Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -0,0 +1,169 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from datasets import load_dataset
|
2 |
+
from datasets import Dataset
|
3 |
+
from sentence_transformers import SentenceTransformer
|
4 |
+
import faiss
|
5 |
+
import time
|
6 |
+
import json
|
7 |
+
#import torch
|
8 |
+
import pandas as pd
|
9 |
+
from llama_cpp import Llama
|
10 |
+
#from langchain_community.llms import LlamaCpp
|
11 |
+
from threading import Thread
|
12 |
+
from huggingface_hub import Repository, upload_file
|
13 |
+
import os
|
14 |
+
|
15 |
+
|
16 |
+
HF_TOKEN = os.getenv('HF_Token')
|
17 |
+
#Log_Path="./Logfolder"
|
18 |
+
logfile = 'DiabetesChatLog.txt'
|
19 |
+
historylog = [{
|
20 |
+
"Prompt": '',
|
21 |
+
"Output": ''
|
22 |
+
}]
|
23 |
+
|
24 |
+
data = load_dataset("Namitg02/Test", split='train', streaming=False)
|
25 |
+
#Returns a list of dictionaries, each representing a row in the dataset.
|
26 |
+
length = len(data)
|
27 |
+
|
28 |
+
embedding_model = SentenceTransformer("all-MiniLM-L6-v2")
|
29 |
+
embedding_dim = embedding_model.get_sentence_embedding_dimension()
|
30 |
+
# Returns dimensions of embedidng
|
31 |
+
|
32 |
+
|
33 |
+
index = faiss.IndexFlatL2(embedding_dim)
|
34 |
+
data.add_faiss_index("embeddings", custom_index=index)
|
35 |
+
# adds an index column for the embeddings
|
36 |
+
|
37 |
+
#question = "How can I reverse Diabetes?"
|
38 |
+
|
39 |
+
SYS_PROMPT = """You are an assistant for answering questions.
|
40 |
+
You are given the extracted parts of documents and a question. Provide a conversational answer.
|
41 |
+
If you don't know the answer, just say "I do not know." Don't make up an answer. Don't repeat the SYS_PROMPT."""
|
42 |
+
# Provides context of how to answer the question
|
43 |
+
|
44 |
+
#llm_model = "TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF", tinyllama-1.1b-chat-v1.0.Q5_K_M.gguf
|
45 |
+
# TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF , TinyLlama/TinyLlama-1.1B-Chat-v0.6, andrijdavid/TinyLlama-1.1B-Chat-v1.0-GGUF"
|
46 |
+
|
47 |
+
model = Llama(
|
48 |
+
model_path="./llama-2-7b-chat.Q4_K_M.gguf",
|
49 |
+
# chat_format="llama-2",
|
50 |
+
n_gpu_layers = 0,
|
51 |
+
temperature=0.75,
|
52 |
+
n_ctx = 4096,
|
53 |
+
max_tokens=500,
|
54 |
+
top_p=0.95 #,
|
55 |
+
# eos_tokens=terminators
|
56 |
+
# callback_manager=callback_manager,
|
57 |
+
# verbose=True, # Verbose is required to pass to the callback manager
|
58 |
+
)
|
59 |
+
#initiate model and tokenizer
|
60 |
+
|
61 |
+
def search(query: str, k: int = 2 ):
|
62 |
+
"""a function that embeds a new query and returns the most probable results"""
|
63 |
+
embedded_query = embedding_model.encode(query) # create embedding of a new query
|
64 |
+
scores, retrieved_examples = data.get_nearest_examples( # retrieve results
|
65 |
+
"embeddings", embedded_query, # compare our new embedded query with the dataset embeddings
|
66 |
+
k=k # get only top k results
|
67 |
+
)
|
68 |
+
return scores, retrieved_examples
|
69 |
+
# returns scores (List[float]): the retrieval scores from either FAISS (IndexFlatL2 by default) and examples (dict) format
|
70 |
+
# called by talk function that passes prompt
|
71 |
+
|
72 |
+
def format_prompt(prompt,retrieved_documents,k,history,memory_limit=3):
|
73 |
+
"""using the retrieved documents we will prompt the model to generate our responses"""
|
74 |
+
PROMPT = f"Question:{prompt}\nContext:"
|
75 |
+
for idx in range(k) :
|
76 |
+
PROMPT+= f"{retrieved_documents['0'][idx]}\n"
|
77 |
+
print("historyinfo")
|
78 |
+
print(history)
|
79 |
+
if len(history) == 0:
|
80 |
+
return PROMPT
|
81 |
+
|
82 |
+
if len(history) > memory_limit:
|
83 |
+
history = history[-memory_limit:]
|
84 |
+
|
85 |
+
print("checkwohist")
|
86 |
+
# PROMPT = PROMPT + f"{history[0][0]} [/INST] {history[0][1]} </s>"
|
87 |
+
# print("checkwthhist")
|
88 |
+
# print(PROMPT)
|
89 |
+
# Handle conversation history
|
90 |
+
for user_message, bot_message in history[0:]:
|
91 |
+
PROMPT += f"<s>[INST] {user_message} [/INST] {bot_message} </s>"
|
92 |
+
print("checkwthhist2")
|
93 |
+
print(PROMPT)
|
94 |
+
return PROMPT
|
95 |
+
|
96 |
+
|
97 |
+
# Called by talk function to add retrieved documents to the prompt. Keeps adding text of retrieved documents to string that are retreived
|
98 |
+
|
99 |
+
def talk(prompt, history):
|
100 |
+
k = 2 # number of retrieved documents
|
101 |
+
scores , retrieved_documents = search(prompt, k) # get retrival scores and examples in dictionary format based on the prompt passed
|
102 |
+
print(retrieved_documents.keys())
|
103 |
+
# print("check4")
|
104 |
+
formatted_prompt = format_prompt(prompt,retrieved_documents,k,history,memory_limit=3) # create a new prompt using the retrieved documents
|
105 |
+
print("check5")
|
106 |
+
pd.options.display.max_colwidth = 4000
|
107 |
+
# print(retrieved_documents['0'])
|
108 |
+
# print(formatted_prompt)
|
109 |
+
# formatted_prompt_with_history = add_history(formatted_prompt, history)
|
110 |
+
|
111 |
+
# formatted_prompt_with_history = formatted_prompt_with_history[:600] # to avoid memory issue
|
112 |
+
# print(formatted_prompt_with_history)
|
113 |
+
messages = [{"role":"system","content":SYS_PROMPT},{"role":"user","content":formatted_prompt}]
|
114 |
+
print(messages)
|
115 |
+
# binding the system context and new prompt for LLM
|
116 |
+
# the chat template structure should be based on text generation model format
|
117 |
+
|
118 |
+
# indicates the end of a sequence
|
119 |
+
stream = model.create_chat_completion(messages = messages, max_tokens=1000, stop=["</s>"], stream=False)
|
120 |
+
# print(f"{stream}")
|
121 |
+
print("check 7")
|
122 |
+
print(stream['choices'][0]['message']['content'])
|
123 |
+
return(stream['choices'][0]['message']['content'])
|
124 |
+
# text = ""
|
125 |
+
# for output in stream:
|
126 |
+
# text += output['choices'][0]['message']['content']
|
127 |
+
# print(f"{output}")
|
128 |
+
# print("check3H")
|
129 |
+
# print(text)
|
130 |
+
# yield text
|
131 |
+
|
132 |
+
|
133 |
+
|
134 |
+
# calling the model to generate response based on message/ input
|
135 |
+
# do_sample if set to True uses strategies to select the next token from the probability distribution over the entire vocabulary
|
136 |
+
# temperature controls randomness. more renadomness with higher temperature
|
137 |
+
# only the tokens comprising the top_p probability mass are considered for responses
|
138 |
+
# This output is a data structure containing all the information returned by generate(), but that can also be used as tuple or dictionary.
|
139 |
+
|
140 |
+
|
141 |
+
|
142 |
+
TITLE = "AI Copilot for Diabetes Patients"
|
143 |
+
|
144 |
+
DESCRIPTION = "I provide answers to concerns related to Diabetes"
|
145 |
+
|
146 |
+
import gradio as gr
|
147 |
+
# Design chatbot
|
148 |
+
demo = gr.ChatInterface(
|
149 |
+
fn=talk,
|
150 |
+
chatbot=gr.Chatbot(
|
151 |
+
show_label=True,
|
152 |
+
show_share_button=True,
|
153 |
+
show_copy_button=True,
|
154 |
+
likeable=True,
|
155 |
+
layout="bubble",
|
156 |
+
bubble_full_width=False,
|
157 |
+
),
|
158 |
+
theme="Soft",
|
159 |
+
examples=[["what is Diabetes?"]],
|
160 |
+
title=TITLE,
|
161 |
+
description=DESCRIPTION,
|
162 |
+
)
|
163 |
+
# launch chatbot and calls the talk function which in turn calls other functions
|
164 |
+
print("check14")
|
165 |
+
#print(historylog)
|
166 |
+
#memory_panda = pd.DataFrame(historylog)
|
167 |
+
#Logfile = Dataset.from_pandas(memory_panda)
|
168 |
+
#Logfile.push_to_hub("Namitg02/Logfile",token = HF_TOKEN)
|
169 |
+
demo.launch()
|