inference/worker/chat_chain_utils.py - main - githubmirror/Open-Assistant

inference/worker/chat_chain_utils.py
374 строки14 KB
Oliver Stanley
Support ChatML prompt format in worker (#3668)
29 авг 2023, 01:56
29 авг 2023, 01:56d613c81
import json
import re
from typing import Callable

import requests
import transformers
from chat_chain_prompts import INSTRUCTIONS, OBSERVATION_SEQ, TOOLS_PREFIX
from hf_langchain_inference import HFInference
from langchain.agents import Tool
from langchain.memory import ConversationBufferMemory
from langchain.prompts import PromptTemplate
from loguru import logger
from oasst_shared.schemas import inference
from openapi_parser import prepare_plugin_for_llm
from settings import settings
from utils import shared_tokenizer_lock, special_tokens

RESPONSE_MAX_LENGTH = 2048
DESCRIPTION_FOR_MODEL_MAX_LENGTH = 512

llm_json_parser = HFInference(
    inference_server_url=settings.inference_server_url,
    max_new_tokens=512,
    stop_sequences=[special_tokens["end"] if special_tokens["end"] else "</s>"],
    top_k=5,
    temperature=0.20,
    repetition_penalty=(1 / 0.83),
)


# This algo should be fine but possible improvements could be levenshtein or vector distance
def similarity(ts1: str, ts2: str) -> float:
    """Compute Jaro-Winkler distance between two strings."""
    if ts1 == ts2:
        return 1

    match = 0
    len1, len2 = len(ts1), len(ts2)
    max_dist = (max(len1, len2) // 2) - 1

    hash_ts1 = [0] * len1
    hash_ts2 = [0] * len2

    for i in range(len1):
        for j in range(max(0, i - max_dist), min(len2, i + max_dist + 1)):
            if ts1[i] == ts2[j] and hash_ts2[j] == 0:
                hash_ts1[i] = 1
                hash_ts2[j] = 1
                match += 1
                break

    if match == 0:
        return 0

    t = 0
    point = 0

    for i in range(len1):
        if hash_ts1[i] == 1:
            while hash_ts2[point] == 0:
                point += 1
            if ts1[i] != ts2[point]:
                t += 1
            point += 1

    t /= 2
    return (match / len1 + match / len2 + (match - t) / match) / 3.0


def extract_tool_and_input(llm_output: str, ai_prefix: str) -> tuple[str, str]:
    """
    Extract tool name and tool input from LLM output. If LLM chose not to use a tool, `ai_prefix` is returned instead of tool name, and LLM output is returned instead of tool input.
    """
    llm_output = llm_output.strip().replace("```", "")
    if f"{ai_prefix}:" in llm_output:
        # No tool used, return LLM prefix and LLM output
        return ai_prefix, llm_output.split(f"{ai_prefix}:")[-1].strip()

    regex = r"Action: (.*?)[\n]*Action Input:\n?(.*)"
    match = re.search(regex, llm_output, re.MULTILINE | re.DOTALL)
    if not match:
        if OBSERVATION_SEQ in llm_output:
            return ai_prefix, llm_output.split(OBSERVATION_SEQ)[-1].strip()
        return ai_prefix, llm_output

    action = match.group(1)
    action_input = match.group(2)
    return action.strip().replace("'", ""), action_input.strip().strip(" ")


# Truncate, but append closing bracket if string starts with [ or { or (
# Helps prevent LLM from just generating output continuously
def truncate_str(output: str, max_length: int = 1024) -> str:
    if len(output) > max_length:
        if output[0] == "(":
            return output[:max_length] + "...)"
        elif output[0] == "[":
            return output[:max_length] + "...]"
        elif output[0] == "{":
            return output[:max_length] + "...}"
        else:
            return output[:max_length] + "..."
    return output


# Parse JSON and try to fix it if invalid
def prepare_json(json_str: str) -> str:
    json_str = json_str.strip()
    fixed_json = json_str
    try:
        json.loads(json_str)
    except json.decoder.JSONDecodeError:
        # Fix missing quotes around keys and replace Python's True, False, and None
        fixed_json = re.sub(r"(?<=\{|\,)(\s*)(\w+)(\s*):", r'\1"\2"\3:', json_str)
        fixed_json = fixed_json.replace("True", "true").replace("False", "false").replace("None", "null")

        # Remove excessive closing braces/brackets
        brace_count = bracket_count = 0
        result = []
        for c in fixed_json:
            if c == "{":
                brace_count += 1
            elif c == "}":
                brace_count -= 1
            elif c == "[":
                bracket_count += 1
            elif c == "]":
                bracket_count -= 1

            if brace_count >= 0 and bracket_count >= 0:
                result.append(c)
        # Add missing closing braces/brackets
        result.extend(["}"] * brace_count)
        result.extend(["]"] * bracket_count)
        fixed_json = "".join(result)

        try:
            json.loads(fixed_json)
        except json.decoder.JSONDecodeError as e:
            logger.warning(f"JSON is still not valid, trying to fix it with LLM {fixed_json}")
            # If still invalid, try using LLM to fix
            prompt = f"""{special_tokens['prompter']}Below is malformed JSON object string:
--------------
{json_str}
--------------
Parsing error:
--------------
{e}

RULES:
1. If malformed JSON object string contains multiple objects, you merge them into one.
2. You will never made up or add any new data, you will only fix the malformed JSON object string.

Here is the fixed JSON object string:{special_tokens['end'] or '</s>'}{special_tokens['assistant']}"""
            logger.warning(f"JSON Fix Prompt: {prompt}")
            out = llm_json_parser.generate(prompts=[prompt]).generations[0][0].text
            out = out[: out.find("}") + 1]
            logger.warning(f"JSON Fix Output: {out}")
            return out

    return fixed_json


def select_tool(tool_name: str, tools: list[Tool]) -> Tool | None:
    tool = next((t for t in tools if t.name in tool_name), None)
    if tool:
        return tool
    tool, tool_similarity = max(
        ((t, similarity(t.name, tool_name)) for t in tools),
        key=lambda x: x[1],
        default=(None, 0),
    )
    # TODO: make stricter with better models
    if tool and tool_similarity > 0.75:
        return tool
    return None


def use_tool(tool_name: str, tool_input: str, tools: list[Tool]) -> str:
    tool = select_tool(tool_name, tools)
    if not tool:
        return f"ERROR! {tool_name} is not a valid tool. Try again with different tool!"
    prepared_input = prepare_json(tool_input)
    tool_output = tool.func(prepared_input)
    return tool_output


# Needs more work for errors, error-prompt tweaks are currently based on
# `OpenAssistant/oasst-sft-6-llama-30b-epoch-1 model`
class RequestsForLLM:
    def run(self, params: str, url: str, param_location: str, type: str, payload: str | None = None) -> str:
        return self.run_request(params, url, param_location, type, payload)

    def run_request(self, params: str, url: str, param_location: str, type: str, payload: str = None) -> str:
        try:
            query_params = params
            if param_location == "path":
                for key, value in query_params.items():
                    url = url.replace(f"{{{key}}}", value)
                query_params = {}

            headers = {"Content-Type": "application/json"} if payload else None

            if type.lower() == "get":
                logger.info(
                    f"Running {type.upper()} request on {url} with\nparams: {params}\nparam_location: {param_location}\npayload: {payload}"
                )
                res = requests.get(url, params=query_params, headers=headers)
            elif type.lower() == "post":
                # if model did not generate payload object, use params as payload
                data = json.dumps(payload) if payload else json.dumps(params)
                logger.info(
                    f"Running {type.upper()} request on {url} with\nparams: {params}\nparam_location: {param_location}\npayload: {data}"
                )
                res = requests.post(url, params=query_params, data=data, headers=headers)
            else:
                return f"ERROR! Unsupported request type: {type}. Only GET and POST are supported. Try again!"

            return self.process_response(res)
        except Exception as e:
            return f"ERROR! That didn't work, try modifying Action Input.\n{e}. Try again!"

    def process_response(self, res: requests.Response) -> str:
        logger.info(f"Request response: {res.text}")
        if res.status_code != 200:
            return f"ERROR! Please modify Action Input. according to this error message: \n{res.text}. Try again!"

        if res.text is None or len(res.text) == 0:
            return "ERROR! That didn't work, try modifying Action Input.\nEmpty response. Try again!"

        if "null" in res.text.lower() and len(res.text) < 10:
            return "ERROR! That didn't work, try modifying Action Input.\nEmpty response. Try again!"

        return truncate_str(res.text, RESPONSE_MAX_LENGTH)


def compose_tools_from_plugin(plugin: inference.PluginEntry | None) -> tuple[str, list[Tool]]:
    if not plugin:
        return "", []

    llm_plugin: inference.PluginConfig = prepare_plugin_for_llm(plugin.url)
    if not llm_plugin:
        return "", []

    tools = []
    request_tool = RequestsForLLM()

    def create_tool_func(endpoint: inference.PluginOpenAPIEndpoint, param_location: str) -> Callable[..., str]:
        def func(req) -> str:
            try:
                json_obj = json.loads(req)
                request = json_obj.get("request", {})
                params = request.get("params", {})
                payload = request.get("payload", None)
            except json.JSONDecodeError:
                print("Error: Invalid JSON input")
                request, params, payload = {}, {}, None
            except Exception as e:
                print(f"Error: {e}")
                request, params, payload = {}, {}, None

            return request_tool.run(
                url=endpoint.url, params=params, param_location=param_location, type=endpoint.type, payload=payload
            )

        return func

    # Generate tool for each plugin endpoint. Helps LLM use tools as it does not choose API server URL etc on its own
    # LLM only chooses endpoint, parameters and values to use. Modifying this can degrade or improve tool usage
    for endpoint in llm_plugin.endpoints:
        params = "\n\n".join(
            [
                f""" name: "{param.name}",\n in: "{param.in_}",\n description: "{truncate_str(param.description, 128)}",\n schema: {param.schema_},\n required: {param.required}"""
                for param in endpoint.params
            ]
        )

        # LangChain uses {input_name} for templating
        # Some plugins can have {some_word} in their description
        params = params.replace("{", "{{").replace("}", "}}")
        payload_description = ""
        if endpoint.payload:
            try:
                payload_description = "payload: " + truncate_str(json.dumps(endpoint.payload, indent=4), 256)
                payload_description = payload_description.replace("{", "{{").replace("}", "}}")
            except Exception as e:
                logger.warning(f"Failed to convert payload to json string: {e}")

        payload_description += "" if not payload_description or payload_description.endswith("\n") else "\n"
        if len(payload_description) > 0:
            payload_description = "\n" + payload_description + "\n"

        parameters_description = f"params:\n{params}\n" if params else "\n"

        openapi_specification_title = (
            "\nOpenAPI specification\n" if len(payload_description) > 0 or len(params) > 0 else ""
        )

        param_location = endpoint.params[0].in_ if len(endpoint.params) > 0 else "query"

        # If plugin has no operation_id, use path as fallback
        path = endpoint.path[1:] if endpoint.path and len(endpoint.path) > 0 else endpoint.path

        tool = Tool(
            name=endpoint.operation_id if endpoint.operation_id != "" else path,
            # Could be path, e.g /api/v1/endpoint but can lead LLM to invent URLs
            # Problem with EP description is that it is too long for some plugins
            func=create_tool_func(endpoint, param_location),
            description=f"{openapi_specification_title}{parameters_description}{payload_description}tool description: {endpoint.summary}\n",
        )
        tools.append(tool)

    tools_string = "\n".join([f"> {tool.name}{tool.description}" for tool in tools])
    # This can be long for some plugins, we need to truncate due to ctx limitations
    plugin_description_for_model = truncate_str(llm_plugin.description_for_model, DESCRIPTION_FOR_MODEL_MAX_LENGTH)
    return (
        f"{TOOLS_PREFIX}{tools_string}\n\n{llm_plugin.name_for_model} plugin description:\n{plugin_description_for_model}\n\n{INSTRUCTIONS}",
        tools,
    )


def prepare_prompt(
    input_prompt: str,
    prompt_template: PromptTemplate,
    memory: ConversationBufferMemory,
    tools_names: list[str] | None,
    current_time: str,
    language: str,
    tokenizer: transformers.PreTrainedTokenizer,
    worker_config: inference.WorkerConfig,
    action_input_format: str,
    custom_instructions: str = "",
) -> str:
    max_input_length = worker_config.model_config.max_input_length

    args = {
        "input": input_prompt,
        "language": language,
        "current_time": current_time,
        "chat_history": memory.buffer,
        "custom_instructions": custom_instructions,
    }

    if tools_names:
        args["tools_names"] = tools_names
        args["action_input_format"] = action_input_format

    out_prompt = prompt_template.format(**args)

    with shared_tokenizer_lock:
        ids = tokenizer.encode(out_prompt)

    # soft truncation (delete whole messages)
    while len(ids) > max_input_length and len(memory.chat_memory.messages) > 0:
        memory.chat_memory.messages.pop(0)
        args = {
            "input": input_prompt,
            "language": language,
            "current_time": current_time,
            "chat_history": memory.buffer,
            "custom_instructions": custom_instructions,
        }

        if tools_names:
            args["tools_names"] = tools_names
            args["action_input_format"] = action_input_format

        out_prompt = prompt_template.format(**args)

        with shared_tokenizer_lock:
            ids = tokenizer.encode(out_prompt)
        logger.warning(f"Prompt too long, deleting chat history. New length: {len(ids)}")

    return out_prompt