shishirpatil--gorilla/berkeley-function-call-leaderboard/bfcl_eval/model_handler/api_inference/mistral.py

import json
import os
import time
from typing import Any

from bfcl_eval.constants.enums import ModelStyle
from bfcl_eval.constants.type_mappings import GORILLA_TO_OPENAPI
from bfcl_eval.model_handler.base_handler import BaseHandler
from bfcl_eval.model_handler.utils import (
    convert_to_function_call,
    convert_to_tool,
    default_decode_ast_prompting,
    default_decode_execute_prompting,
    format_execution_results_prompting,
    retry_with_backoff,
    system_prompt_pre_processing_chat_model,
)
from mistralai import Mistral


class MistralHandler(BaseHandler):
    def __init__(
        self,
        model_name,
        temperature,
        registry_name,
        is_fc_model,
        **kwargs,
    ) -> None:
        super().__init__(model_name, temperature, registry_name, is_fc_model, **kwargs)
        self.model_style = ModelStyle.MISTRAL

        self.client = Mistral(api_key=os.getenv("MISTRAL_API_KEY"))

    def decode_ast(self, result, language, has_tool_call_tag):
        if self.is_fc_model:
            decoded_output = []
            for invoked_function in result:
                name = list(invoked_function.keys())[0]
                params = json.loads(invoked_function[name])
                decoded_output.append({name: params})
            return decoded_output
        else:
            return default_decode_ast_prompting(result, language, has_tool_call_tag)

    def decode_execute(self, result, has_tool_call_tag):
        if self.is_fc_model:
            function_call = convert_to_function_call(result)
            return function_call
        else:
            return default_decode_execute_prompting(result, has_tool_call_tag)

    @retry_with_backoff(error_message_pattern=r".*Status 429.*")
    def generate_with_backoff(self, **kwargs):
        start_time = time.time()
        api_response = self.client.chat.complete(**kwargs)
        end_time = time.time()

        return api_response, end_time - start_time

    #### FC methods ####

    def _query_FC(self, inference_data: dict):
        message = inference_data["message"]
        tool = inference_data["tools"]
        inference_data["inference_input_log"] = {
            "message": message,
            "tools": tool,
        }

        return self.generate_with_backoff(
            model=self.model_name,
            messages=message,
            tools=tool,
            temperature=self.temperature,
        )

    def _pre_query_processing_FC(self, inference_data: dict, test_entry: dict) -> dict:
        inference_data["message"] = []
        return inference_data

    def _compile_tools(self, inference_data: dict, test_entry: dict) -> dict:
        functions: list = test_entry["function"]

        tools = convert_to_tool(functions, GORILLA_TO_OPENAPI, self.model_style)

        inference_data["tools"] = tools

        return inference_data

    def _parse_query_response_FC(self, api_response: Any) -> dict:
        try:
            model_responses = [
                {func_call.function.name: func_call.function.arguments}
                for func_call in api_response.choices[0].message.tool_calls
            ]
            tool_call_func_names = [
                func_call.function.name
                for func_call in api_response.choices[0].message.tool_calls
            ]
            tool_call_ids = [
                func_call.id for func_call in api_response.choices[0].message.tool_calls
            ]
        except:
            model_responses = api_response.choices[0].message.content
            tool_call_func_names = []
            tool_call_ids = []

        return {
            "model_responses": model_responses,
            "model_responses_message_for_chat_history": api_response.choices[0].message,
            "tool_call_func_names": tool_call_func_names,
            "tool_call_ids": tool_call_ids,
            "input_token": api_response.usage.prompt_tokens,
            "output_token": api_response.usage.completion_tokens,
        }

    def add_first_turn_message_FC(
        self, inference_data: dict, first_turn_message: list[dict]
    ) -> dict:
        inference_data["message"].extend(first_turn_message)
        return inference_data

    def _add_next_turn_user_message_FC(
        self, inference_data: dict, user_message: list[dict]
    ) -> dict:
        inference_data["message"].extend(user_message)
        return inference_data

    def _add_assistant_message_FC(
        self, inference_data: dict, model_response_data: dict
    ) -> dict:
        inference_data["message"].append(
            model_response_data["model_responses_message_for_chat_history"]
        )
        return inference_data

    def _add_execution_results_FC(
        self, inference_data: dict, execution_results: list[str], model_response_data: dict
    ) -> dict:
        for execution_result, func_name, tool_call_id in zip(
            execution_results,
            model_response_data["tool_call_func_names"],
            model_response_data["tool_call_ids"],
        ):
            tool_message = {
                "role": "tool",
                "name": func_name,
                "content": execution_result,
                "tool_call_id": tool_call_id,
            }
            inference_data["message"].append(tool_message)
        return inference_data

    #### Prompting methods ####

    def _query_prompting(self, inference_data: dict):
        message = inference_data["message"]
        inference_data["inference_input_log"] = {"message": message}

        return self.generate_with_backoff(
            model=self.model_name,
            messages=message,
            temperature=self.temperature,
        )

    def _pre_query_processing_prompting(self, test_entry: dict) -> dict:
        functions: list = test_entry["function"]
        test_entry_id: str = test_entry["id"]

        test_entry["question"][0] = system_prompt_pre_processing_chat_model(
            test_entry["question"][0], functions, test_entry_id
        )

        return {"message": []}

    def _parse_query_response_prompting(self, api_response: Any) -> dict:
        return {
            "model_responses": api_response.choices[0].message.content,
            "model_responses_message_for_chat_history": api_response.choices[0].message,
            "input_token": api_response.usage.prompt_tokens,
            "output_token": api_response.usage.completion_tokens,
        }

    def add_first_turn_message_prompting(
        self, inference_data: dict, first_turn_message: list[dict]
    ) -> dict:
        inference_data["message"].extend(first_turn_message)
        return inference_data

    def _add_next_turn_user_message_prompting(
        self, inference_data: dict, user_message: list[dict]
    ) -> dict:
        inference_data["message"].extend(user_message)
        return inference_data

    def _add_assistant_message_prompting(
        self, inference_data: dict, model_response_data: dict
    ) -> dict:
        inference_data["message"].append(
            model_response_data["model_responses_message_for_chat_history"]
        )
        return inference_data

    def _add_execution_results_prompting(
        self, inference_data: dict, execution_results: list[str], model_response_data: dict
    ) -> dict:
        formatted_results_message = format_execution_results_prompting(
            inference_data, execution_results, model_response_data
        )
        inference_data["message"].append(
            {"role": "user", "content": formatted_results_message}
        )

        return inference_data