app.py

import gradio as gr
import urllib.request
import requests
import bs4
import lxml
import os
#import subprocess
from huggingface_hub import InferenceClient,HfApi
import random
import json
import datetime
import dl
#from query import tasks
from prompts import (
    FINDER,
    MAIN_PROMPT,
    READ_FILE_CODE,
    COMPRESS_HISTORY_PROMPT,
    COMPRESS_DATA_PROMPT,
    COMPRESS_DATA_PROMPT_SMALL,
    LOG_PROMPT,
    LOG_RESPONSE,
    PREFIX,
    TASK_PROMPT,
)
api=HfApi()



client = InferenceClient(
    "mistralai/Mixtral-8x7B-Instruct-v0.1"
)

def parse_action(string: str):
    print("PARSING:")
    print(string)
    assert string.startswith("action:")
    idx = string.find("action_input=")
    print(idx)
    if idx == -1:
        print ("idx == -1")
        print (string[8:])
        return string[8:], None

    print ("last return:")
    print (string[8 : idx - 1])
    print (string[idx + 13 :].strip("'").strip('"'))
    return string[8 : idx - 1], string[idx + 13 :].strip("'").strip('"')



VERBOSE = True
MAX_HISTORY = 100
MAX_DATA = 20000

def format_prompt(message, history):
  prompt = "<s>"
  for user_prompt, bot_response in history:
    prompt += f"[INST] {user_prompt} [/INST]"
    prompt += f" {bot_response}</s> "
  prompt += f"[INST] {message} [/INST]"
  return prompt


def run_gpt(
    prompt_template,
    stop_tokens,
    max_tokens,
    seed,
    purpose,
    files,
    **prompt_kwargs,
):
    timestamp=datetime.datetime.now()

    print(seed)
    generate_kwargs = dict(
        temperature=0.9,
        max_new_tokens=max_tokens,
        top_p=0.95,
        repetition_penalty=1.0,
        do_sample=True,
        seed=seed,
    )
    
    content = PREFIX.format(
        timestamp=timestamp,
        purpose=purpose,
        files=files,
    ) + prompt_template.format(**prompt_kwargs)
    if VERBOSE:
        print(LOG_PROMPT.format(content))
    
    
    #formatted_prompt = format_prompt(f"{system_prompt}, {prompt}", history)
    #formatted_prompt = format_prompt(f'{content}', **prompt_kwargs['history'])

    stream = client.text_generation(content, **generate_kwargs, stream=True, details=True, return_full_text=False)
    resp = ""
    for response in stream:
        resp += response.token.text
        #yield resp

    if VERBOSE:
        print(LOG_RESPONSE.format(resp))
    return resp

def compress_data(c,purpose, task, history, result,repo,space,file_name):
    seed=random.randint(1,1000000000)
    
    print (c)
    #tot=len(purpose)
    #print(tot)
    divr=int(c)/MAX_DATA
    divi=int(divr)+1 if divr != int(divr) else int(divr)
    chunk = int(int(c)/divr)
    print(f'chunk:: {chunk}')
    print(f'divr:: {divr}')
    print (f'divi:: {divi}')
    out = []
    #out=""
    s=0
    e=chunk
    print(f'e:: {e}')
    new_history=""
    task = f'Compile this data to fulfill the task: {task}, and complete the purpose: {purpose}\n'
    for z in range(divi):
        print(f's:e :: {s}:{e}')
        
        hist = history[s:e]
        
        resp = run_gpt(
            COMPRESS_DATA_PROMPT,
            stop_tokens=["observation:", "task:", "action:", "thought:"],
            max_tokens=2048,
            seed=seed,
            purpose=purpose,
            files=file_name,
            task=task,
            knowledge=new_history,
            history=hist,
        ).strip('\n')
        new_history = resp
        print (resp)
        out+=resp
        e=e+chunk
        s=s+chunk
    '''
    resp = run_gpt(
        COMPRESS_DATA_PROMPT,
        stop_tokens=["observation:", "task:", "action:", "thought:"],
        max_tokens=2048,
        seed=seed,
        purpose=purpose,
        task=task,
        knowledge=new_history,
        history=result,
    )
    '''
    print ("final" + resp)
    history = "result: {}\n".format(resp)
    return history

def read_code(purpose,task,history,action_input,result,repo,space,file_name):
    print("WORKING ON CODE")
    seed=random.randint(1,1000000000)
    out=dl.show_file_content(repo,space,action_input)
    out = str(out)
    rl = len(out)
    print(f'rl:: {rl}')
    c=0
    for i in str(out):
        if i == " " or i=="," or i=="\n" or i=="/" or i=="." or i=="<":
            c +=1  
    print (c)
    #tot=len(purpose)
    #print(tot)
    divr=int(c)/MAX_DATA
    divi=int(divr)+1 if divr != int(divr) else int(divr)
    chunk = int(int(c)/divr)
    print(f'chunk:: {chunk}')
    print(f'divr:: {divr}')
    print (f'divi:: {divi}')
    s=0
    e=chunk
    print(f'e:: {e}')
    new_history=""
    task = f'Compile this data to fulfill the task: {task}, and complete the purpose: {purpose}\n'
    for z in range(divi):
        print(f's:e :: {s}:{e}')
        hist = out[s:e]
        
        resp = run_gpt(
            READ_FILE_CODE,
            stop_tokens=["observation:", "task:", "action:", "thought:"],
            max_tokens=4096,
            seed=seed,
            purpose=purpose,
            files=file_name,
            task=task,
            file_name=action_input,
            file_contents=hist,
        ).strip('\n')
        new_history = resp
        print (resp)
        out+=resp
        e=e+chunk-1000
        s=s+chunk-1000
        history += f'observation: the new code is: {resp}'
        result += f'\n{resp}\n'
    return "MAIN", None, history, task, result


def compress_history(purpose, task, history,file_name):
    resp = run_gpt(
        COMPRESS_HISTORY_PROMPT,
        stop_tokens=["observation:", "task:", "action:", "thought:"],
        max_tokens=1024,
        seed=random.randint(1,1000000000),
        purpose=purpose,
        files=file_name,
        task=task,
        history=history,
    )
    history = "observation: {}\n".format(resp)
    return history


def call_main(purpose, task, history, action_input, result,repo,space,file_name):
    '''
    out=dl.show_file_content(repo,space,action_input)
    resp = run_gpt(
        MAIN_PROMPT,
        stop_tokens=["observation:", "task:"],
        max_tokens=1024,
        seed=random.randint(1,1000000000),
        purpose=purpose,
        files=file_name,
        task=task,
        history=history,
        file_name=action_input,
        file_contents=out,
    )
    '''
    
    resp = run_gpt(
        MAIN_PROMPT,
        stop_tokens=["observation:", "task:"],
        max_tokens=1024,
        seed=random.randint(1,1000000000),
        purpose=purpose,
        files=file_name,
        task=task,
        history=history,
    )
    lines = resp.strip().strip("\n").split("\n")
    #history=""
    for line in lines:
        if line == "":
            continue
        if line.startswith("thought: "):
            history += "{}\n".format(line)
            
        if line.startswith("action: "):
            action_name, action_input = parse_action(line)
            print(f'ACTION::{action_name} -- INPUT :: {action_input}')
            #history += "{}\n".format(line)
            return action_name, action_input, history, task, result
        else:
            pass
            #history += "{}\n".format(line)
            #assert False, "unknown action: {}".format(line)
            #return "UPDATE-TASK", None, history, task
    if "VERBOSE":
        print(history)
    return "MAIN", None, history, task, result


def call_set_task(purpose, task, history, action_input, result,repo,space,file_name):
    task = run_gpt(
        TASK_PROMPT,
        stop_tokens=[],
        max_tokens=1024,
        seed=random.randint(1,1000000000),
        purpose=purpose,
        files=file_name,
        task=task,
        history=history,
    ).strip("\n")
    history += "observation: task has been updated to: {}\n".format(task)
    return "MAIN", None, history, task, result



###########################################################
def search_all(url):
    source=""
    return source



def find_all(purpose,task,history, url, result,repo,space,file_name):
    return_list=[]
    print (url)
    print (f"trying URL:: {url}")        
    try:
        if url != "" and url != None:    
            out = []
            source = requests.get(url)
            if source.status_code ==200:
                soup = bs4.BeautifulSoup(source.content,'lxml')
               
                rawp=(f'RAW TEXT RETURNED: {soup.text}')
                cnt=0
                cnt+=len(rawp)
                out.append(rawp)
                out.append("HTML fragments: ")
                q=("a","p","span","content","article")
                for p in soup.find_all("a"):
                    out.append([{"LINK TITLE":p.get('title'),"URL":p.get('href'),"STRING":p.string}])
                c=0
                out = str(out)
                rl = len(out)
                print(f'rl:: {rl}')
                for i in str(out):
                    if i == " " or i=="," or i=="\n" or i=="/" or i=="." or i=="<":
                        c +=1
                print (f'c:: {c}')
                if c > MAX_HISTORY:
                    print("compressing...")
                    rawp = compress_data(c,purpose,task,out,result,repo,space,file_name)  
                    result += rawp
                else:
                    rawp = out 
                    
                #print (rawp)
                #print (f'out:: {out}')
                history += "observation: the search results are:\n {}\n".format(rawp)
                task = "compile report or complete?"
                return "MAIN", None, history, task, result
            else:
                history += f"observation: That URL string returned an error: {source.status_code}, I should try a different URL string\n"
                #result="Still Working..."
                return "MAIN", None, history, task, result
            
        else: 
            history += "observation: An Error occured\nI need to trigger a search using the following syntax:\naction: SCRAPE_WEBSITE action_input=URL\n"
            return "MAIN", None, history, task, result
    except Exception as e:
        print (e)
        history += "observation: I need to trigger a search using the following syntax:\naction: SCRAPE_WEBSITE action_input=URL\n"
        return "MAIN", None, history, task, result

        #else:
    #    history = "observation: The search query I used did not return a valid response"
        
    return "MAIN", None, history, task, result



#################################

NAME_TO_FUNC = {
    "MAIN": call_main,
    "UPDATE-TASK": call_set_task,
    "SEARCH_ENGINE": find_all,
    "SCRAPE_WEBSITE": find_all,
    "READ_CODE": read_code,
}


def run_action(purpose, task, history, action_name, action_input,result,repo,space,file_name):
    if "COMPLETE" in action_name:
        print("Complete - Exiting")
        #exit(0) 
        return "COMPLETE", None, history, task, result

    # compress the history when it is long
    if len(history.split("\n")) > MAX_HISTORY:
        if VERBOSE:
            print("COMPRESSING HISTORY")
        history = compress_history(purpose, task, history,file_name)
    if action_name in NAME_TO_FUNC:
        
        assert action_name in NAME_TO_FUNC

        print(f"RUN: {action_name}  ACTION_INPUT: {action_input}")
        return NAME_TO_FUNC[action_name](purpose, task, history, action_input, result,repo,space,file_name)
    else:
        history += "observation: The TOOL I tried to use returned an error, I need to select a tool from: (UPDATE-TASK, SEARCH_ENGINE, SCRAPE_WEBSITE, COMPLETE)\n"

        return "MAIN", None, history, task, result

def run(purpose,history,repo,space,f_name,file_name):
    yield [(purpose,"Searching...")]
    task=None
    result=""
    #history = ""
    if not history:
        history = ""
    else:
        history=str(history)
    action_name = "MAIN"
    action_input = f_name
    while True:
        print("")
        print("")
        print("---")
        #print("purpose:", purpose)
        print("task:", task)
        print("---")
        #print(history)
        print("---")

        action_name, action_input, history, task, result = run_action(
            purpose,
            task,
            history,
            action_name,
            action_input,
            result,
            repo,
            space,
            file_name
        )
            
        if not result:
            yield [(purpose,"More Searching...")]
        else:
            yield [(purpose,result)]
        if action_name == "COMPLETE":
            yield [(purpose,result)]

            break
    #return [(purpose,result)]


examples =[
    "What is the current weather in Florida?",
    "Find breaking news about Texas",
    "Find the best deals on flippers for scuba diving",
    "Teach me to fly a helicopter"
]
def clear_fn():
    return None,None
rand_val=random.randint(1,99999999999)

def check_rand(inp,val):
    if inp==True:
        return gr.Slider(label="Seed", minimum=1, maximum=99999999999, value=random.randint(1,99999999999))
    else:
        return gr.Slider(label="Seed", minimum=1, maximum=99999999999, value=int(val))
        
with gr.Blocks() as app:
    gr.HTML("""<center><h1>Mixtral 8x7B RPG</h1><h3>HF Co-pilot (development)</h3>""")
    with gr.Group():

        with gr.Row():
            r_name = gr.Textbox(label="Repo") 
            token = gr.Textbox(label="auth (optional)")        
            s_btn = gr.Button("Show Spaces")
        with gr.Row():
            s_name = gr.Dropdown(label="Spaces", choices=[])
            f_name = gr.Dropdown(label="Files", choices=[])            
            l_btn = gr.Button("Load Files")
                
        with gr.Row():
            with gr.Column(scale=2):
                chatbot=gr.Chatbot(show_label=False, show_share_button=True, show_copy_button=True, likeable=True, layout="panel", height="800px")
                with gr.Row():
                    with gr.Column(scale=3):
                        opt=gr.Dropdown(label="Choices",choices=examples,allow_custom_value=True, value="Start a new game", interactive=True)
                        #prompt=gr.Textbox(label = "Prompt", value="Start a new game")
                    with gr.Column(scale=2):
                        
                        rand = gr.Checkbox(label="Random", value=True)
                        seed=gr.Slider(label="Seed", minimum=1, maximum=99999999999, value=rand_val)
                    #models_dd=gr.Dropdown(choices=[m for m in return_list],interactive=True)
                with gr.Row():
                    button=gr.Button()
                    stop_button=gr.Button("Stop")
                    clear_btn = gr.Button("Clear")
                with gr.Row():
                    tokens = gr.Slider(label="Max new tokens",value=2096,minimum=0,maximum=1048*10,step=64,interactive=False, visible=False,info="The maximum numbers of new tokens")
            with gr.Column(scale=1):
                files=gr.File(file_count='directory')
                space_info_json=gr.JSON()
        file_list=gr.Textbox()
        file_frame=gr.HTML()
                
        json_out=gr.JSON()

    s_btn.click(dl.show_s,[r_name,token],[s_name])    
    #l_btn.click(dl.show_f,[r_name,s_name,token], [f_name, files, space_info_json])
    s_name.change(dl.show_f,[r_name,s_name,token], [f_name, files, file_list, space_info_json])
    #s_name.change(dl.show_f_frame2,[r_name,s_name,f_name],[file_frame])
    #s_name.change(dl.show_f_frame2,[r_name,s_name,f_name],[file_frame])
    
    #space_radio.change(show_f,[r_name,space_radio,token],[f_name, files,file_radio,space_info_json])    
    #file_radio.change(show_f_cont,[r_name,space_radio,file_radio,token],[file_contents])
    
    
    clear_btn.click(clear_fn,None,[opt,chatbot])
    #go=button.click(check_rand,[rand,seed],seed).then(run,[opt,chatbot,tokens,char_stats,seed],[chatbot,char_stats,json_out,opt])
    go=button.click(check_rand,[rand,seed],seed).then(run,[opt,chatbot,r_name,s_name,f_name,file_list],[chatbot])
    stop_button.click(None,None,None,cancels=[go])
app.queue(default_concurrency_limit=20).launch(show_api=False)