|
| 1 | +"""Module with the Openlayer callback handler for LangChain.""" |
| 2 | + |
| 3 | +# pylint: disable=unused-argument |
| 4 | +import time |
| 5 | +from typing import Any, Dict, List, Optional, Union |
| 6 | + |
| 7 | +from langchain import schema as langchain_schema |
| 8 | +from langchain.callbacks.base import BaseCallbackHandler |
| 9 | + |
| 10 | +from .. import constants |
| 11 | +from ..tracing import tracer |
| 12 | + |
| 13 | +LANGCHAIN_TO_OPENLAYER_PROVIDER_MAP = {"openai-chat": "OpenAI"} |
| 14 | +PROVIDER_TO_STEP_NAME = {"OpenAI": "OpenAI Chat Completion"} |
| 15 | + |
| 16 | + |
| 17 | +class OpenlayerHandler(BaseCallbackHandler): |
| 18 | + """LangChain callback handler that logs to Openlayer.""" |
| 19 | + |
| 20 | + def __init__( |
| 21 | + self, |
| 22 | + **kwargs: Any, |
| 23 | + ) -> None: |
| 24 | + super().__init__() |
| 25 | + |
| 26 | + self.start_time: float = None |
| 27 | + self.end_time: float = None |
| 28 | + self.prompt: List[Dict[str, str]] = None |
| 29 | + self.latency: float = None |
| 30 | + self.provider: str = None |
| 31 | + self.model: Optional[str] = None |
| 32 | + self.model_parameters: Dict[str, Any] = None |
| 33 | + self.cost: Optional[float] = None |
| 34 | + self.prompt_tokens: int = None |
| 35 | + self.completion_tokens: int = None |
| 36 | + self.total_tokens: int = None |
| 37 | + self.output: str = None |
| 38 | + self.metatada: Dict[str, Any] = kwargs or {} |
| 39 | + |
| 40 | + def on_llm_start( |
| 41 | + self, serialized: Dict[str, Any], prompts: List[str], **kwargs: Any |
| 42 | + ) -> Any: |
| 43 | + """Run when LLM starts running.""" |
| 44 | + |
| 45 | + def on_chat_model_start( |
| 46 | + self, |
| 47 | + serialized: Dict[str, Any], |
| 48 | + messages: List[List[langchain_schema.BaseMessage]], |
| 49 | + **kwargs: Any, |
| 50 | + ) -> Any: |
| 51 | + """Run when Chat Model starts running.""" |
| 52 | + self.model_parameters = kwargs.get("invocation_params", {}) |
| 53 | + |
| 54 | + provider = self.model_parameters.get("_type", None) |
| 55 | + if provider in LANGCHAIN_TO_OPENLAYER_PROVIDER_MAP: |
| 56 | + self.provider = LANGCHAIN_TO_OPENLAYER_PROVIDER_MAP[provider] |
| 57 | + self.model_parameters.pop("_type") |
| 58 | + |
| 59 | + self.model = self.model_parameters.get("model_name", None) |
| 60 | + self.output = "" |
| 61 | + self.prompt = self._langchain_messages_to_prompt(messages) |
| 62 | + self.start_time = time.time() |
| 63 | + |
| 64 | + @staticmethod |
| 65 | + def _langchain_messages_to_prompt( |
| 66 | + messages: List[List[langchain_schema.BaseMessage]], |
| 67 | + ) -> List[Dict[str, str]]: |
| 68 | + """Converts Langchain messages to the Openlayer prompt format (similar to |
| 69 | + OpenAI's.)""" |
| 70 | + prompt = [] |
| 71 | + for message in messages: |
| 72 | + for m in message: |
| 73 | + if m.type == "human": |
| 74 | + prompt.append({"role": "user", "content": m.content}) |
| 75 | + elif m.type == "system": |
| 76 | + prompt.append({"role": "system", "content": m.content}) |
| 77 | + elif m.type == "ai": |
| 78 | + prompt.append({"role": "assistant", "content": m.content}) |
| 79 | + return prompt |
| 80 | + |
| 81 | + def on_llm_new_token(self, token: str, **kwargs: Any) -> Any: |
| 82 | + """Run on new LLM token. Only available when streaming is enabled.""" |
| 83 | + |
| 84 | + def on_llm_end(self, response: langchain_schema.LLMResult, **kwargs: Any) -> Any: |
| 85 | + """Run when LLM ends running.""" |
| 86 | + self.end_time = time.time() |
| 87 | + self.latency = (self.end_time - self.start_time) * 1000 |
| 88 | + |
| 89 | + if response.llm_output and "token_usage" in response.llm_output: |
| 90 | + self.prompt_tokens = response.llm_output["token_usage"].get( |
| 91 | + "prompt_tokens", 0 |
| 92 | + ) |
| 93 | + self.completion_tokens = response.llm_output["token_usage"].get( |
| 94 | + "completion_tokens", 0 |
| 95 | + ) |
| 96 | + self.cost = self._get_cost_estimate( |
| 97 | + num_input_tokens=self.prompt_tokens, |
| 98 | + num_output_tokens=self.completion_tokens, |
| 99 | + ) |
| 100 | + self.total_tokens = response.llm_output["token_usage"].get( |
| 101 | + "total_tokens", 0 |
| 102 | + ) |
| 103 | + |
| 104 | + for generations in response.generations: |
| 105 | + for generation in generations: |
| 106 | + self.output += generation.text.replace("\n", " ") |
| 107 | + |
| 108 | + self._add_to_trace() |
| 109 | + |
| 110 | + def _get_cost_estimate( |
| 111 | + self, num_input_tokens: int, num_output_tokens: int |
| 112 | + ) -> float: |
| 113 | + """Returns the cost estimate for a given model and number of tokens.""" |
| 114 | + if self.model not in constants.OPENAI_COST_PER_TOKEN: |
| 115 | + return None |
| 116 | + cost_per_token = constants.OPENAI_COST_PER_TOKEN[self.model] |
| 117 | + return ( |
| 118 | + cost_per_token["input"] * num_input_tokens |
| 119 | + + cost_per_token["output"] * num_output_tokens |
| 120 | + ) |
| 121 | + |
| 122 | + def _add_to_trace(self) -> None: |
| 123 | + """Adds to the trace.""" |
| 124 | + name = PROVIDER_TO_STEP_NAME.get(self.provider, "Chat Completion Model") |
| 125 | + tracer.add_openai_chat_completion_step_to_trace( |
| 126 | + name=name, |
| 127 | + provider=self.provider, |
| 128 | + inputs={"prompt": self.prompt}, |
| 129 | + output=self.output, |
| 130 | + cost=self.cost, |
| 131 | + tokens=self.total_tokens, |
| 132 | + latency=self.latency, |
| 133 | + start_time=self.start_time, |
| 134 | + end_time=self.end_time, |
| 135 | + model=self.model, |
| 136 | + model_parameters=self.model_parameters, |
| 137 | + prompt_tokens=self.prompt_tokens, |
| 138 | + completion_tokens=self.completion_tokens, |
| 139 | + metadata=self.metatada, |
| 140 | + ) |
| 141 | + |
| 142 | + def on_llm_error( |
| 143 | + self, error: Union[Exception, KeyboardInterrupt], **kwargs: Any |
| 144 | + ) -> Any: |
| 145 | + """Run when LLM errors.""" |
| 146 | + |
| 147 | + def on_chain_start( |
| 148 | + self, serialized: Dict[str, Any], inputs: Dict[str, Any], **kwargs: Any |
| 149 | + ) -> Any: |
| 150 | + """Run when chain starts running.""" |
| 151 | + |
| 152 | + def on_chain_end(self, outputs: Dict[str, Any], **kwargs: Any) -> Any: |
| 153 | + """Run when chain ends running.""" |
| 154 | + |
| 155 | + def on_chain_error( |
| 156 | + self, error: Union[Exception, KeyboardInterrupt], **kwargs: Any |
| 157 | + ) -> Any: |
| 158 | + """Run when chain errors.""" |
| 159 | + |
| 160 | + def on_tool_start( |
| 161 | + self, serialized: Dict[str, Any], input_str: str, **kwargs: Any |
| 162 | + ) -> Any: |
| 163 | + """Run when tool starts running.""" |
| 164 | + |
| 165 | + def on_tool_end(self, output: str, **kwargs: Any) -> Any: |
| 166 | + """Run when tool ends running.""" |
| 167 | + |
| 168 | + def on_tool_error( |
| 169 | + self, error: Union[Exception, KeyboardInterrupt], **kwargs: Any |
| 170 | + ) -> Any: |
| 171 | + """Run when tool errors.""" |
| 172 | + |
| 173 | + def on_text(self, text: str, **kwargs: Any) -> Any: |
| 174 | + """Run on arbitrary text.""" |
| 175 | + |
| 176 | + def on_agent_action( |
| 177 | + self, action: langchain_schema.AgentAction, **kwargs: Any |
| 178 | + ) -> Any: |
| 179 | + """Run on agent action.""" |
| 180 | + |
| 181 | + def on_agent_finish( |
| 182 | + self, finish: langchain_schema.AgentFinish, **kwargs: Any |
| 183 | + ) -> Any: |
| 184 | + """Run on agent end.""" |
0 commit comments