yijinaembed.py

# app.py
import os
import re
import uuid
import gradio as gr
import torch
import torch.nn.functional as F
from dotenv import load_dotenv
from typing import List, Tuple, Dict, Any
from transformers import AutoTokenizer, AutoModel
from openai import OpenAI
from langchain_community.document_loaders import UnstructuredFileLoader
from langchain_chroma import Chroma
from chromadb import Documents, EmbeddingFunction, Embeddings
from chromadb.config import Settings
import chromadb
from utils import load_env_variables, parse_and_route, escape_special_characters
from globalvars import API_BASE, intention_prompt, tasks, system_message, metadata_prompt, model_name
import spaces
from langchain.retrievers.contextual_compression import ContextualCompressionRetriever
from langchain_community.document_compressors.jina_rerank import JinaRerank
from langchain import hub
from langchain.chains.retrieval import create_retrieval_chain
from langchain.chains.combine_documents.stuff  import create_stuff_documents_chain

load_dotenv()

# os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'max_split_size_mb:180'
# os.environ['CUDA_LAUNCH_BLOCKING'] = '1'
# os.environ['CUDA_CACHE_DISABLE'] = '1'

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

hf_token, yi_token = load_env_variables()

tokenizer = AutoTokenizer.from_pretrained(model_name, token=hf_token, trust_remote_code=True)
model = None

@spaces.GPU
def load_model():
    global model
    if model is None:
        model = AutoModel.from_pretrained(model_name, token=hf_token, trust_remote_code=True).to(device)
    return model

# Load model
jina_model = load_model()

def clear_cuda_cache():
    torch.cuda.empty_cache()

client = OpenAI(api_key=yi_token, base_url=API_BASE)

chroma_client = chromadb.Client(Settings())

chroma_collection = chroma_client.create_collection("all-my-documents")

class JinaEmbeddingFunction(EmbeddingFunction):
    def __init__(self, model, tokenizer, intention_client):
        self.model = model
        self.tokenizer = tokenizer
        self.intention_client = intention_client

    def __call__(self, input: Documents) -> Tuple[List[List[float]], List[Dict[str, Any]]]:
        embeddings_with_metadata = [self.compute_embeddings(doc) for doc in input]
        embeddings = [item[0] for item in embeddings_with_metadata]
        metadata = [item[1] for item in embeddings_with_metadata]
        return embeddings, metadata

    @spaces.GPU
    def compute_embeddings(self, input_text: str):
        escaped_input_text = escape_special_characters(input_text)

        # Get the intention
        intention_completion = self.intention_client.chat.completions.create(
            model="yi-large",
            messages=[
                {"role": "system", "content": escape_special_characters(intention_prompt)},
                {"role": "user", "content": escaped_input_text}
            ]
        )
        intention_output = intention_completion.choices[0].message.content
        parsed_task = parse_and_route(intention_output)
        selected_task = parsed_task if parsed_task in tasks else "DEFAULT"
        task = tasks[selected_task]

        # Get the metadata
        metadata_completion = self.intention_client.chat.completions.create(
            model="yi-large",
            messages=[
                {"role": "system", "content": escape_special_characters(metadata_prompt)},
                {"role": "user", "content": escaped_input_text}
            ]
        )
        metadata_output = metadata_completion.choices[0].message.content
        metadata = self.extract_metadata(metadata_output)

        # Compute embeddings using Jina model
        encoded_input = self.tokenizer(escaped_input_text, padding=True, truncation=True, return_tensors="pt").to(device)
        with torch.no_grad():
            model_output = self.model(**encoded_input, task=task)

        embeddings = self.mean_pooling(model_output, encoded_input["attention_mask"])
        embeddings = F.normalize(embeddings, p=2, dim=1)

        return embeddings.cpu().numpy().tolist()[0], metadata

    def extract_metadata(self, metadata_output: str) -> Dict[str, str]:
        pattern = re.compile(r'\"(\w+)\": \"([^\"]+)\"')
        matches = pattern.findall(metadata_output)
        metadata = {key: value for key, value in matches}
        return metadata

    @staticmethod
    def mean_pooling(model_output, attention_mask):
        token_embeddings = model_output[0]
        input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
        return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)

def load_documents(file_path: str, mode: str = "elements"):
    loader = UnstructuredFileLoader(file_path, mode=mode)
    docs = loader.load()
    return [doc.page_content for doc in docs]

def initialize_chroma(collection_name: str, embedding_function: JinaEmbeddingFunction):
    db = Chroma(client=chroma_client, collection_name=collection_name, embedding_function=embedding_function)
    return db

@spaces.GPU
def add_documents_to_chroma(documents: list, embedding_function: JinaEmbeddingFunction):
    for doc in documents:
        embeddings, metadata = embedding_function.compute_embeddings(doc)
        chroma_collection.add(
            ids=[str(uuid.uuid1())],
            documents=[doc],
            embeddings=[embeddings],
            metadatas=[metadata]
        )

@spaces.GPU
def rerank_documents(query: str, documents: List[str]) -> List[str]:
    compressor = JinaRerank()
    retriever = chroma_db.as_retriever(search_kwargs={"k": 15})
    compression_retriever = ContextualCompressionRetriever(
        base_compressor=compressor, base_retriever=retriever
    )
    
    compressed_docs = compression_retriever.get_relevant_documents(query)
    
    return [doc.page_content for doc in compressed_docs]

def query_chroma(query_text: str, embedding_function: JinaEmbeddingFunction):
    query_embeddings, query_metadata = embedding_function.compute_embeddings(query_text)
    result_docs = chroma_collection.query(
        query_embeddings=[query_embeddings],
        n_results=5
    )
    return result_docs

@spaces.GPU
def answer_query(message: str, chat_history: List[Tuple[str, str]], system_message: str, max_new_tokens: int, temperature: float, top_p: float):
    # Query Chroma for relevant documents
    results = query_chroma(message, embedding_function)
    context = "\n\n".join([result['document'] for result in results['documents'][0]])

    # Rerank the documents
    reranked_docs = rerank_documents(message, context.split("\n\n"))
    reranked_context = "\n\n".join(reranked_docs)

    # Prepare the prompt for YI model
    prompt = f"{system_message}\n\nContext: {reranked_context}\n\nHuman: {message}\n\nAssistant:"

    # Generate response using YI model
    response = client.chat.completions.create(
        model="yi-large",
        messages=[
            {"role": "system", "content": system_message},
            {"role": "user", "content": f"Context: {reranked_context}\n\nHuman: {message}"}
        ],
        max_tokens=max_new_tokens,
        temperature=temperature,
        top_p=top_p
    )

    assistant_response = response.choices[0].message.content
    chat_history.append((message, assistant_response))
    return "", chat_history

# Initialize clients
intention_client = OpenAI(api_key=yi_token, base_url=API_BASE)
embedding_function = JinaEmbeddingFunction(jina_model, tokenizer, intention_client)
chroma_db = initialize_chroma(collection_name="Jina-embeddings", embedding_function=embedding_function)

@spaces.GPU
def upload_documents(files):
    for file in files:
        loader = UnstructuredFileLoader(file.name)
        documents = loader.load()
        add_documents_to_chroma([doc.page_content for doc in documents], embedding_function)
    return "Documents uploaded and processed successfully!"

@spaces.GPU
def query_documents(query):
    results = query_chroma(query, embedding_function)
    reranked_docs = rerank_documents(query, [result for result in results['documents'][0]])
    return "\n\n".join(reranked_docs)

with gr.Blocks() as demo:
    with gr.Tab("Upload Documents"):
        document_upload = gr.File(file_count="multiple", file_types=["document"])
        upload_button = gr.Button("Upload and Process")
        upload_button.click(upload_documents, inputs=document_upload, outputs=gr.Text())

    with gr.Tab("Ask Questions"):
        with gr.Row():
            chat_interface = gr.ChatInterface(
                answer_query,
                additional_inputs=[
                    gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
                    gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
                    gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
                    gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p (nucleus sampling)"),
                ],
            )
            query_input = gr.Textbox(label="Query")
            query_button = gr.Button("Query")
            query_output = gr.Textbox()
            query_button.click(query_documents, inputs=query_input, outputs=query_output)

if __name__ == "__main__":
    demo.launch()