Global LLM usage tracker (singleton).
| 44 | |
| 45 | |
| 46 | class LLMUsageTracker: |
| 47 | """Global LLM usage tracker (singleton).""" |
| 48 | |
| 49 | _instance = None |
| 50 | |
| 51 | def __new__(cls): |
| 52 | if cls._instance is None: |
| 53 | cls._instance = super().__new__(cls) |
| 54 | cls._instance._init() |
| 55 | return cls._instance |
| 56 | |
| 57 | def _init(self): |
| 58 | self._memorize_usage = TokenUsage() |
| 59 | self._query_usage = TokenUsage() |
| 60 | self._current_phase = "unknown" |
| 61 | |
| 62 | def set_phase(self, phase: str) -> None: |
| 63 | self._current_phase = phase |
| 64 | |
| 65 | def record(self, response: "LLMResponse") -> None: |
| 66 | if self._current_phase == "memorize": |
| 67 | self._memorize_usage.add(response.input_tokens, response.output_tokens, response.latency) |
| 68 | else: |
| 69 | self._query_usage.add(response.input_tokens, response.output_tokens, response.latency) |
| 70 | |
| 71 | def reset(self) -> None: |
| 72 | self._memorize_usage = TokenUsage() |
| 73 | self._query_usage = TokenUsage() |
| 74 | self._current_phase = "unknown" |
| 75 | |
| 76 | def get_stats(self) -> Dict[str, Any]: |
| 77 | total = TokenUsage() |
| 78 | total.merge(self._memorize_usage) |
| 79 | total.merge(self._query_usage) |
| 80 | return { |
| 81 | "memorize_phase": self._memorize_usage.to_dict(), |
| 82 | "query_phase": self._query_usage.to_dict(), |
| 83 | "total": total.to_dict(), |
| 84 | } |
| 85 | |
| 86 | |
| 87 | _usage_tracker: Optional[LLMUsageTracker] = None |