Spaces:
Sleeping
Sleeping
| ######################################################################################### | |
| # Title: Gradio AI-Interface with Memory-RAG | |
| # Author: Andreas Fischer | |
| # Date: October 15th, 2023 | |
| # Last update: May 27th, 2024 | |
| ########################################################################################## | |
| #https://github.com/abetlen/llama-cpp-python/issues/306 | |
| #sudo apt install libclblast-dev | |
| #CMAKE_ARGS="-DLLAMA_CLBLAST=on" FORCE_CMAKE=1 pip install llama-cpp-python --force-reinstall --upgrade --no-cache-dir -v | |
| # Prepare resources | |
| #------------------- | |
| import torch | |
| import gc | |
| torch.cuda.empty_cache() | |
| gc.collect() | |
| import os | |
| from datetime import datetime | |
| global filename | |
| filename=f"./{datetime.now().strftime('%Y%m%d')}_history.json" # where to store the history as json-file | |
| if(os.path.exists(filename)==True): os.remove(filename) | |
| # Chroma-DB | |
| #----------- | |
| import os | |
| import chromadb | |
| dbPath = "/home/af/Schreibtisch/Code/gradio/Chroma/db" | |
| onPrem = True if(os.path.exists(dbPath)) else False | |
| if(onPrem==False): dbPath="/home/user/app/db" | |
| #onPrem=True # uncomment to override automatic detection | |
| 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") | |
| embeddingModel = embedding_functions.SentenceTransformerEmbeddingFunction(model_name="T-Systems-onsite/cross-en-de-roberta-sentence-transformer", device="cuda" if(onPrem) else "cpu") | |
| print(str(client.list_collections())) | |
| global collection | |
| dbName="historicalChromaDB1" | |
| if("name="+dbName in str(client.list_collections())): client.delete_collection(name=dbName) # deletes collection | |
| if("name="+dbName in str(client.list_collections())): | |
| print(dbName+" found!") | |
| collection = client.get_collection(name=dbName, embedding_function=embeddingModel) #sentence_transformer_ef) | |
| else: | |
| #client.delete_collection(name=dbName) | |
| print(dbName+" created!") | |
| collection = client.create_collection( | |
| dbName, | |
| embedding_function=embeddingModel, | |
| metadata={"hnsw:space": "cosine"}) | |
| print("Database ready!") | |
| print(collection.count()) | |
| x=collection.get(include=[])["ids"] | |
| if(len(x)==0): | |
| message="Ich bin der User." | |
| response="Hallo User, wie kann ich dienen?" | |
| x=collection.get(include=[])["ids"] | |
| collection.add( | |
| documents=[message,response], | |
| metadatas=[ | |
| {"source": "ICH", "dialog": f"ICH: {message}\nDU: {response}"}, | |
| {"source": "DU", "dialog": f"ICH: {message}\nDU: {response}"} | |
| ], | |
| ids=[str(len(x)+1),str(len(x)+2)] | |
| ) | |
| RAGResults=collection.query( | |
| query_texts=[message], | |
| n_results=1, | |
| #where={"source": "USER"} | |
| ) | |
| RAGResults["metadatas"][0][0]["dialog"] | |
| # Model | |
| #------- | |
| #onPrem=False | |
| if(onPrem==False): | |
| modelPath="mistralai/Mixtral-8x7B-Instruct-v0.1" | |
| from huggingface_hub import InferenceClient | |
| import gradio as gr | |
| client = InferenceClient( | |
| modelPath | |
| #"mistralai/Mixtral-8x7B-Instruct-v0.1" | |
| #"mistralai/Mistral-7B-Instruct-v0.1" | |
| ) | |
| else: | |
| import os | |
| import requests | |
| import subprocess | |
| #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!") | |
| #import llama_cpp | |
| #llama_cpp.llama_backend_init(numa=False) | |
| #params=llama_cpp.llama_context_default_params() | |
| #params.n_ctx | |
| # Gradio-GUI | |
| #------------ | |
| 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(| open)>.*?</details>","", bot_response, flags=re.DOTALL) # 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 | |
| import gradio as gr | |
| import requests | |
| import json | |
| from datetime import datetime | |
| import os | |
| import re | |
| def response(message, history,customSysPrompt,settings): | |
| #settings="Memory Off" | |
| removeHTML=True | |
| system=customSysPrompt | |
| message=message.replace("[INST]","") | |
| message=message.replace("[/INST]","") | |
| message=message.replace("</s>","") | |
| message=re.sub("<[|](im_start|im_end|end_of_turn)[|]>", '', message) | |
| if (settings=="Memory On"): | |
| if((len(history)==0)&(os.path.isfile(filename))): history=json.load(open(filename,'r',encoding="utf-8")) # retrieve history (if available) | |
| x=collection.get(include=[])["ids"] | |
| rag=None # RAG is turned off until history gets too long | |
| historylimit=0 #4 | |
| if(len(x)>(historylimit*2)): # turn on RAG when the database contains entries that are not shown within historylimit | |
| RAGResults=collection.query( | |
| query_texts=[message], | |
| n_results=1, | |
| #where={"source": "USER"} | |
| ) | |
| bestMatch=str(RAGResults["metadatas"][0][0]["dialog"]) | |
| #print("Message: "+message+"\n\nBest Match: "+bestMatch) | |
| rag="\n\n" | |
| rag += "Mit Blick auf den aktuellen Stand des Dialogs erinnerst du dich insb. an folgende Episode:\n" | |
| rag += bestMatch | |
| rag += "\n\nIm Folgenden siehst du den aktuellen Stand des Dialogs." | |
| system2=None # system2 can be used as fictive first words of the AI, which are not displayed or stored | |
| #print("RAG: "+rag) | |
| #print("System: "+system+"\n\nMessage: "+message) | |
| prompt=extend_prompt( | |
| message, # current message of the user | |
| history, # complete history | |
| system, # system prompt | |
| rag, # RAG-component added to the system prompt | |
| system2, # fictive first words of the AI (neither displayed nor stored) | |
| historylimit=historylimit,# number of past messages to consider for response to current message | |
| removeHTML=removeHTML # remove HTML-components from History (to prevent bugs with Markdown) | |
| ) | |
| #print("\n\nMESSAGE:"+str(message)) | |
| #print("\n\nHISTORY:"+str(history)) | |
| #print("\n\nSYSTEM:"+str(system)) | |
| #print("\n\nRAG:"+str(rag)) | |
| #print("\n\nSYSTEM2:"+str(system2)) | |
| #print("\n\n*** Prompt:\n"+prompt+"\n***\n\n") | |
| ## Request response from model | |
| #------------------------------ | |
| print("AI running on prem!" if(onPrem) else "AI running HFHub!") | |
| if(onPrem==False): | |
| temperature=float(0.9) | |
| max_new_tokens=500 | |
| top_p=0.95 | |
| repetition_penalty=1.0 | |
| 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, | |
| ) | |
| stream = client.text_generation(prompt, **generate_kwargs, stream=True, details=True, return_full_text=False) | |
| response = "" | |
| #print("User: "+message+"\nAI: ") | |
| for text in stream: | |
| part=text.token.text | |
| #print(part, end="", flush=True) | |
| response += part | |
| if removeHTML==True: response = re.sub("<(.*?)>","\n", response) # remove HTML-components in general (may cause bugs with markdown-rendering) | |
| yield response | |
| # Store current state in DB if memory is turned on | |
| if (settings=="Memory On"): | |
| x=collection.get(include=[])["ids"] # add current dialog to db | |
| collection.add( | |
| documents=[message,response], | |
| metadatas=[ | |
| { "source": "ICH", "dialog": f"ICH: {message.strip()}\n DU: {response.strip()}", "type":"episode"}, | |
| { "source": "DU", "dialog": f"ICH: {message.strip()}\n DU: {response.strip()}", "type":"episode"} | |
| ], | |
| ids=[str(len(x)+1),str(len(x)+2)] | |
| ) | |
| json.dump(history,open(filename,'w',encoding="utf-8"),ensure_ascii=False) | |
| if(onPrem==True): | |
| # url="https://afischer1985-wizardlm-13b-v1-2-q4-0-gguf.hf.space/v1/completions" | |
| url="http://0.0.0.0:2600/v1/completions" | |
| body={"prompt":prompt,"max_tokens":None, "echo":"False","stream":"True"} # e.g. Mixtral-Instruct | |
| if("Discolm_german_7b" in modelPath): body.update({"stop": ["<|im_end|>"]}) # fix stop-token of DiscoLM | |
| if("Gemma-" in modelPath): body.update({"stop": ["<|im_end|>","</end_of_turn>"]}) # fix stop-token of Gemma | |
| response="" #+"("+myType+")\n" | |
| buffer="" | |
| #print("URL: "+url) | |
| #print("User: "+message+"\nAI: ") | |
| for text in requests.post(url, json=body, stream=True): #-H 'accept: application/json' -H 'Content-Type: application/json' | |
| if buffer is None: buffer="" | |
| buffer=str("".join(buffer)) | |
| # print("*** Raw String: "+str(text)+"\n***\n") | |
| text=text.decode('utf-8') | |
| if((text.startswith(": ping -")==False) & (len(text.strip("\n\r"))>0)): buffer=buffer+str(text) | |
| # print("\n*** Buffer: "+str(buffer)+"\n***\n") | |
| buffer=buffer.split('"finish_reason": null}]}') | |
| if(len(buffer)==1): | |
| buffer="".join(buffer) | |
| pass | |
| if(len(buffer)==2): | |
| part=buffer[0]+'"finish_reason": null}]}' | |
| if(part.lstrip('\n\r').startswith("data: ")): part=part.lstrip('\n\r').replace("data: ", "") | |
| try: | |
| part = str(json.loads(part)["choices"][0]["text"]) | |
| #print(part, end="", flush=True) | |
| response=response+part | |
| buffer="" # reset buffer | |
| except Exception as e: | |
| print("Exception:"+str(e)) | |
| pass | |
| if removeHTML==True: response = re.sub("<(.*?)>","\n", response) # remove HTML-components in general (may cause bugs with markdown-rendering) | |
| yield response | |
| history.append((message, response)) # add current dialog to history | |
| # Store current state in DB if memory is turned on | |
| if (settings=="Memory On"): | |
| x=collection.get(include=[])["ids"] # add current dialog to db | |
| collection.add( | |
| documents=[message,response], | |
| metadatas=[ | |
| { "source": "ICH", "dialog": f"ICH: {message.strip()}\n DU: {response.strip()}", "type":"episode"}, | |
| { "source": "DU", "dialog": f"ICH: {message.strip()}\n DU: {response.strip()}", "type":"episode"} | |
| ], | |
| ids=[str(len(x)+1),str(len(x)+2)] | |
| ) | |
| json.dump(history,open(filename,'w',encoding="utf-8"),ensure_ascii=False) | |
| gr.ChatInterface( | |
| response, | |
| chatbot=gr.Chatbot(value=[[None,"Herzlich willkommen! Ich bin ein KI-basiertes Assistenzsystem, mit dem du zu unterschiedlichen Themen sprechen kannst.<br>Wenn du mich lokal bzw. \"on premises\" betreibst, kann ich unseren Dialog auf Wunsch auch speichern und bei meinen Antworten auch lange zurück liegende Episoden berücksichtigen.<br>Was ist dein Anliegen?"]], render_markdown=True), | |
| title="AI-Interface (on prem)" if onPrem else "AI-Interface (HFHub)", | |
| additional_inputs=[ | |
| gr.Textbox(value=None,label="System Prompt"), | |
| gr.Dropdown(["Memory On","Memory Off"],value="Memory Off",label="Memory") | |
| ] | |
| ).queue().launch(share=True) #False, server_name="0.0.0.0", server_port=7864) | |
| print("Interface up and running!") | |