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
#from query import tasks
from prompts import (
    FINDER,
    COMPRESS_HISTORY_PROMPT,
    COMPRESS_DATA_PROMPT,
    LOG_PROMPT,
    LOG_RESPONSE,
    PREFIX,
    TASK_PROMPT,
)
api=HfApi()



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

def parse_action(string: str):
    assert string.startswith("action:")
    idx = string.find("action_input=")
    if idx == -1:
        return string[8:], None
    return string[8 : idx - 1], string[idx + 13 :].strip("'").strip('"')



VERBOSE = False
MAX_HISTORY = 100
MAX_DATA = 100

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 call_search(purpose, task, history, action_input):
    return_list=[]
    print (action_input)
    #if action_input in query.tasks:
    print ("trying")        
    try:
        if action_input != "" and action_input != None:
            action_input.strip('""')
            #model_list = api.list_models(filter=f"{action_input}",sort="last_modified",limit=1000,direction=-1)
            #model_list = api.list_models(filter=f"{action_input}",limit=1000)
            model_list = api.list_models(filter=f"{action_input}")
            this_obj = list(model_list)
            print(f'THIS_OBJ :: {this_obj[0]}')
            for i,eb in enumerate(this_obj):
                #return_list.append(this_obj[i].id)
                return_list.append({"id":this_obj[i].id,
                                    "author":this_obj[i].author,
                                    "created_at":this_obj[i].created_at,
                                    "last_modified":this_obj[i].last_modified,
                                    "private":this_obj[i].private,
                                    "gated":this_obj[i].gated,
                                    "disabled":this_obj[i].disabled,
                                    "downloads":this_obj[i].downloads,
                                    "likes":this_obj[i].likes,
                                    "library_name":this_obj[i].library_name,
                                    "tags":this_obj[i].tags,
                                    "pipeline_tag":this_obj[i].pipeline_tag,
                                   })
            #print (return_list)
            c=0
            rl = len(return_list)
            print(rl)
            for i in str(return_list):
                if i == " " or i==",":
                    c +=1
            
            print (c)
            if rl > MAX_DATA:
                print("compressing...")
                return_list = compress_data(rl,purpose,task,return_list)
            history = "observation: the search results are:\n {}\n".format(return_list)
            return "COMPLETE", None, history, task
        else: 
            history = "observation: I need to trigger a search using the following syntax:\naction: SEARCH action_input=SEARCH_QUERY\n"
            return "UPDATE-TASK", None, history, task
    except Exception as e:
        print (e)
        history = "observation: I need to trigger a search using the following syntax:\naction: SEARCH action_input=SEARCH_QUERY\n"
        return "UPDATE-TASK", None, history, task

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


def run_gpt(
    prompt_template,
    stop_tokens,
    max_tokens,
    seed,
    purpose,
    **prompt_kwargs,
):
    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(
        purpose=purpose,
    ) + 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}', 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):
    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_SMALL,
            stop_tokens=["observation:", "task:", "action:", "thought:"],
            max_tokens=2048,
            seed=seed,
            purpose=purpose,
            task=task,
            knowledge=new_history,
            history=hist,
        )
        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=1024,
        seed=seed,
        purpose=purpose,
        task=task,
        knowledge=new_history,
        history="All data has been recieved.",
    )'''
    print ("final" + resp)
    history = "observation: {}\n".format(resp)
    return history




def compress_history(purpose, task, history):
    resp = run_gpt(
        COMPRESS_HISTORY_PROMPT,
        stop_tokens=["observation:", "task:", "action:", "thought:"],
        max_tokens=512,
        seed=random.randint(1,1000000000),
        purpose=purpose,
        task=task,
        history=history,
    )
    history = "observation: {}\n".format(resp)
    return history


def call_main(purpose, task, history, action_input):
    resp = run_gpt(
        FINDER,
        stop_tokens=["observation:", "task:"],
        max_tokens=2048,
        seed=random.randint(1,1000000000),
        purpose=purpose,
        task=task,
        history=history,
    )
    lines = resp.strip().strip("\n").split("\n")
    for line in lines:
        if line == "":
            continue
        if line.startswith("thought: "):
            history += "{}\n".format(line)
        elif line.startswith("action: COMPLETE"):
            return "COMPLETE", None, history, task
        elif line.startswith("action: "):
            action_name, action_input = parse_action(line)
            history += "{}\n".format(line)
            return action_name, action_input, history, task
        else:
            
            #history += "observation: {}\n".format(line)
            #assert False, "unknown action: {}".format(line)
            return "UPDATE-TASK", None, history, task
            
    return "MAIN", None, history, task


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



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



def find_all(purpose,task,history, url):
    return_list=[]
    print (url)
    #if action_input in query.tasks:
    print ("trying")        
    try:
        if url != "" and url != None:    
            rawp = []
            source = urllib.request.urlopen(url).read()
            soup = bs4.BeautifulSoup(source,'lxml')
            # title of the page
            print(soup.title)
            # get attributes:
            print(soup.title.name)
            # get values:
            print(soup.title.string)
            # beginning navigation:
            print(soup.title.parent.name)
            #rawp.append([tag.name for tag in soup.find_all()] )
            print([tag.name for tag in soup.find_all()])
            rawp=soup.text
            c=0
            rl = len(rawp)
            print(rl)
            for i in str(rawp):
                if i == " " or i==",":
                    c +=1
            
            print (c)
            if c > MAX_DATA:
                print("compressing...")
                rawp = compress_data(c,purpose,task,rawp)    
            print (rawp)
            history += "observation: the search results are:\n {}\n".format(rawp)
            task = "complete?"
            return "UPDATE-TASK", None, history, task
        else: 
            history += "observation: I need to trigger a search using the following syntax:\naction: WEBSITE_SCRAPE action_input=SEARCH_QUERY\n"
            return "UPDATE-TASK", None, history, task
    except Exception as e:
        print (e)
        history += "observation: I need to trigger a search using the following syntax:\naction: WEBSITE_SCRAPE action_input=SEARCH_QUERY\n"
        return "UPDATE-TASK", None, history, task

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


def find_it(url,q=None,num=None):
    out = []
    out_l = []
    z=""
    source = urllib.request.urlopen(url).read()
    soup = bs4.BeautifulSoup(source,'lxml')
    
    for p in soup.find_all(f'{q}'):
        if num != "":
            z=p.get(f'{num}')
        
        try:
            test = soup.select(f'{p.name}:first-child')
            
            #print(p.findChildren())
        except Exception as e:
            print (e)
        #out.append(p)
        out.append([{q:p.string,"additional":z,"parent":p.parent.name,"previous":[b for b in p.previous],"first-child":[b.name for b in p.children],"content":p}])
        if p.string !=None:
            out_l.append(p.string)
        else:
            out_l.append(z)
        #out.append(p.parent.name)
        print(dir(p))
        print(p.parent.name)
    for url in soup.find_all('a'):
        print(url.get('href'))
        
    #print(soup.get_text())
    return out,out_l
    
def find_it2(url):
    response = requests.get(url,a1=None,q2=None,q3=None)
    try:
        response.raise_for_status()
        soup = BeautifulSoup(response.content, 'lxml')
        out = 'URL Links:\n'.join([p.text for p in soup.find_all('a')])
        return out
    except Exception as e:
        print (e)
        return e
#################################

NAME_TO_FUNC = {
    "MAIN": call_main,
    "UPDATE-TASK": call_set_task,
    "SEARCH_ENGINE": find_all,
    "WEBSITE_SCRAPE": find_all,
    

}


def run_action(purpose, task, history, action_name, action_input):
    if action_name == "COMPLETE":
        print("Complete - Exiting")
        #exit(0) 
        return "COMPLETE", None, history, task

    # 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)
    if action_name in NAME_TO_FUNC:
        
        assert action_name in NAME_TO_FUNC

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

        return "MAIN", None, history, task

def run(purpose,history):
    task=None
    history = ""
    #if not history:
    #    history = []
    action_name = "SEARCH_ENGINE" if task is None else "MAIN"
    action_input = None
    while True:
        print("")
        print("")
        print("---")
        #print("purpose:", purpose)
        print("task:", task)
        print("---")
        #print(history)
        print("---")

        action_name, action_input, history, task = run_action(
            purpose,
            task,
            history,
            action_name,
            action_input,
        )
        yield history
        if action_name == "COMPLETE":
            return history




examples =[
    "find the most popular model that I can use to generate an image by providing a text prompt",
    "return the top 10 models that I can use to identify objects in images",
    "which models have the most likes from each category?"
]


gr.ChatInterface(
    fn=run,
    chatbot=gr.Chatbot(show_label=False, show_share_button=False, show_copy_button=True, likeable=True, layout="panel"),
    title="Mixtral 46.7B Powered <br> Search",
    examples=examples,
    concurrency_limit=20,
).launch(show_api=False)

'''
with gr.Blocks() as app:
    with gr.Row():
        with gr.Column(scale=1):
            inp = gr.Textbox()
        with gr.Column(scale=2):
            q = gr.Textbox(value="p")
        with gr.Column(scale=2):
            num = gr.Textbox()
    with gr.Row():
        all_btn = gr.Button("Load")
        find_btn = gr.Button("Find")
    with gr.Row():
        rawp = gr.JSON()
        outp = gr.JSON()
        outl = gr.Textbox()
    
    all_btn.click(find_all,[inp,q,num],[rawp])
    find_btn.click(find_it,[inp,q,num],[outp,outl])
    
app.launch()

'''