceeds remaining budget "
+                        f"({budget_remaining} tokens). Skipping."
+                    )
+
+        # Phase 2: Add truncatable components
+        truncatable = [c for c in components if c.truncatable]
+
+        for comp in truncatable:
+            if comp.token_count <= budget_remaining:
+                # Add complete component
+                result.append(comp)
+                budget_remaining -= comp.token_count
+            elif budget_remaining > 100:  # Minimum 100 tokens to be useful
+                # Truncate and add
+                truncated_content = self._truncate_content(
+                    comp.content,
+                    budget_remaining - 20  # Reserve 20 tokens for "... (truncated)" marker
+                )
+                truncated_comp = ContextComponent(
+                    name=comp.name,
+                    content=truncated_content,
+                    priority=comp.priority,
+                    token_count=self.token_counter(truncated_content),
+                    truncatable=comp.truncatable
+                )
+                result.append(truncated_comp)
+                budget_remaining = 0
+                break
+            else:
+                # Not enough budget left, skip remaining components
+                break
+
+        return result
+
+    def _truncate_content(self, content: str, max_tokens: int) -> str:
+        """
+        Truncate content to fit within token limit.
+
+        Strategy: Proportional character truncation with conservative estimation
+
+        Args:
+            content: Content to truncate
+            max_tokens: Maximum tokens allowed
+
+        Returns:
+            Truncated content with marker
+        """
+        current_tokens = self.token_counter(content)
+
+        if current_tokens <= max_tokens:
+            return content
+
+        # Calculate target character count (conservative)
+        ratio = max_tokens / current_tokens
+        target_chars = int(len(content) * ratio * 0.95)  # 5% safety margin
+
+        if target_chars < 100:
+            # Too small to be useful
+            return ""
+
+        # Truncate and add marker
+        truncated = content[:target_chars].rsplit(' ', 1)[0]  # Truncate at word boundary
+        return f"{truncated}\n\n... (truncated due to token limit)"
+
+    def _estimate_tokens(self, text: str) -> int:
+        """
+        Simple token estimation.
+
+        Rule of thumb: 1 token ≈ 4 characters for English text
+        This is a conservative estimate that works reasonably well.
+
+        For precise counting, use a model-specific tokenizer.
+
+        Args:
+            text: Text to estimate
+
+        Returns:
+            Estimated token count
+        """
+        return len(text) // 4
+
+    def get_summary(self) -> Dict:
+        """
+        Get assembly summary for debugging and monitoring.
+
+        Returns:
+            Dictionary containing:
+            - components: List of component info (name, priority, tokens, truncatable)
+            - total_tokens: Sum of all component tokens
+            - budget: Maximum token budget
+            - overflow: Tokens over budget (0 if within budget)
+            - utilization: Budget utilization percentage
+        """
+        total_tokens = sum(c.token_count for c in self.components)
+        overflow = max(0, total_tokens - self.max_tokens)
+        utilization = (total_tokens / self.max_tokens * 100) if self.max_tokens > 0 else 0
+
+        return {
+            "components": [
+                {
+                    "name": c.name,
+                    "priority": c.priority,
+                    "tokens": c.token_count,
+                    "truncatable": c.truncatable
+                }
+                for c in sorted(self.components, key=lambda x: x.priority, reverse=True)
+            ],
+            "total_tokens": total_tokens,
+            "budget": self.max_tokens,
+            "overflow": overflow,
+            "utilization": round(utilization, 2)
+        }
+
+    def clear(self) -> None:
+        """Clear all components."""
+        self.components.clear()
+
+    def __len__(self) -> int:
+        """Return number of components."""
+        return len(self.components)
+
+    def __repr__(self) -> str:
+        """String representation."""
+        summary = self.get_summary()
+        return (
+            f"ContextAssembler(components={len(self.components)}, "
+            f"tokens={summary['total_tokens']}/{summary['budget']}, "
+            f"utilization={summary['utilization']}%)"
+        )