chains.py

import asyncio
import openai

from langchain.chat_models import ChatOpenAI
from langchain.chat_models.openai import _create_retry_decorator
# from langchain.llms.base import create_base_retry_decorator
from langchain.prompts.chat import ChatPromptTemplate
from langchain.schema import BaseOutputParser
from application import *

from utility import read_pdf,aterminal_print

llm = ChatOpenAI(
    temperature=0.0,
    model_name="gpt-3.5-turbo-16k",
    openai_api_key=openai.api_key,
    max_retries=2,
    timeout=30 # this one may help me to solve all the old problems...
    )

# retry_decorator = _create_retry_decorator(llm)

class Replacement(BaseOutputParser):
    """Parse the output of an LLM call to a comma-separated list."""
    def parse(self, text: str, **kwargs):
        """Parse the output of an LLM call."""
        if kwargs:
            print(kwargs)
        return text.strip().split(", ")

# @retry_decorator
@aterminal_print
async def async_generate(article,name,chain,input_variables={}):
    try:
        res = await chain.ainvoke(input_variables)
    except Exception as e:
        print("API Error",str(e))
        article[name] = f"API Error: {str(e)}"
        return
    
    print("completed",name)
    article[name] = res.content


@aterminal_print
async def execute_concurrent(article,prompts):
    tasks = []

    prompt_list = list(prompts.keys())

    while prompt_list:
        name = prompt_list.pop(0)
        p = prompts[name]

        # recurse on the missing tables until all input condiitons are met
        missing_inputs = [s for s in p["inputs"] if s not in article]
        for x in missing_inputs:
            await execute_concurrent(article,{x:app_data["prompts"][x]})

        print("executing",p["assessment"],name)
        tasks.append(gen_task(p,article))

    await asyncio.gather(*tasks)

def replace_term(res,**kwargs):
    if "map" in kwargs:
        for key,term in kwargs["map"].items():
            res.content = res.content.replace(key,term)
    return res

post_prompt_maping = {}
post_replace_term = lambda res,map=post_prompt_maping:replace_term(res,map=map)

def gen_task(prompt,article):
    chat_prompt = gen_chat_prompt(prompt,article)
    chain = chat_prompt | llm | post_replace_term
    input_variables = gen_input_variables (chat_prompt,prompt)

    return async_generate(article=article,name=prompt["name"],chain=chain,input_variables=input_variables)

def gen_chat_prompt(prompt,article):
    input_text = "".join([article[s] for s in prompt["inputs"]])
    
    messages = [
        ("system","You are a helpful AI that can answer questions about clinical trail and operation studies."),
        ("human",input_text)
    ]

    if len(prompt["chain"]) > 1:
        for i in article["logic"]["chain id"]:
            messages.append(("human",prompt["chain"][i]))
    else:
        messages.append(("human",prompt["chain"][0]))
    # messages.append(("system",reformat_inst))
    
    return ChatPromptTemplate.from_messages(messages=messages)

def gen_input_variables(chat_prompt,prompt):
    input_variables = {}
    if "term" in prompt:
        app_data["current"]["term"] = prompt["term"][0]["prompting_term"]
    
    if "n_col" in chat_prompt.input_variables:
        input_variables["n_col"] = len(prompt["chain"])
    if "term" in chat_prompt.input_variables:
        input_variables["term"] = app_data["current"]["term"]
    return input_variables

if __name__ == "__main__":
    # lets try the Blood Loss, Operation Time, and Need for ICU in other folder
    sample_artice = ".samples/Ha SK, 2008.pdf"
    sample_content,_ = read_pdf(sample_artice)

    with open(".prompts/other/Operation Time.txt") as f:
        prompt = f.read()
        name = "Operation Time"
    
    post_prompt_maping = {}
    post_replace_term = lambda res,map=post_prompt_maping:replace_term(res,map=map)

    chat_prompt = ChatPromptTemplate.from_messages([
        ("human",sample_artice),
        ("system",prompt),
    ])

    # experiment with cascading the chain
    chain = chat_prompt | llm 
    chain2 = chain | post_replace_term

    # lets try remove from chain
    chain2.last.with_retry = True
    res = chain2.invoke({"term":name})
    print(res.content)