app.py

from typing import List, Dict, Any
import zipfile
import os
import warnings
from openai import OpenAI
from dotenv import load_dotenv
import bm25s
from fastapi.staticfiles import StaticFiles
from nltk.stem import WordNetLemmatizer
import nltk
from fastapi import FastAPI
from fastapi.responses import FileResponse
from fastapi.middleware.cors import CORSMiddleware
import numpy as np
from pydantic import BaseModel
from sklearn.preprocessing import MinMaxScaler

load_dotenv()

nltk.download('wordnet')
if os.path.exists("bm25s.zip"):
    with zipfile.ZipFile("bm25s.zip", 'r') as zip_ref:
        zip_ref.extractall(".")
    bm25_engine = bm25s.BM25.load("3gpp_bm25_docs", load_corpus=True)
lemmatizer = WordNetLemmatizer()
llm = OpenAI(api_key=os.environ.get("GEMINI"), base_url="https://generativelanguage.googleapis.com/v1beta/openai/")

warnings.filterwarnings("ignore")

app = FastAPI(title="RAGnarok", 
              description="API to search specifications for RAG")

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

origins = [
    "*",
]

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

class SearchRequest(BaseModel):
    keyword: str
    threshold: int

class SearchResponse(BaseModel):
    results: List[Dict[str, Any]]

class ChatRequest(BaseModel):
    messages: List[Dict[str, str]]
    model: str

class ChatResponse(BaseModel):
    response: str

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

@app.post("/chat", response_model=ChatResponse)
def question_the_sources(req: ChatRequest):
    model = req.model
    resp = llm.chat.completions.create(
        messages=req.messages,
        model=model
    )
    return ChatResponse(response=resp.choices[0].message.content)

@app.post("/search", response_model=SearchResponse)
def search_specifications(req: SearchRequest):
    keywords = req.keyword
    threshold = req.threshold
    query = lemmatizer.lemmatize(keywords)
    results_out = []
    query_tokens = bm25s.tokenize(query)
    results, scores = bm25_engine.retrieve(query_tokens, k=len(bm25_engine.corpus))

    def calculate_boosted_score(metadata, score, query):
        title = {lemmatizer.lemmatize(metadata['title']).lower()}
        q = {query.lower()}
        spec_id_presence = 0.5 if len(q & {metadata['id']}) > 0 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, query)

        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)