ChatAbstractions

This repo is a collection of chat model abstractions that demonstrates how to wrap (subclass) LangChain's BaseChatModel in order to add functionality to a chain without breaking existing chat model interfaces. The use cases for wrapping chat models in this manner are mostly focused on dynamic model selection. However, other use cases are possible as well.

Subclassing BaseChatModel requires implementing 2 methods: _llm_type() and _generate().

from typing import Any, List, Optional

from langchain.callbacks.manager import CallbackManagerForLLMRun
from langchain.chat_models.base import BaseChatModel
from langchain.schema import ChatResult
from langchain.schema.messages import BaseMessage


class ChatSubclass(BaseChatModel):

    @property
    def _llm_type(self) -> str:
        """Return type of chat model."""
        raise NotImplementedError

    def _generate(
        self,
        messages: List[BaseMessage],
        stop: Optional[List[str]] = None,
        run_manager: Optional[CallbackManagerForLLMRun] = None,
        **kwargs: Any,
    ) -> ChatResult:
        """Add custom logic here."""
        raise NotImplementedError

ChatDynamic

The implementation of ChatDynamic demonstrates the ability to select a chat model at runtime based on environment variable configuration. In the event of an outage or degraded performance by an LLM provider, this functionality (i.e. failover) may be desirable.

# set environment variable DYNAMIC_CHAT_MODEL_ID=gpt-4

# initialize chat models
gpt_4_model = ChatOpenAI(model="gpt-4")
gpt_3_5_model = ChatOpenAI(model="gpt-3.5-turbo")

# specify all models that can be selected in the ChatDynamic instance
chat_dynamic_model = ChatDynamic(
    models={
        "gpt-4": gpt_4_model,
        "gpt-3_5": gpt_3_5_model,
    },
    default_model="gpt-4",
)

Reading: Dynamic Failover and Load Balancing LLMs With LangChain

ChatLoadBalance

The implementation of ChatLoadBalance demonstrates the ability to select a method of load balancing (random, round robin, least rate limited) between LLM models. In the event of rate limiting or peak usage times, this functionality may be desirable.

# initialize chat models
gpt_4_model = ChatOpenAI(model="gpt-4")
gpt_3_5_model = ChatOpenAI(model="gpt-3.5-turbo")

# specify all models that can be selected in the ChatLoadBalance instance
chat_load_balance_model = ChatLoadBalance(
    models=[gpt_4_model, gpt_3_5_model],
    load_balance_type=1,  # 0 - random, 1 - round robin, 2 - least rate limited
)

Reading: Dynamic Failover and Load Balancing LLMs With LangChain

ChatChaos

The implementation of ChatChaos demonstrates the ability to substitute normal LLM behavior with chaotic behavior. The purpose of this abstraction is to promote the Principles of Chaos Engineering in the context of LLM applications. This abstraction is inspired by Netflix's Chaos Monkey.

# initialize chat model
gpt_3_5_model = ChatOpenAI(model="gpt-3.5-turbo")

# configure ChatChaos
chat_chaos_model = ChatChaos(
    model=gpt_3_5_model,
    enabled=True,
    cron=croniter("0 * * * *"),
    duration_mins=60,
    ratio=1.0,
    enable_malformed_json=False,
    enable_hallucination=True,
    enable_latency=False,
    hallucination_prompt="Write a poem about the Python programming language.",
)

Reading: ChatChaos: The Good, the Bad, and the Ugly

ChatNotDiamond

The implementation of ChatNotDiamond demonstrates the abi