import os
import re
import time
import msgpack
import numpy as np
import streamlit as st
from io import StringIO
from numpy import ndarray
from langchain_core.documents.base import Document
from scipy.sparse import csr_matrix, save_npz, load_npz, vstack
from qdrant_client import QdrantClient, models
from langchain_community.document_loaders.wikipedia import WikipediaLoader
from langchain_experimental.text_splitter import SemanticChunker
from fastembed.sparse.splade_pp import supported_splade_models
from fastembed import SparseTextEmbedding, SparseEmbedding
from fastembed_ext import FastEmbedEmbeddingsLc
from huggingface_hub import hf_hub_download
from unstructured.partition.auto import partition
from qdrant_client.models import (
    NamedSparseVector,
    NamedVector,
    SparseVector,
    PointStruct,
    SearchRequest,
    ScoredPoint,
)
from llama_cpp import Llama

MAP_PROMPT = """
You will be given a single passage of a book. This section will be enclosed in triple backticks (```)
Your goal is to give a summary of this section so that a reader will have a full understanding of what happened.
Your response should be at least three paragraphs and fully encompass what said in the passage.

```{text}```
FULL SUMMARY:
"""

COMBINE_PROMPT = """
You will be given a series of summaries from a book. The summaries will be enclosed in triple backticks (```)
Your goal is to give a verbose summary of what happened in the story.
The reader should be able to grasp what happened in the book.

```{text}```
VERBOSE SUMMARY:
"""


def make_points(chunks: list[str], dense: list[ndarray], sparse: list[SparseEmbedding])-> list[PointStruct]:
    points = []
    for idx, (sparse_vector, chunk, dense_vector) in enumerate(zip(sparse, chunks, dense)):
        sparse_vec = SparseVector(indices=sparse_vector.indices.tolist(), values=sparse_vector.values.tolist())
        point = PointStruct(
            id=idx,
            vector={
                "text-sparse": sparse_vec,
                "text-dense": dense_vector,
            },
            payload={
                "text": chunk
            }
        )
        points.append(point)
    return points

def search(client: QdrantClient, collection_name: str, dense, sparse):
    search_results = client.search_batch(
        collection_name,
        [
            SearchRequest(
                vector=NamedVector(
                    name="text-dense",
                    vector=dense,
                ),
                limit=10
            ),
            SearchRequest(
                vector=NamedSparseVector(
                    name="text-sparse",
                    vector=SparseVector(
                        indices=sparse[0].indices.tolist(),
                        values=sparse[0].values.tolist(),
                    ),
                ),
                limit=10
            ),
        ],
    )

    return search_results

def rank_list(search_result: list[ScoredPoint]):
    return [(point.id, rank + 1) for rank, point in enumerate(search_result)]

def rrf(rank_lists, alpha=60, default_rank=1000):
    """
    Optimized Reciprocal Rank Fusion (RRF) using NumPy for large rank lists.

    :param rank_lists: A list of rank lists. Each rank list should be a list of (item, rank) tuples.
    :param alpha: The parameter alpha used in the RRF formula. Default is 60.
    :param default_rank: The default rank assigned to items not present in a rank list. Default is 1000.
    :return: Sorted list of items based on their RRF scores.
    """
    all_items = set(item for rank_list in rank_lists for item, _ in rank_list)
    item_to_index = {item: idx for idx, item in enumerate(all_items)}
    rank_matrix = np.full((len(all_items), len(rank_lists)), default_rank)

    for list_idx, rank_list in enumerate(rank_lists):
        for item, rank in rank_list:
            rank_matrix[item_to_index[item], list_idx] = rank

    rrf_scores = np.sum(1.0 / (alpha + rank_matrix), axis=1)
    sorted_indices = np.argsort(-rrf_scores)
    sorted_items = [(list(item_to_index.keys())[idx], rrf_scores[idx]) for idx in sorted_indices]

    return sorted_items


def main(query: str, client: QdrantClient, collection_name: str, llm, dense_model, sparse_model):
    dense_query = list(dense_model.embed_query(query, 32))
    sparse_query = list(sparse_model.embed(query, 32))

    search_results = search(
        client,
        collection_name,
        dense_query,
        sparse_query
    )

    dense_rank_list, sparse_rank_list = rank_list(search_results[0]), rank_list(search_results[1])
    rrf_rank_list = rrf([dense_rank_list, sparse_rank_list])

    records_list = client.retrieve(
        collection_name,
        [item[0] for item in rrf_rank_list]
    )
    
    docs = [record.payload['text'] for record in records_list[:3]]

    combined_docs = "\n".join(docs)

    response = llm.create_chat_completion(
        messages = [
          {"role": "system", "content": f"""Use the following pieces of context to answer the user question.
          If you don't know the answer, just say that you don't know, don't try to make up an answer.
          Use three sentences maximum and keep the answer as concise as possible.
          
          {combined_docs}"""
          },
          {
              "role": "user",
              "content": f"Question: {query}"
          }
      ], stop=["</s>"], temperature=0.7)

    text = response["choices"][0]["message"]['content']
    print(f'TEXT: {text}')

    output = llm.create_chat_completion(
        messages = [
          {"role": "system", "content": """You are an assistant that provides summaries of texts. 
          Your task is to create a summary that includes the main points and any important details. 
          """
          },
          {
              "role": "user",
              "content": f"""Write a summary of the following text delimited by triple backquotes. Ensure the summary covers the key points of the text. Avoid introducing any sentences before the summary.
              ```{text}```"""
          }
      ], stop=["</s>"], temperature=0.7, max_tokens=3000)['choices'][0]['message']['content']
    print(f'OUTPUT: {output}')
    return output

@st.cache_resource
def load_models_and_documents():
    print('load')
    supported_splade_models[0] = {
        "model": "prithivida/Splade_PP_en_v2",
        "vocab_size": 30522,
        "description": "Implementation of SPLADE++ Model for English v2",
        "size_in_GB": 0.532,
        "sources": {
            "hf": "devve1/Splade_PP_en_v2_onnx"
        },
        "model_file": "model.onnx"
    }
    
    with st.spinner('Load models...'):
        llm = Llama.from_pretrained(
            repo_id="MaziyarPanahi/Llama-3-8B-Instruct-32k-v0.1-GGUF",
            filename="*Q8_0.gguf",
            verbose=False,
            chat_format="chatml",
            n_ctx=16000,
            n_gpu_layers=32
        )

        provider = ['CPUExecutionProvider']
        
        dense_model = FastEmbedEmbeddingsLc(
            model_name='mixedbread-ai/mxbai-embed-large-v1',
            providers=provider,
            cache_dir=os.getenv('HF_HOME'),
            batch_size=32
        )
        
        sparse_model = SparseTextEmbedding(
            'prithivida/Splade_PP_en_v2',
            cache_dir=os.getenv('HF_HOME'),
            providers=provider
        )

    client = QdrantClient(':memory:')
    collection_name = 'collection_demo'
    
    client.create_collection(
        collection_name,
        {
            "text-dense": models.VectorParams(
                size=1024,
                distance=models.Distance.COSINE,
                on_disk=False,
                quantization_config=models.BinaryQuantization(
                    binary=models.BinaryQuantizationConfig(
                        always_ram=True
                    )
                )
            )
        },
        {
            "text-sparse": models.SparseVectorParams(
                index=models.SparseIndexParams(
                    on_disk=False
                )
            )
        },
        2,
        optimizers_config=models.OptimizersConfigDiff(
            indexing_threshold=0
        ),
        hnsw_config=models.HnswConfigDiff(
            on_disk=False, 
            m=16, 
            ef_construct=100
        )
    )
        
    with st.spinner('Parse and chunk documents...'):
        name = 'action_rpg'
        embeddings_path = os.path.join(os.getenv('HF_HOME'), 'embeddings')

        chunks_path = os.path.join(embeddings_path, name + '_chunks.msgpack')
        dense_path = os.path.join(embeddings_path, name + '_dense.npz')
        sparse_path = os.path.join(embeddings_path, name + '_sparse.npz')

        if not os.path.exists(embeddings_path):
            os.mkdir(embeddings_path)

            docs_1 = WikipediaLoader(query='Action-RPG').load()
            docs_2 = WikipediaLoader(query='Dark Souls').load()
            docs_3 = WikipediaLoader(query='Zelda II: The Adventure of Link').load()
            #docs_4 = WikipediaLoader(query='The Witcher').load()
            #docs_5 = WikipediaLoader(query='Genshin Impact ').load()
            #docs_6 = WikipediaLoader(query='The Elder Scrolls').load()
            #docs_7 = WikipediaLoader(query='Fallout').load()
            docs = docs_1 + docs_2 + docs_3 #+ docs_4 + docs_5 + docs_6 + docs_7
            chunks, dense_embeddings, sparse_embeddings = chunk_documents(docs, dense_model, sparse_model)

            with open(chunks_path, "wb") as outfile:
                packed = msgpack.packb(chunks, use_bin_type=True)
                outfile.write(packed)
            
            np.savez_compressed(dense_path, *dense_embeddings)
            max_index = max(np.max(embedding.indices) for embedding in sparse_embeddings)
        
            sparse_matrices = []
            for embedding in sparse_embeddings:
                data = embedding.values
                indices = embedding.indices
                indptr = np.array([0, len(data)])
                matrix = csr_matrix((data, indices, indptr), shape=(1, max_index + 1))
                sparse_matrices.append(matrix)
            
            combined_sparse_matrix = vstack(sparse_matrices)
            save_npz(sparse_path, combined_sparse_matrix)
        else:
            with open(chunks_path, "rb") as data_file:
                byte_data = data_file.read()
                
            chunks = msgpack.unpackb(byte_data, raw=False)

            dense_embeddings = list(np.load(dense_path).values())

            sparse_embeddings = []
            loaded_sparse_matrix = load_npz(sparse_path)
        
            for i in range(loaded_sparse_matrix.shape[0]):
                row = loaded_sparse_matrix.getrow(i)
                values = row.data
                indices = row.indices
                embedding = SparseEmbedding(values, indices)
                sparse_embeddings.append(embedding)

    with st.spinner('Save documents...'):
        client.upsert(
            collection_name,
            make_points(
                chunks, 
                dense_embeddings, 
                sparse_embeddings
            )
        )
    client.update_collection(
        collection_name=collection_name,
        optimizer_config=models.OptimizersConfigDiff(indexing_threshold=20000)
    )

    return client, collection_name, llm, dense_model, sparse_model

def chunk_documents(docs, dense_model, sparse_model):
    text_splitter = SemanticChunker(
        dense_model,
        breakpoint_threshold_type='standard_deviation'
    )

    documents = [doc.page_content for doc in text_splitter.transform_documents(list(docs))]

    dense_embeddings = dense_model.embed_documents(documents,32)
    sparse_embeddings = list(sparse_model.embed(documents, 32))
    
    return documents, dense_embeddings, sparse_embeddings
    
if __name__ == '__main__':
    st.set_page_config(page_title="Video Game Assistant",
                       layout="wide"
                       )
    st.title("Video Game Assistant")
    
    client, collection_name, llm, dense_model, sparse_model = load_models_and_documents()

    if "messages" not in st.session_state:
        st.session_state.messages = []

    for message in st.session_state.messages:
        with st.chat_message(message["role"]):
            st.markdown(message["content"])

    if prompt := st.chat_input("Message Video Game Assistant"):
        st.chat_message("user").markdown(prompt)
        st.session_state.messages.append({"role": "user", "content": prompt})

        ai_response = main(prompt, client, collection_name, llm, dense_model, sparse_model)
        response = f"Echo: {ai_response}"
        with st.chat_message("assistant"):
            message_placeholder = st.empty()
            full_response = ""
            for chunk in re.split(r'(\s+)', response):
                full_response += chunk + " "
                time.sleep(0.01)
                message_placeholder.markdown(full_response + "▌")
        st.session_state.messages.append({"role": "assistant", "content": full_response})

    st.sidebar.title("Upload your file")
    uploaded_files = st.sidebar.file_uploader("Choose a file", accept_multiple_files=True, type=['docx', 'doc', 'odt', 'pptx', 'ppt', 'xlsx', 'csv', 'tsv', 'eml', 'msg', 'rtf', 'epub', 'html', 'xml', 'pdf', 'png', 'jpg', 'heic','txt'])
    print(f'uploaded-files : {uploaded_files}')
    for uploaded_file in uploaded_files:
        print('count')
        elements = partition(file=uploaded_file,
                             strategy='hi_res',
                             skip_infer_table_types=['png', 'pdf', 'jpg', 'xls', 'xlsx', 'heic'],
                             hi_res_model_name='yolox',
                             include_page_breaks=True
                            )
        docs = [Document(elem.text) for elem in elements]

        chunks, dense_embeddings, sparse_embeddings = chunk_documents(docs, dense_model, sparse_model)
        client.upsert(
            collection_name,
            make_points(
                chunks, 
                dense_embeddings, 
                sparse_embeddings
            )
        )