run.py

#############################################################################################################
# Title:  Gradio Interface to LLM-chatbot (for recommending AI) with RAG-funcionality and ChromaDB on HF-Hub 
# Author: Andreas Fischer
# Date:   December 30th, 2023
# Last update: May 27th, 2024
##############################################################################################################


# Chroma-DB
#-----------
import os
import chromadb
dbPath="/home/af/Schreibtisch/gradio/Chroma/db" 
if(os.path.exists(dbPath)==False): 
  dbPath="/home/user/app/db"
print(dbPath)
#client = chromadb.Client()
path=dbPath
client = chromadb.PersistentClient(path=path)
print(client.heartbeat()) 
print(client.get_version())  
print(client.list_collections()) 
from chromadb.utils import embedding_functions
default_ef = embedding_functions.DefaultEmbeddingFunction()
sentence_transformer_ef = embedding_functions.SentenceTransformerEmbeddingFunction(model_name="T-Systems-onsite/cross-en-de-roberta-sentence-transformer")
#instructor_ef = embedding_functions.InstructorEmbeddingFunction(model_name="hkunlp/instructor-large", device="cuda")
print(str(client.list_collections()))

global collection
if("name=ChromaDB1" in str(client.list_collections())):
  print("ChromaDB1 found!")
  collection = client.get_collection(name="ChromaDB1", embedding_function=sentence_transformer_ef)
else:
  print("ChromaDB1 created!")
  collection = client.create_collection(
    "ChromaDB1",
    embedding_function=sentence_transformer_ef,
    metadata={"hnsw:space": "cosine"})
  
  collection.add(
    documents=[
      "Text generating AI model mistralai/Mixtral-8x7B-Instruct-v0.1: Suitable for text generation, e.g., social media content, marketing copy, blog posts, short stories, etc.",
      "Image generating AI model stabilityai/sdxl-turbo: Suitable for image generation, e.g., illustrations, graphics, AI art, etc.",
      "Audio transcribing AI model openai/whisper-large-v3: Suitable for audio-transcription in different languages",
      "Speech synthesizing AI model coqui/XTTS-v2: Suitable for generating audio from text and for voice-cloning",
      "Code generating AI model deepseek-ai/deepseek-coder-6.7b-instruct: Suitable for programming in Python, JavaScript, PHP, Bash and many other programming languages.",
      "Translation AI model Helsinki-NLP/opus-mt: Suitable for translating text, e.g., from English to German or vice versa",
      "Search result-integrating AI model phind/phind-v9-model: Suitable for researching current topics and for obtaining precise and up-to-date answers to questions based on web search results"
    ], 
    metadatas=[{"source": "AF"}, {"source": "AF"}, {"source": "AF"}, {"source": "AF"}, {"source": "AF"}, {"source": "AF"}, {"source": "AF"}], 
    ids=["ai1", "ai2", "ai3", "ai4", "ai5", "ai6", "ai7"], 
  )

print("Database ready!")
print(collection.count()) 


# Model
#-------
onPrem=False
myModel="mistralai/Mixtral-8x7B-Instruct-v0.1" 
if(onPrem==False): 
  modelPath=myModel
  from huggingface_hub import InferenceClient
  import gradio as gr
  client = InferenceClient(
    model=modelPath,
    #token="hf_..."
  )
else:
  import os
  import requests
  import subprocess
  #modelPath="/home/af/gguf/models/c4ai-command-r-v01-Q4_0.gguf"
  #modelPath="/home/af/gguf/models/Discolm_german_7b_v1.Q4_0.gguf"
  modelPath="/home/af/gguf/models/Mixtral-8x7b-instruct-v0.1.Q4_0.gguf"
  if(os.path.exists(modelPath)==False):
    #url="https://huggingface.co/TheBloke/DiscoLM_German_7b_v1-GGUF/resolve/main/discolm_german_7b_v1.Q4_0.gguf?download=true"
    url="https://huggingface.co/TheBloke/Mixtral-8x7B-Instruct-v0.1-GGUF/resolve/main/mixtral-8x7b-instruct-v0.1.Q4_0.gguf?download=true"
    response = requests.get(url)
    with open("./Mixtral-8x7b-instruct.gguf", mode="wb") as file:
      file.write(response.content)
    print("Model downloaded")  
    modelPath="./Mixtral-8x7b-instruct.gguf"
  print(modelPath)
  n="20" 
  if("Mixtral-8x7b-instruct" in modelPath): n="0" # mixtral seems to cause problems here...
  command = ["python3", "-m", "llama_cpp.server", "--model", modelPath, "--host", "0.0.0.0", "--port", "2600", "--n_threads", "8", "--n_gpu_layers", n]
  subprocess.Popen(command)
  print("Server ready!")


# Check template
#----------------
if(False):
  from transformers import AutoTokenizer
  #mod="mistralai/Mixtral-8x22B-Instruct-v0.1"
  #mod="mistralai/Mixtral-8x7b-instruct-v0.1"
  mod="VAGOsolutions/Llama-3-SauerkrautLM-8b-Instruct"
  tok=AutoTokenizer.from_pretrained(mod) #,token="hf_...")
  cha=[{"role":"system","content":"A"},{"role":"user","content":"B"},{"role":"assistant","content":"C"}]
  res=tok.apply_chat_template(cha)
  print(tok.decode(res))
  cha=[{"role":"user","content":"U1"},{"role":"assistant","content":"A1"},{"role":"user","content":"U2"},{"role":"assistant","content":"A2"}]
  res=tok.apply_chat_template(cha)
  print(tok.decode(res))


# Gradio-GUI
#------------

import gradio as gr
import json
import re

def extend_prompt(message="", history=None, system=None, RAGAddon=None, system2=None, zeichenlimit=None,historylimit=4, removeHTML=True): 
  startOfString=""
  if zeichenlimit is None: zeichenlimit=1000000000 # :-)
  template0=" [INST]{system}\n  [/INST] </s>" 
  template1=" [INST] {message} [/INST]"
  template2=" {response}</s>"
  if("command-r" in modelPath): #https://huggingface.co/CohereForAI/c4ai-command-r-v01
    ## <BOS_TOKEN><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Hello, how are you?<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>
    template0="<BOS_TOKEN><|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|> {system}<|END_OF_TURN_TOKEN|>" 
    template1="<|START_OF_TURN_TOKEN|><|USER_TOKEN|>{message}<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>"
    template2="{response}<|END_OF_TURN_TOKEN|>"
  if("Gemma-" in modelPath): # https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1
    template0="<start_of_turn>user{system}</end_of_turn>" 
    template1="<start_of_turn>user{message}</end_of_turn><start_of_turn>model"
    template2="{response}</end_of_turn>"      
  if("Mixtral-8x22B-Instruct" in modelPath): # AutoTokenizer: <s>[INST] U1[/INST] A1</s>[INST] U2[/INST] A2</s>
    startOfString="<s>"
    template0="[INST]{system}\n  [/INST] </s>"  
    template1="[INST] {message}[/INST]"
    template2=" {response}</s>"
  if("Mixtral-8x7b-instruct" in modelPath): # https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1
    startOfString="<s>"                     # AutoTokenzizer: <s> [INST] U1 [/INST]A1</s> [INST] U2 [/INST]A2</s>
    template0=" [INST]{system}\n  [/INST] </s>"  
    template1=" [INST] {message} [/INST]"
    template2=" {response}</s>"
  if("Mistral-7B-Instruct" in modelPath): #https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2
    startOfString="<s>"
    template0="[INST]{system}\n [/INST]</s>"
    template1="[INST] {message} [/INST]"
    template2=" {response}</s>"
  if("Openchat-3.5" in modelPath): #https://huggingface.co/TheBloke/openchat-3.5-0106-GGUF
    template0="GPT4 Correct User: {system}<|end_of_turn|>GPT4 Correct Assistant: Okay.<|end_of_turn|>"
    template1="GPT4 Correct User: {message}<|end_of_turn|>GPT4 Correct Assistant: "
    template2="{response}<|end_of_turn|>"
  if(("Discolm_german_7b" in modelPath) or ("SauerkrautLM-7b-HerO" in modelPath)):  #https://huggingface.co/VAGOsolutions/SauerkrautLM-7b-HerO
    template0="<|im_start|>system\n{system}<|im_end|>\n"
    template1="<|im_start|>user\n{message}<|im_end|>\n<|im_start|>assistant\n"
    template2="{response}<|im_end|>\n"    
  if("Llama-3-SauerkrautLM-8b-Instruct" in modelPath):  #https://huggingface.co/VAGOsolutions/SauerkrautLM-7b-HerO
    template0="<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{system}<|eot_id|>"
    template1="<|start_header_id|>user<|end_header_id|>\n\n{message}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
    template2="{response}<|eot_id|>\n"        
  if("WizardLM-13B-V1.2" in modelPath): #https://huggingface.co/WizardLM/WizardLM-13B-V1.2
    template0="{system} " #<s>
    template1="USER: {message} ASSISTANT: "
    template2="{response}</s>"
  if("Phi-2" in modelPath): #https://huggingface.co/TheBloke/phi-2-GGUF
    template0="Instruct: {system}\nOutput: Okay.\n"
    template1="Instruct: {message}\nOutput:"
    template2="{response}\n"  
  prompt = ""
  if RAGAddon is not None:
    system += RAGAddon
  if system is not None:
    prompt += template0.format(system=system) #"<s>"
  if history is not None:
    for user_message, bot_response in history[-historylimit:]:
      if user_message is None: user_message = "" 
      if bot_response is None: bot_response = ""
      bot_response = re.sub("\n\n<details>((.|\n)*?)</details>","", bot_response) # remove RAG-compontents
      if removeHTML==True: bot_response = re.sub("<(.*?)>","\n", bot_response) # remove HTML-components in general (may cause bugs with markdown-rendering)
      if user_message is not None: prompt += template1.format(message=user_message[:zeichenlimit])  
      if bot_response is not None: prompt += template2.format(response=bot_response[:zeichenlimit]) 
  if message is not None: prompt += template1.format(message=message[:zeichenlimit])                
  if system2 is not None:
    prompt += system2
  return startOfString+prompt



def response(
    message, history, temperature=0.9, max_new_tokens=500, top_p=0.95, repetition_penalty=1.0,
):
    temperature = float(temperature)
    if temperature < 1e-2: temperature = 1e-2
    top_p = float(top_p)
    generate_kwargs = dict(
        temperature=temperature,
        max_new_tokens=max_new_tokens,
        top_p=top_p,
        repetition_penalty=repetition_penalty,
        do_sample=True,
        seed=42,
    )
    addon=""
    results=collection.query(
      query_texts=[message],
      n_results=2,
      #where={"source": "google-docs"}
      #where_document={"$contains":"search_string"}
    )
    dists=["<br><small>(relevance: "+str(round((1-d)*100)/100)+";" for d in results['distances'][0]]
    sources=["source: "+s["source"]+")</small>" for s in results['metadatas'][0]]
    results=results['documents'][0]
    combination = zip(results,dists,sources)
    combination = [' '.join(triplets) for triplets in combination]
    print(combination)
    if(len(results)>1):
      addon=" Bitte berücksichtige bei deiner Antwort ggf. folgende Auszüge aus unserer Datenbank, sofern sie für die Antwort erforderlich sind. Beantworte die Frage knapp und präzise. Ignoriere unpassende Datenbank-Auszüge OHNE sie zu kommentieren, zu erwähnen oder aufzulisten:\n"+"\n".join(results)
    system="Du bist ein deutschsprachiges KI-basiertes Assistenzsystem, das zu jedem Anliegen möglichst geeignete KI-Tools empfiehlt." #+addon #+"\n\nUser-Anliegen:"   
    #body={"prompt":system+"### Instruktion:\n"+message+"\n\n### Antwort:","max_tokens":500, "echo":"False","stream":"True"} #e.g. SauerkrautLM
    #formatted_prompt = extend_prompt(system+"\n"+prompt,  None) #history)
    prompt=extend_prompt(
      message,                  # current message of the user
      history,                  # complete history 
      system,                   # system prompt
      addon,                    # RAG-component added to the system prompt
      None,                     # fictive first words of the AI (neither displayed nor stored)
      historylimit=4,           # number of past messages to consider for response to current message
      removeHTML=True           # remove HTML-components from History (to prevent bugs with Markdown)
    )
    stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)
    output = ""
    for response in stream:
        output += response.token.text
        yield output
    output=output+"\n\n<br><details open><summary><strong>Sources</strong></summary><br><ul>"+ "".join(["<li>" + s + "</li>" for s in combination])+"</ul></details>"
    yield output

gr.ChatInterface(response, chatbot=gr.Chatbot(value=[[None,"Herzlich willkommen! Ich bin ein KI-basiertes Assistenzsystem, das für jede Anfrage die am besten geeigneten KI-Tools empfiehlt.<br>Aktuell bin ich wenig mehr als eine Tech-Demo und kenne nur 7 KI-Modelle - also sei bitte nicht zu streng mit mir.<br>Was ist dein Anliegen?"]],render_markdown=True),title="German AI-RAG-Interface to the Hugging Face Hub").queue().launch(share=True) #False, server_name="0.0.0.0", server_port=7864)
print("Interface up and running!")