app.py

from datasets import load_dataset
from langchain.docstore.document import Document as LangchainDocument
from langchain.text_splitter import RecursiveCharacterTextSplitter
from sentence_transformers import SentenceTransformer
from langchain_community.embeddings import HuggingFaceEmbeddings
#from langchain_community.vectorstores import faiss
import faiss
from langchain.prompts import PromptTemplate
#from langchain.chains import ConversationalRetrievalChain
#from transformers import pipeline
#from langchain_community.llms.huggingface_pipeline import HuggingFacePipeline
#from langchain_core.messages import SystemMessage
import time
from transformers import AutoTokenizer
from transformers import AutoModelForCausalLM
from transformers import TextIteratorStreamer
from threading import Thread



#dataset = load_dataset("Namitg02/Test", split='train', streaming=False)
#dataset = load_dataset("not-lain/wikipedia",revision = "embedded")
dataset = load_dataset("epfl-llm/guidelines"",)

#Returns a list of dictionaries, each representing a row in the dataset.
#print(dataset[1])
# splitter = RecursiveCharacterTextSplitter(chunk_size=150, chunk_overlap=25,separators=["\n"]) # ["\n\n", "\n", " ", ""])


#docs = splitter.create_documents(str(dataset))
# Returns a list of documents
#print(docs)
embedding_model = SentenceTransformer("mixedbread-ai/mxbai-embed-large-v1")
#all-MiniLM-L6-v2, BAAI/bge-base-en-v1.5,infgrad/stella-base-en-v2, BAAI/bge-large-en-v1.5 working with default dimensions
#docs_text = [doc.text for doc in docs]
#embed = embedding_model.embed_documents(docs_text)
embedding_dim = embedding_model.get_sentence_embedding_dimension()
print(embedding_dim)
#data = FAISS.from_embeddings(embed, embedding_model)
#data = FAISS.from_texts(docs, embedding_model)

# Returns a FAISS wrapper vector store. Input is a list of strings. from_documents method used documents to Return VectorStore

data = dataset["clean_text"]
#data = dataset["train"]

#print(data)
d = 384  # vectors dimension
m = 32  # hnsw parameter. Higher is more accurate but takes more time to index (default is 32, 128 should be ok)
#index = faiss.IndexHNSWFlat(d, m)
#index =  faiss.IndexFlatL2(embedding_dim)
#data.add_faiss_index(embeddings.shape[1], custom_index=index) 
data.add_faiss_index("embeddings")
# adds an index column that for the embeddings

print("check1")
#question = "How can I reverse Diabetes?"

SYS_PROMPT = """You are an assistant for answering questions.
You are given the extracted parts of a long document and a question. Provide a conversational answer.
If you don't know the answer, just say "I do not know." Don't make up an answer."""
# Provides context of how to answer the question

print("check2")


llm_model = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
tokenizer = AutoTokenizer.from_pretrained(llm_model)
# pulling tokeinzer for text generation model
model = AutoModelForCausalLM.from_pretrained(llm_model)
# Initializing the text generation model

terminators = [
    tokenizer.eos_token_id, # End-of-Sequence Token that indicates where the model should consider the text sequence to be complete
    tokenizer.convert_tokens_to_ids("<|eot_id|>") # Converts a token strings in a single/ sequence of integer id using the vocabulary
]
# indicates the end of a sequence


def search(query: str, k: int = 3 ):
    """a function that embeds a new query and returns the most probable results"""
    embedded_query = embedding_model.encode(query) # create embedding of a new query
    scores, retrieved_examples = data.get_nearest_examples( # retrieve results
        "embeddings", embedded_query, # compare our new embedded query with the dataset embeddings
        k=k # get only top k results
    )
    return scores, retrieved_examples
# returns scores (List[float]): the retrieval scores from either FAISS (IndexFlatL2 by default) and examples (dict) format
# called by talk function that passes prompt

#print(scores, retrieved_examples)
print("check2A")


def format_prompt(prompt,retrieved_documents,k):
    """using the retrieved documents we will prompt the model to generate our responses"""
    PROMPT = f"Question:{prompt}\nContext:"
    for idx in range(k) :
        PROMPT+= f"{retrieved_documents['text'][idx]}\n"
    return PROMPT

# Called by talk function to add retrieved documents to the prompt. Keeps adding text of retrieved documents to string taht are retreived

print("check3")
#print(PROMPT)

print("check3A")


def talk(prompt,history):
    k = 1 # number of retrieved documents
    scores , retrieved_documents = search(prompt, k) # get retrival scores and examples in dictionary format based on the promt passed
    formatted_prompt = format_prompt(prompt,retrieved_documents,k) # create a new prompt using the retrieved documents
    formatted_prompt = formatted_prompt[:400] # to avoid memory issue
    messages = [{"role":"system","content":SYS_PROMPT},{"role":"user","content":formatted_prompt}] # binding the system context and new prompt for LLM
    # the chat template structure should be based on text generation model format
    print("check3B")    
    input_ids = tokenizer.apply_chat_template(
      messages,
      add_generation_prompt=True,
      return_tensors="pt"
    ).to(model.device)
    # tell the model to generate
    # add_generation_prompt argument tells the template to add tokens that indicate the start of a bot response
    print("check3C")
    outputs = model.generate(
      input_ids,
      max_new_tokens=300,
      eos_token_id=terminators,
      do_sample=True,
      temperature=0.6,
      top_p=0.9,
    )
    # calling the model to generate response based on message/ input
    # do_sample if set to True uses strategies to select the next token from the probability distribution over the entire vocabulary
    # temperature controls randomness. more renadomness with higher temperature
    # only the tokens comprising the top_p probability mass are considered for responses
    # This output is a data structure containing all the information returned by generate(), but that can also be used as tuple or dictionary.
    print("check3D")
    streamer = TextIteratorStreamer(
            tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True
            )
    # stores print-ready text in a queue, to be used by a downstream application as an iterator. removes specail tokens in generated text. 
    # timeout for text queue. tokenizer for decoding tokens
    # called by generate_kwargs
    print("check3E")
    generate_kwargs = dict(
        input_ids= input_ids,
        streamer=streamer,
        max_new_tokens= 512,
        do_sample=True,
        top_p=0.95,
        temperature=0.75,
        eos_token_id=terminators,
    )
    # send additional parameters to model for generation
    print("check3F")
    t = Thread(target=model.generate, kwargs=generate_kwargs)
    # to process multiple instances
    t.start()
    # start a thread
    print("check3G")
    outputs = []
    for text in streamer:
        outputs.append(text)
        print(outputs)
        yield "".join(outputs)
        print("check3H")


TITLE = "AI Copilot for Diabetes Patients"

DESCRIPTION = ""

import gradio as gr
# Design chatbot
demo = gr.ChatInterface(
    fn=talk,
    chatbot=gr.Chatbot(
        show_label=True,
        show_share_button=True,
        show_copy_button=True,
        likeable=True,
        layout="bubble",
        bubble_full_width=False,
    ),
    theme="Soft",
    examples=[["what is Diabetes? "]],
    title=TITLE,
    description=DESCRIPTION,
    
)
# launch chatbot and calls the talk function which in turn calls other functions
print("check3I")
demo.launch(share=True)