IntegraChat / backend /api /services /context_engineer.py
nothingworry's picture
feat: Implement Anthropic context engineering with compaction, structured prompts, and tool result clearing
09e23a2
raw
history blame
20.2 kB
# =============================================================
# File: backend/api/services/context_engineer.py
# =============================================================
"""
Context Engineering Service
Implements write, select, compress, and isolate strategies for managing agent context.
Based on LangChain's context engineering best practices.
"""
import time
from typing import Dict, Any, List, Optional
from collections import deque
class ContextScratchpad:
"""Scratchpad for saving context during agent execution.
Based on Anthropic's structured note-taking strategy:
- Agents write notes persisted outside context window
- Notes pulled back into context when needed
- Enables tracking progress across complex tasks
"""
def __init__(self, max_size: int = 50):
self.notes: deque = deque(maxlen=max_size)
self.plan: Optional[str] = None
self.key_facts: List[str] = []
self.objectives: List[Dict[str, Any]] = [] # Track objectives like Claude playing Pokémon
self.architectural_decisions: List[str] = [] # Track design decisions
self.unresolved_issues: List[str] = [] # Track bugs/issues
def add_note(self, note: str, category: str = "general"):
"""Add a note to the scratchpad."""
self.notes.append({
"timestamp": time.time(),
"note": note,
"category": category
})
def set_plan(self, plan: str):
"""Save the agent's plan."""
self.plan = plan
def add_fact(self, fact: str):
"""Add a key fact."""
if fact not in self.key_facts:
self.key_facts.append(fact)
if len(self.key_facts) > 20: # Limit facts
self.key_facts.pop(0)
def get_recent_notes(self, limit: int = 10, category: Optional[str] = None) -> List[str]:
"""Get recent notes, optionally filtered by category."""
notes = list(self.notes)
if category:
notes = [n for n in notes if n.get("category") == category]
return [n["note"] for n in notes[-limit:]]
def add_objective(self, objective: str, progress: str = "", target: str = ""):
"""Add or update an objective (like Claude playing Pokémon tracking)."""
# Update existing or add new
for obj in self.objectives:
if objective in obj.get("objective", ""):
obj["progress"] = progress
obj["target"] = target
return
self.objectives.append({
"objective": objective,
"progress": progress,
"target": target
})
if len(self.objectives) > 10:
self.objectives.pop(0)
def add_architectural_decision(self, decision: str):
"""Add an architectural decision (preserved during compaction)."""
if decision not in self.architectural_decisions:
self.architectural_decisions.append(decision)
if len(self.architectural_decisions) > 10:
self.architectural_decisions.pop(0)
def add_unresolved_issue(self, issue: str):
"""Add an unresolved issue (preserved during compaction)."""
if issue not in self.unresolved_issues:
self.unresolved_issues.append(issue)
if len(self.unresolved_issues) > 10:
self.unresolved_issues.pop(0)
def get_summary(self) -> str:
"""Get a structured summary of scratchpad contents.
Based on Anthropic's structured note-taking approach."""
parts = []
if self.plan:
parts.append(f"## Plan\n{self.plan}")
if self.objectives:
obj_text = "\n".join([f"- {o['objective']}: {o.get('progress', '')} (target: {o.get('target', 'N/A')})"
for o in self.objectives[-5:]])
parts.append(f"## Objectives\n{obj_text}")
if self.architectural_decisions:
parts.append(f"## Architectural Decisions\n" + "\n".join([f"- {d}" for d in self.architectural_decisions[-5:]]))
if self.unresolved_issues:
parts.append(f"## Unresolved Issues\n" + "\n".join([f"- {i}" for i in self.unresolved_issues[-5:]]))
if self.key_facts:
parts.append(f"## Key Facts\n" + ", ".join(self.key_facts[:5]))
if self.notes:
recent = self.get_recent_notes(5)
parts.append(f"## Recent Notes\n" + "\n".join([f"- {n}" for n in recent]))
return "\n\n".join(parts) if parts else ""
class ContextCompressor:
"""Compresses context to reduce token usage.
Based on Anthropic's context engineering best practices:
- Compaction: Summarize conversations nearing context limit
- Tool result clearing: Remove raw tool outputs once processed
- High-fidelity summarization preserving critical details
"""
def __init__(self, llm_client):
self.llm = llm_client
async def compact_conversation(self, messages: List[Dict[str, Any]], preserve_recent: int = 5, max_tokens: int = 1000) -> List[Dict[str, Any]]:
"""
Compact a conversation using Anthropic's compaction strategy.
Preserves architectural decisions, unresolved issues, and implementation details
while discarding redundant tool outputs.
Args:
messages: List of message dicts with 'role' and 'content'
preserve_recent: Number of recent messages to keep verbatim
max_tokens: Target token count for summary
Returns:
Compacted message list with summary + recent messages
"""
if len(messages) <= preserve_recent + 2:
return messages
# Keep first message (system/initial context) and last N messages
first = messages[:1] if messages else []
recent = messages[-preserve_recent:] if len(messages) > preserve_recent else messages
middle = messages[1:-preserve_recent] if len(messages) > preserve_recent + 1 else []
if not middle:
return messages
# Extract key information for compaction
user_queries = [m.get("content", "") for m in middle if m.get("role") == "user"]
assistant_responses = [m.get("content", "") for m in middle if m.get("role") == "assistant"]
tool_calls = [m for m in middle if m.get("role") == "tool" or "tool" in str(m.get("content", "")).lower()]
# Compaction prompt based on Anthropic's guidance
prompt = f"""You are compacting a conversation history. Preserve:
1. Architectural decisions and design choices
2. Unresolved bugs or issues
3. Implementation details and progress
4. Key facts and information shared
5. User preferences and requirements
Discard:
- Redundant tool outputs (raw results already processed)
- Repetitive information
- Verbose explanations that don't add value
- Tool call details that are no longer needed
Conversation to compact:
{chr(10).join([f"{m.get('role', 'user')}: {str(m.get('content', ''))[:400]}" for m in middle[:20]])}
Provide a high-fidelity summary that preserves critical context (max {max_tokens} tokens):"""
try:
summary = await self.llm.simple_call(prompt, temperature=0.0)
summary_msg = {
"role": "system",
"content": f"[Compacted conversation history: {summary}]",
"_compacted": True,
"_original_length": len(middle)
}
return first + [summary_msg] + recent
except Exception:
# Fallback: simple trimming
return first + recent
async def summarize_conversation(self, messages: List[Dict[str, Any]], max_tokens: int = 500) -> str:
"""
Summarize a conversation while preserving key decisions and facts.
Uses Anthropic's compaction principles.
Args:
messages: List of message dicts with 'role' and 'content'
max_tokens: Target token count for summary
Returns:
Summarized conversation
"""
if len(messages) <= 2:
return "\n".join([f"{m.get('role', 'user')}: {m.get('content', '')[:200]}" for m in messages])
# Extract key information
user_queries = [m.get("content", "") for m in messages if m.get("role") == "user"]
assistant_responses = [m.get("content", "") for m in messages if m.get("role") == "assistant"]
prompt = f"""Summarize this conversation using high-fidelity compaction. Preserve:
1. Key user questions/requests
2. Important decisions made (architectural, design, implementation)
3. Critical facts or information shared
4. Unresolved issues or bugs
5. Implementation progress
Discard redundant tool outputs and repetitive information.
Conversation:
{chr(10).join([f"User: {q[:300]}" for q in user_queries[-5:]])}
{chr(10).join([f"Assistant: {r[:300]}" for r in assistant_responses[-5:]])}
Provide a concise, high-fidelity summary (max {max_tokens} tokens):"""
try:
summary = await self.llm.simple_call(prompt, temperature=0.0)
return summary[:max_tokens * 4] # Rough token limit
except Exception:
# Fallback: simple truncation
return "\n".join([f"{m.get('role', 'user')}: {m.get('content', '')[:100]}..." for m in messages[-5:]])
def trim_messages(self, messages: List[Dict[str, Any]], keep_first: int = 2, keep_last: int = 10) -> List[Dict[str, Any]]:
"""
Trim messages, keeping first N and last M.
Based on Anthropic's guidance: preserve system context and recent interactions.
Args:
messages: List of messages
keep_first: Number of initial messages to keep (system context)
keep_last: Number of recent messages to keep
Returns:
Trimmed message list
"""
if len(messages) <= keep_first + keep_last:
return messages
return messages[:keep_first] + messages[-keep_last:]
def clear_tool_results(self, messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
"""
Clear tool call results from messages (safest form of compaction).
Based on Anthropic's recommendation: once a tool has been called deep in history,
the raw result is often no longer needed.
Args:
messages: List of messages
Returns:
Messages with tool results cleared (tool calls kept, results removed)
"""
cleared = []
for msg in messages:
# Keep tool calls but clear large results
if msg.get("role") == "tool" or "tool" in str(msg.get("content", "")).lower():
# Keep tool metadata but truncate large results
content = str(msg.get("content", ""))
if len(content) > 500:
msg_copy = msg.copy()
msg_copy["content"] = content[:200] + "... [tool result truncated]"
msg_copy["_tool_result_cleared"] = True
cleared.append(msg_copy)
else:
cleared.append(msg)
else:
cleared.append(msg)
return cleared
async def compress_tool_output(self, tool_name: str, output: Dict[str, Any], max_length: int = 500) -> Dict[str, Any]:
"""
Compress tool output to reduce tokens.
Args:
tool_name: Name of the tool
output: Tool output dict
max_length: Max characters for compressed output
Returns:
Compressed output
"""
if tool_name == "web":
# Compress web search results
hits = output.get("results", [])
if len(hits) > 5:
# Keep only top 5 results
output["results"] = hits[:5]
output["_compressed"] = True
output["_original_count"] = len(hits)
elif tool_name == "rag":
# Compress RAG results
hits = output.get("results", [])
if len(hits) > 5:
output["results"] = hits[:5]
output["_compressed"] = True
output["_original_count"] = len(hits)
# Summarize long text fields
for key in ["text", "content", "snippet"]:
if key in output and len(str(output[key])) > max_length:
text = str(output[key])
output[key] = text[:max_length] + "..."
output[f"{key}_compressed"] = True
return output
class ContextSelector:
"""Selects relevant context for agent steps."""
def __init__(self, llm_client):
self.llm = llm_client
async def select_relevant_memories(self, query: str, memories: List[Dict[str, Any]], limit: int = 5) -> List[Dict[str, Any]]:
"""
Select most relevant memories for a query.
Args:
query: User query
memories: List of memory dicts
limit: Max memories to return
Returns:
Selected memories
"""
if not memories or len(memories) <= limit:
return memories
# Simple keyword-based selection (can be enhanced with embeddings)
query_lower = query.lower()
scored = []
for mem in memories:
content = str(mem.get("content", "")).lower()
score = sum(1 for word in query_lower.split() if word in content)
scored.append((score, mem))
# Sort by score and return top N
scored.sort(reverse=True, key=lambda x: x[0])
return [mem for score, mem in scored[:limit] if score > 0]
def select_relevant_tools(self, query: str, available_tools: List[Dict[str, Any]], limit: int = 5) -> List[Dict[str, Any]]:
"""
Select most relevant tools for a query.
Args:
query: User query
available_tools: List of tool dicts with descriptions
limit: Max tools to return
Returns:
Selected tools
"""
if not available_tools or len(available_tools) <= limit:
return available_tools
# Simple keyword matching (can be enhanced with semantic search)
query_lower = query.lower()
scored = []
for tool in available_tools:
desc = str(tool.get("description", "")).lower()
name = str(tool.get("name", "")).lower()
score = sum(1 for word in query_lower.split() if word in desc or word in name)
scored.append((score, tool))
scored.sort(reverse=True, key=lambda x: x[0])
return [tool for score, tool in scored[:limit]]
class ContextIsolator:
"""Isolates context to prevent token bloat."""
def __init__(self):
self.isolated_data: Dict[str, Any] = {}
def isolate_tool_output(self, tool_name: str, output: Any, key: Optional[str] = None) -> str:
"""
Isolate tool output, storing it separately and returning a reference.
Args:
tool_name: Name of the tool
output: Tool output
key: Optional key for storage
Returns:
Reference string to use in context
"""
storage_key = key or f"{tool_name}_{int(time.time())}"
self.isolated_data[storage_key] = {
"tool": tool_name,
"output": output,
"timestamp": time.time()
}
return f"[ISOLATED:{storage_key}]"
def get_isolated(self, key: str) -> Optional[Any]:
"""Retrieve isolated data by key."""
return self.isolated_data.get(key, {}).get("output")
def clear_old_isolated(self, max_age_seconds: int = 3600):
"""Clear isolated data older than max_age_seconds."""
current_time = time.time()
keys_to_remove = [
key for key, data in self.isolated_data.items()
if current_time - data.get("timestamp", 0) > max_age_seconds
]
for key in keys_to_remove:
del self.isolated_data[key]
class ContextEngineer:
"""Main context engineering service combining all strategies."""
def __init__(self, llm_client):
self.scratchpad = ContextScratchpad()
self.compressor = ContextCompressor(llm_client)
self.selector = ContextSelector(llm_client)
self.isolator = ContextIsolator()
self.llm = llm_client
def write_to_scratchpad(self, note: str, category: str = "general"):
"""Write to scratchpad."""
self.scratchpad.add_note(note, category)
def save_plan(self, plan: str):
"""Save agent plan."""
self.scratchpad.set_plan(plan)
def save_fact(self, fact: str):
"""Save key fact."""
self.scratchpad.add_fact(fact)
def get_scratchpad_context(self, limit: int = 10) -> str:
"""Get relevant scratchpad context."""
return self.scratchpad.get_summary()
async def compress_if_needed(self, messages: List[Dict[str, Any]], max_tokens: int = 8000,
use_compaction: bool = True) -> List[Dict[str, Any]]:
"""
Compress messages if they exceed token limit.
Uses Anthropic's compaction strategy: high-fidelity summarization
preserving architectural decisions, unresolved issues, and implementation details.
Args:
messages: List of messages
max_tokens: Token limit
use_compaction: Use full compaction vs simple trimming
Returns:
Compressed messages
"""
# Rough token estimate (4 chars per token)
total_chars = sum(len(str(m.get("content", ""))) for m in messages)
estimated_tokens = total_chars // 4
if estimated_tokens > max_tokens:
# First, try tool result clearing (safest form of compaction)
cleared = self.compressor.clear_tool_results(messages)
cleared_chars = sum(len(str(m.get("content", ""))) for m in cleared)
cleared_tokens = cleared_chars // 4
if cleared_tokens <= max_tokens:
return cleared
# If still over limit, use full compaction
if use_compaction and len(messages) > 10:
return await self.compressor.compact_conversation(messages, preserve_recent=5, max_tokens=1000)
else:
# Fallback: simple trimming
return self.compressor.trim_messages(messages, keep_first=2, keep_last=5)
return messages
async def select_context(self, query: str, available_context: Dict[str, Any]) -> Dict[str, Any]:
"""
Select relevant context for a query.
Args:
query: User query
available_context: Dict with keys like 'memories', 'tools', etc.
Returns:
Selected context dict
"""
selected = {}
# Select memories
if "memories" in available_context:
selected["memories"] = await self.selector.select_relevant_memories(
query, available_context["memories"]
)
# Select tools
if "tools" in available_context:
selected["tools"] = self.selector.select_relevant_tools(
query, available_context["tools"]
)
return selected
def isolate_large_output(self, tool_name: str, output: Any) -> str:
"""Isolate large tool output."""
return self.isolator.isolate_tool_output(tool_name, output)
def get_isolated_context(self, key: str) -> Optional[Any]:
"""Get isolated context."""
return self.isolator.get_isolated(key)