vdb.py

import hashlib
import json
import re
from pathlib import Path

from dotenv import load_dotenv
from llama_index.core import (QueryBundle)
from llama_index.core.postprocessor import LLMRerank
from nest_asyncio import apply
from openai import OpenAI
from tqdm import tqdm

from llama_index.core import VectorStoreIndex, Settings
from llama_index.embeddings.openai import OpenAIEmbedding

from llama_index.embeddings.huggingface import HuggingFaceEmbedding

from llama_index.core import Document


embed_model = HuggingFaceEmbedding(
        model_name="sentence-transformers/all-MiniLM-L6-v2"
    )
Settings.embed_model = embed_model

# Load variables from .env
load_dotenv()

def build_documents(sections):
    docs = []
    for s in sections:
        metadata = {"section_title": s["title"]}
        docs.append(Document(text=s["content"], metadata=metadata))
    return docs

def create_vector_index(docs):
    # embed_model = OpenAIEmbedding()
    # index = VectorStoreIndex.from_documents(docs, embed_model=embed_model)
    index = VectorStoreIndex.from_documents(docs)
    return index

def split_markdown_by_section(md_path: str):
    text = Path(md_path).read_text(encoding="utf-8")
    sections = re.split(r"(?m)^# ", text)
    chunks = []
    for section in sections:
        if not section.strip():
            continue
        title, *content = section.split("\n", 1)
        body = content[0].strip() if content else ""
        chunks.append({"title": title.strip(), "content": body})
    return chunks



client = OpenAI()

apply()

tqdm.pandas()


def hash_data(data):
    json_str = json.dumps(data, sort_keys=True)

    json_bytes = json_str.encode('utf-8')

    hash_hex = hashlib.sha256(json_bytes).hexdigest()

    return hash_hex


def get_retrieved_nodes(query, index, vector_top_k=10, reranker_top_n=3, with_reranker=True):
    query_bundle = QueryBundle(query)
    retriever = index.as_retriever(similarity_top_k=vector_top_k)
    retrieved_nodes = retriever.retrieve(query_bundle)

    if with_reranker:
        reranker = LLMRerank(choice_batch_size=5, top_n=reranker_top_n)
        retrieved_nodes = reranker.postprocess_nodes(retrieved_nodes, query_bundle)

    return retrieved_nodes


def get_all_text(nodes):
    return ' '.join(f"\n- {node.get_text()}" for node in nodes)


async def further_retrieve(query, index, messages):
    try:
        retrieved_nodes = get_retrieved_nodes(query, index, vector_top_k=10, reranker_top_n=3, with_reranker=False)
        return completion(query, get_all_text(retrieved_nodes), messages)
    except Exception as e:
        print(e)
        return None


async def completion(query, docs, messages):
    messages.extend([
        {
            "role": "system",
            "content": f"""
Given tone and voice guidelines and customer support help documents, act as a customer support bot. 
Answer any further questions as if you are customer support bot.
TONE AND VOICE:
promote the society, be gentle, be kind always positive.

DOCUMENT:
{docs}



INSTRUCTIONS:

- Answer the users QUESTION using the DOCUMENT text above.
- Format formula into latex format between $...$ or \[...\]
- Keep your answer ground in the facts of the DOCUMENT or chat history.
- If document has an image markdown ,use it in your answer
- Respond in same language as user Question
- Use Markdown Structure
- DOCUMENT can have images with there descriptions
- if a text is followed by an image dont skip  the image
QUESTION:
              """
        },
        {
            "role": "system",
            "content": query
        }
    ])
    completion = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=messages,
        stream=True
    )
    for chunk in completion:
        if chunk.choices[0].delta.content:
            yield chunk.choices[0].delta.content