import os, warnings
from dotenv import load_dotenv
from schemas import *

os.environ["CURL_CA_BUNDLE"] = ""
warnings.filterwarnings("ignore")
load_dotenv()

from datasets import load_dataset
import bm25s
from bm25s.hf import BM25HF

from fastapi import FastAPI
from fastapi.middleware.cors import CORSMiddleware
from fastapi.responses import FileResponse
from fastapi.staticfiles import StaticFiles

from sklearn.preprocessing import MinMaxScaler
import numpy as np

import litellm

bm25_index = BM25HF.load_from_hub("OrganizedProgrammers/3GPPBM25IndexSections", load_corpus=True, token=os.environ["HF_TOKEN"])

app = FastAPI(title="RAGnarok", 
              description="Speak with the specifications")

app.mount("/static", StaticFiles(directory="static"), name="static")

origins = [
    "*",
]

app.add_middleware(
    CORSMiddleware,
    allow_origins=origins,
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

@app.get("/")
def main_menu():
    return FileResponse(os.path.join("templates", "index.html"))

@app.post("/search", response_model=SearchResponse)
def search_specifications(req: SearchRequest):
    keywords = req.keyword
    threshold = req.threshold
    results_out = []
    query_tokens = bm25s.tokenize(keywords)
    results, scores = bm25_index.retrieve(query_tokens, k=len(bm25_index.corpus))

    def calculate_boosted_score(metadata, score, query):
        title = set(metadata['title'].lower().split())
        q = set(query.lower().split())
        spec_id_presence = 0.5 if metadata['id'].lower() in q else 0
        booster = len(q & title) * 0.5
        return score + spec_id_presence + booster

    spec_scores = {}
    spec_indices = {}
    spec_details = {}

    for i in range(results.shape[1]):
        doc = results[0, i]
        score = scores[0, i]
        spec = doc["metadata"]["id"]

        boosted_score = calculate_boosted_score(doc['metadata'], score, keywords)

        if spec not in spec_scores or boosted_score > spec_scores[spec]:
            spec_scores[spec] = boosted_score
            spec_indices[spec] = i
            spec_details[spec] = {
                'original_score': score,
                'boosted_score': boosted_score,
                'doc': doc
            }

    def normalize_scores(scores_dict):
        if not scores_dict:
            return {}
            
        scores_array = np.array(list(scores_dict.values())).reshape(-1, 1)
        scaler = MinMaxScaler()
        normalized_scores = scaler.fit_transform(scores_array).flatten()
            
        normalized_dict = {}
        for i, spec in enumerate(scores_dict.keys()):
            normalized_dict[spec] = normalized_scores[i]
            
        return normalized_dict

    normalized_scores = normalize_scores(spec_scores)

    for spec in spec_details:
        spec_details[spec]["normalized_score"] = normalized_scores[spec]

    unique_specs = sorted(normalized_scores.keys(), key=lambda x: normalized_scores[x], reverse=True)

    for rank, spec in enumerate(unique_specs, 1):
        details = spec_details[spec]
        metadata = details['doc']['metadata']
        if details['normalized_score'] < threshold / 100:
            break
        results_out.append({'id': metadata['id'], 'title': metadata['title'], 'section': metadata['section_title'], 'content': details['doc']['text'], 'similarity': int(details['normalized_score']*100)})
    
    return SearchResponse(results=results_out)

@app.post("/chat", response_model=ChatResponse)
def questions_the_sources(req: ChatRequest):
    model = req.model
    resp = litellm.completion(
        model=f"gemini/{model}",
        messages=req.messages,
        api_key=os.environ["GEMINI"]
    )
    return ChatResponse(response=resp.choices[0].message.content)