scripts/config.py

import os
import google.generativeai as genai
from llama_index.embeddings.huggingface import HuggingFaceEmbedding
from llama_index.llms.google_genai import GoogleGenAI
from llama_index.core import Settings
from llama_index.core.llms import ChatMessage, MessageRole
import os



EMBEDDING_MODEL = "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"
RETRIEVER_TOP_K = 10
RETRIEVER_SIMILARITY_CUTOFF = 0.7
RAG_FILES_DIR = "processed_data"
PROCESSED_DATA_FILE = "processed_data/processed_chunks.csv"

UPLOAD_FOLDER = "UPLOADED_DOCUMENTS"
PROCESSED_DATA_FILE = "processed_data/processed_chunks.csv"
INDEX_STATE_FILE = "processed_data/index_store.json"
RAG_FILES_DIR = "rag_files"

GOOGLE_API_KEY = os.getenv('GOOGLE_API_KEY')
LLM_MODEL = "gemini-2.5-flash"



CHUNK_SIZE = 1024
CHUNK_OVERLAP = 256
MAX_CHUNK_SIZE = 2048
MIN_CHUNK_SIZE = 750
SIMILARITY_THRESHOLD = 0.7


RETRIEVER_TOP_K = 15
RETRIEVER_SIMILARITY_CUTOFF = 0.7

CUSTOM_PROMPT = """
You are a highly specialized Document Analysis Assistant (AIEXP). Your purpose is to provide precise, accurate, and contextually relevant answers by analyzing a set of normal regulatory documents (НД). Your responses must be entirely based on the provided context, without any external knowledge or assumptions.

Core Tasks:
Based on the user's query, perform one of the following tasks:
- Information Retrieval: Find and present specific information.
- Summarization: Provide a concise summary of a document or a section.
- Semantic Analysis: Compare a provided text against the requirements of the ND.
- Action Planning: Create a step-by-step plan based on ND requirements.

Strict Rules for Response Generation:
1. Source Attribution is Mandatory: Every answer must explicitly cite its source from the provided context. Use one of the following formats:
   - For content from a specific section/subsection:
     Согласно разделу [X] и подразделу [X.X]: [Ваш ответ]
   - For content that is not part of a specific subsection (e.g., from a general section, table, or figure):
     Согласно [Название документа] - [Номер и наименование пункта/таблицы/изображения]: [Ваш ответ]
   - If the source chunk has metadata for both section and subsection, always include both.
   - If the source chunk has only a section, use the format Согласно разделу [X]: [Ваш ответ].

2. No Hallucinations: If the requested information is not explicitly found within the provided context, you must state that the information is not available. Do not attempt to infer, guess, or create a response. The correct response in this case is:
   Информация по вашему запросу не была найдена в нормативной документации.

3. Use ND Language: When possible, use terminology and phrasing directly from the ND to maintain accuracy and fidelity to the source document.

4. Prioritize Precision: When answering, provide the most specific and direct information possible, avoiding vague or overly broad summaries unless explicitly asked to summarize.

Context:
{context_str}

Question:
{query_str}

Answer:
"""

def setup_llm_settings():
    Settings.embed_model = HuggingFaceEmbedding(model_name=EMBEDDING_MODEL)
    Settings.llm = GoogleGenAI(model=LLM_MODEL, api_key=GOOGLE_API_KEY)
    Settings.llm.system_prompt = CUSTOM_PROMPT


LLM_MODEL_PREPROCESS = "gemini-1.5-flash"

def preprocess_query_with_context(user_query, chat_history=None, llm=None):
    if not chat_history:
        return user_query
    
    if not llm:
        llm = GoogleGenAI(model=LLM_MODEL_PREPROCESS, temperature=0.1)

    # Format chat history into a string for the prompt
    history_context = "\n".join([
        f"User: {item['user']}\nAssistant: {item['assistant']}"
        for item in chat_history[-3:]  # Consider only the last 3 exchanges for conciseness
    ])

    preprocessing_prompt = f"""Analyze the user's current question in the context of their chat history and improve it for better document retrieval.

Chat History:
{history_context}

Current Question: {user_query}

Tasks:
1. If the question refers to previous context (e.g., "what about that law?"), make it self-contained.
2. Add relevant keywords that would help find documents.
3. Maintain the legal/regulatory focus.
4. Keep it concise but specific.

Return ONLY the improved question:
"""

    try:
        # Create the message and get a response from the LLM
        messages = [ChatMessage(role=MessageRole.USER, content=preprocessing_prompt)]
        response = llm.chat(messages)
        improved_query = response.message.content.strip()

        # Fallback to the original query if the preprocessing fails or provides an overly long response
        if len(improved_query) > len(user_query) * 3 or not improved_query:
            return user_query
        
        return improved_query
    except Exception as e:
        print(f"Query preprocessing failed: {e}")
        return user_query


def create_chat_context_prompt(base_response, chat_history=None):
    if not chat_history:
        return base_response

    base_aware_response = base_response

    if len(chat_history) > 0:
        last_exchange = chat_history[-1]
        if any(keyword in last_exchange['user'].lower() for keyword in ['закон', 'кодекс', 'статья']):
            # Add a conversational prefix
            base_aware_response = f"Продолжая тему нормативных документов: {base_response}"

    return base_aware_response