|
| 1 | +from datetime import datetime, timezone |
| 2 | +from enum import Enum |
| 3 | +from pydantic import BaseModel, Field |
| 4 | +from typing import Any, Dict, List, Optional, Union |
| 5 | +from uuid import UUID |
| 6 | + |
| 7 | +import httpx |
| 8 | + |
| 9 | +try: |
| 10 | + from langchain.callbacks.base import BaseCallbackHandler # type: ignore |
| 11 | + from langchain_core.messages import BaseMessage, HumanMessage, SystemMessage # type: ignore |
| 12 | + from langchain_core.outputs import LLMResult # type: ignore |
| 13 | +except ImportError as exc: |
| 14 | + raise ModuleNotFoundError("Please install langchain to use the orq.ai langchain native integration: 'pip install langchain'") from exc |
| 15 | + |
| 16 | +def get_iso_string(): |
| 17 | + # Get current datetime in UTC, timezone-aware |
| 18 | + current_utc_datetime = datetime.now(timezone.utc) |
| 19 | + # Format it to ISO 8601 string with 'Z' indicating UTC |
| 20 | + return current_utc_datetime.isoformat(timespec="milliseconds").replace( |
| 21 | + "+00:00", "Z" |
| 22 | + ) |
| 23 | + |
| 24 | +class EventType(str, Enum): |
| 25 | + LLM = "llm" |
| 26 | + |
| 27 | +class LlmRole(str, Enum): |
| 28 | + SYSTEM = "system" |
| 29 | + USER = "user" |
| 30 | + |
| 31 | +class LlmUsage(BaseModel): |
| 32 | + input_tokens: int |
| 33 | + output_tokens: int |
| 34 | + |
| 35 | +class ChoiceMessage(BaseModel): |
| 36 | + content: Union[str, List[str]] |
| 37 | + role: LlmRole |
| 38 | + |
| 39 | +class Choice(BaseModel): |
| 40 | + index: int |
| 41 | + message: ChoiceMessage |
| 42 | + finish_reason: Optional[str] = None |
| 43 | + |
| 44 | +class LlmEvent(BaseModel): |
| 45 | + type: EventType |
| 46 | + run_id: str |
| 47 | + parameters: Optional[dict] = {} |
| 48 | + prompts: Optional[List[str]] = [] |
| 49 | + messages: List[ChoiceMessage] = [] |
| 50 | + start_timestamp: str = Field(default_factory=get_iso_string) |
| 51 | + end_timestamp: Optional[str] = None |
| 52 | + response_choices: List[Choice] = [] |
| 53 | + usage: Optional[LlmUsage] = None |
| 54 | + |
| 55 | +class OrqClient(): |
| 56 | + def __init__(self, api_key: str, api_url: str): |
| 57 | + self.api_key = api_key |
| 58 | + self.api_url = api_url |
| 59 | + |
| 60 | + def log_event(self, event: LlmEvent): |
| 61 | + headers = { |
| 62 | + "Authorization": f"Bearer {self.api_key}" |
| 63 | + } |
| 64 | + |
| 65 | + response = httpx.post(f"{self.api_url}/v2/traces/langchain", headers=headers, json=event.model_dump()) |
| 66 | + |
| 67 | +class OrqLangchainCallback(BaseCallbackHandler): |
| 68 | + """Base callback handler that can be used to handle callbacks from langchain.""" |
| 69 | + |
| 70 | + def __init__(self, api_key: str, api_url = "https://my.orq.ai"): |
| 71 | + self.events: Dict[str, LlmEvent] = {} |
| 72 | + self.orq_client = OrqClient(api_key, api_url) |
| 73 | + |
| 74 | + def on_llm_start( |
| 75 | + self, |
| 76 | + serialized: Dict[str, Any], |
| 77 | + prompts: List[str], |
| 78 | + *, |
| 79 | + run_id: UUID, |
| 80 | + metadata: Optional[Dict[str, Any]] = None, |
| 81 | + **kwargs: Any, |
| 82 | + ) -> Any: |
| 83 | + self.events[str(run_id)] = LlmEvent(type=EventType.LLM, parameters={ |
| 84 | + "serialized": serialized, |
| 85 | + "metadata": metadata, |
| 86 | + "kwargs": kwargs, |
| 87 | + }, prompts=prompts, run_id=str(run_id)) |
| 88 | + |
| 89 | + def on_chat_model_start( |
| 90 | + self, |
| 91 | + serialized: Dict[str, Any], |
| 92 | + messages: List[List[BaseMessage]], |
| 93 | + *, |
| 94 | + run_id: UUID, |
| 95 | + metadata: Optional[Dict[str, Any]] = None, |
| 96 | + **kwargs: Any |
| 97 | + ) -> Any: |
| 98 | + normalize_messages: List[ChoiceMessage] = [] |
| 99 | + |
| 100 | + for root_messages in messages: |
| 101 | + for message in root_messages: |
| 102 | + if isinstance(message, HumanMessage): |
| 103 | + normalize_messages.append(ChoiceMessage(role=LlmRole.USER, content=message.content)) |
| 104 | + elif isinstance(message, SystemMessage): |
| 105 | + normalize_messages.append(ChoiceMessage(role=LlmRole.SYSTEM, content=message.content)) |
| 106 | + |
| 107 | + self.events[str(run_id)] = LlmEvent(type=EventType.LLM, parameters={ |
| 108 | + "serialized": serialized, |
| 109 | + "metadata": metadata, |
| 110 | + "kwargs": kwargs, |
| 111 | + }, messages=normalize_messages, run_id=str(run_id)) |
| 112 | + |
| 113 | + def on_llm_end( |
| 114 | + self, |
| 115 | + response: LLMResult, |
| 116 | + *, |
| 117 | + run_id: UUID, |
| 118 | + parent_run_id: Optional[UUID] = None, |
| 119 | + **kwargs: Any, |
| 120 | + ) -> Any: |
| 121 | + event: LlmEvent = self.events[str(run_id)] |
| 122 | + event.end_timestamp = get_iso_string() |
| 123 | + token_usage = response.llm_output['token_usage'] |
| 124 | + event.usage = LlmUsage(input_tokens=token_usage['prompt_tokens'], output_tokens=token_usage['completion_tokens']) |
| 125 | + event.response_choices = [] |
| 126 | + |
| 127 | + for index, choice in enumerate(response.generations[0]): |
| 128 | + event.response_choices.append(Choice(index=index, message=ChoiceMessage(role=LlmRole.SYSTEM, content=choice.text), finish_reason=choice.generation_info['finish_reason'])) |
| 129 | + |
| 130 | + self.orq_client.log_event(event) |
| 131 | + |
| 132 | +__all__ = [ |
| 133 | + "OrqLangchainCallback" |
| 134 | +] |
0 commit comments