feat(context): implement token budget management (Phase 2)

Add TokenCalculator with LLM Gateway integration for accurate token
counting with in-memory caching and fallback character-based estimation.
Implement TokenBudget for tracking allocations per context type with
budget enforcement, and BudgetAllocator for creating budgets based on
model context window sizes.

- TokenCalculator: MCP integration, caching, model-specific ratios
- TokenBudget: allocation tracking, can_fit/allocate/deallocate/reset
- BudgetAllocator: model context sizes, budget creation and adjustment
- 35 comprehensive tests covering all budget functionality

Part of #61 - Context Management Engine

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
2026-01-04 02:13:23 +01:00
parent 22ecb5e989
commit dfa75e682e
5 changed files with 1277 additions and 0 deletions

View File

@@ -0,0 +1,284 @@
"""
Token Calculator for Context Management.
Provides token counting with caching and fallback estimation.
Integrates with LLM Gateway for accurate counts.
"""
import hashlib
import logging
from typing import TYPE_CHECKING, Any, Protocol
if TYPE_CHECKING:
from app.services.mcp.client_manager import MCPClientManager
logger = logging.getLogger(__name__)
class TokenCounterProtocol(Protocol):
"""Protocol for token counting implementations."""
async def count_tokens(
self,
text: str,
model: str | None = None,
) -> int:
"""Count tokens in text."""
...
class TokenCalculator:
"""
Token calculator with LLM Gateway integration.
Features:
- In-memory caching for repeated text
- Fallback to character-based estimation
- Model-specific counting when possible
The calculator uses the LLM Gateway's count_tokens tool
for accurate counting, with a local cache to avoid
repeated calls for the same content.
"""
# Default characters per token ratio for estimation
DEFAULT_CHARS_PER_TOKEN = 4.0
# Model-specific ratios (more accurate estimation)
MODEL_CHAR_RATIOS: dict[str, float] = {
"claude": 3.5,
"gpt-4": 4.0,
"gpt-3.5": 4.0,
"gemini": 4.0,
}
def __init__(
self,
mcp_manager: "MCPClientManager | None" = None,
project_id: str = "system",
agent_id: str = "context-engine",
cache_enabled: bool = True,
cache_max_size: int = 10000,
) -> None:
"""
Initialize token calculator.
Args:
mcp_manager: MCP client manager for LLM Gateway calls
project_id: Project ID for LLM Gateway calls
agent_id: Agent ID for LLM Gateway calls
cache_enabled: Whether to enable in-memory caching
cache_max_size: Maximum cache entries
"""
self._mcp = mcp_manager
self._project_id = project_id
self._agent_id = agent_id
self._cache_enabled = cache_enabled
self._cache_max_size = cache_max_size
# In-memory cache: hash(model:text) -> token_count
self._cache: dict[str, int] = {}
self._cache_hits = 0
self._cache_misses = 0
def _get_cache_key(self, text: str, model: str | None) -> str:
"""Generate cache key from text and model."""
# Use hash for efficient storage
content = f"{model or 'default'}:{text}"
return hashlib.sha256(content.encode()).hexdigest()[:32]
def _check_cache(self, cache_key: str) -> int | None:
"""Check cache for existing count."""
if not self._cache_enabled:
return None
if cache_key in self._cache:
self._cache_hits += 1
return self._cache[cache_key]
self._cache_misses += 1
return None
def _store_cache(self, cache_key: str, count: int) -> None:
"""Store count in cache."""
if not self._cache_enabled:
return
# Simple LRU-like eviction: remove oldest entries when full
if len(self._cache) >= self._cache_max_size:
# Remove first 10% of entries
entries_to_remove = self._cache_max_size // 10
keys_to_remove = list(self._cache.keys())[:entries_to_remove]
for key in keys_to_remove:
del self._cache[key]
self._cache[cache_key] = count
def estimate_tokens(self, text: str, model: str | None = None) -> int:
"""
Estimate token count based on character count.
This is a fast fallback when LLM Gateway is unavailable.
Args:
text: Text to count
model: Optional model for more accurate ratio
Returns:
Estimated token count
"""
if not text:
return 0
# Get model-specific ratio
ratio = self.DEFAULT_CHARS_PER_TOKEN
if model:
model_lower = model.lower()
for model_prefix, model_ratio in self.MODEL_CHAR_RATIOS.items():
if model_prefix in model_lower:
ratio = model_ratio
break
return max(1, int(len(text) / ratio))
async def count_tokens(
self,
text: str,
model: str | None = None,
) -> int:
"""
Count tokens in text.
Uses LLM Gateway for accurate counts with fallback to estimation.
Args:
text: Text to count
model: Optional model for accurate counting
Returns:
Token count
"""
if not text:
return 0
# Check cache first
cache_key = self._get_cache_key(text, model)
cached = self._check_cache(cache_key)
if cached is not None:
return cached
# Try LLM Gateway
if self._mcp is not None:
try:
result = await self._mcp.call_tool(
server="llm-gateway",
tool="count_tokens",
args={
"project_id": self._project_id,
"agent_id": self._agent_id,
"text": text,
"model": model,
},
)
# Parse result
if result.success and result.data:
count = self._parse_token_count(result.data)
if count is not None:
self._store_cache(cache_key, count)
return count
except Exception as e:
logger.warning(f"LLM Gateway token count failed, using estimation: {e}")
# Fallback to estimation
count = self.estimate_tokens(text, model)
self._store_cache(cache_key, count)
return count
def _parse_token_count(self, data: Any) -> int | None:
"""Parse token count from LLM Gateway response."""
if isinstance(data, dict):
if "token_count" in data:
return int(data["token_count"])
if "tokens" in data:
return int(data["tokens"])
if "count" in data:
return int(data["count"])
if isinstance(data, int):
return data
if isinstance(data, str):
# Try to parse from text content
try:
# Handle {"token_count": 123} or just "123"
import json
parsed = json.loads(data)
if isinstance(parsed, dict) and "token_count" in parsed:
return int(parsed["token_count"])
if isinstance(parsed, int):
return parsed
except (json.JSONDecodeError, ValueError):
# Try direct int conversion
try:
return int(data)
except ValueError:
pass
return None
async def count_tokens_batch(
self,
texts: list[str],
model: str | None = None,
) -> list[int]:
"""
Count tokens for multiple texts.
Efficient batch counting with caching.
Args:
texts: List of texts to count
model: Optional model for accurate counting
Returns:
List of token counts (same order as input)
"""
results: list[int] = []
for text in texts:
count = await self.count_tokens(text, model)
results.append(count)
return results
def clear_cache(self) -> None:
"""Clear the token count cache."""
self._cache.clear()
self._cache_hits = 0
self._cache_misses = 0
def get_cache_stats(self) -> dict[str, Any]:
"""Get cache statistics."""
total = self._cache_hits + self._cache_misses
hit_rate = self._cache_hits / total if total > 0 else 0.0
return {
"enabled": self._cache_enabled,
"size": len(self._cache),
"max_size": self._cache_max_size,
"hits": self._cache_hits,
"misses": self._cache_misses,
"hit_rate": round(hit_rate, 3),
}
def set_mcp_manager(self, mcp_manager: "MCPClientManager") -> None:
"""
Set the MCP manager (for lazy initialization).
Args:
mcp_manager: MCP client manager instance
"""
self._mcp = mcp_manager