Best practices for bulk metadata enrichment in Document insertion for Vector Search

Generative AI Vector Search for LLMs
, , ,
1

Hello everyone,

I am working on a project where I need to add a custom field (target_id) to the metadata of a large set of documents before inserting them into MongoDB for vector search indexing. Here is the approach I currently use:

for doc in docs:
    doc.metadata["target_id"] = target_id
# Insert docs into MongoDB and proceed with vector search indexing

This method involves looping over each document to insert the custom field, which seems inefficient for large datasets.

So my question is: is there a better way to populate the vector database with custom metadata field rather than loop docs

here is the full code i implemented

from flask import Blueprint, request, jsonify

from langchain.text_splitter import RecursiveCharacterTextSplitter

from .db import MongoDBHandler

load_blueprint = Blueprint('upload', __name__)

class UnsupportedDocumentType(Exception):
    pass

@load_blueprint.route('/load', methods=['POST'])
def load():
    body = request.json

    type = body.get('type')
    url = body.get('url')
    target_id = body.get('target_id')

    try:
        docs = load_document(type, url)

        for doc in docs:
            doc.metadata["target_id"] = target_id

        db_handler = MongoDBHandler()
        db_handler.vector_search_from_documents(
            "suppalism",
            "knowledges",
            "knowledge_vector_search_index",
            docs
        )

        return jsonify({"message": "Processed "+type+" file"})
    except Exception as e:
        return jsonify({"status": 500, "message": str(e)})
    
def load_document(type, url):
    loaders = {
        'pdf': 'langchain.document_loaders.PyPDFLoader',
        'doc': 'langchain_community.document_loaders.UnstructuredURLLoader',
        'video': 'langchain_community.document_loaders.YoutubeLoader',
        'website': 'langchain_community.document_loaders.WebBaseLoader'
    }

    if type not in loaders:
        raise UnsupportedDocumentType('Document format is not supported!')

    module_name, class_name = loaders[type].rsplit('.', 1)
    LoaderClass = getattr(__import__(module_name, fromlist=[class_name]), class_name)

    if type == 'video':
        loader = LoaderClass.from_youtube_url(url, add_video_info=True)
    elif type == "website":
        loader = LoaderClass(url)
    elif type == 'doc':
        loader = LoaderClass(urls=[url])
    elif type == 'pdf':
        loader = LoaderClass(url)

    data = loader.load()

    text_splitter = RecursiveCharacterTextSplitter(
        chunk_size=200,
        chunk_overlap=20,
    )
    docs = text_splitter.split_documents(data)

    return docs
2

Hi, @work_state, I checked with the LangChain team, and the use of the ‘for’ loop for adding the metadata is the suggested solution and should be pretty quick. The majority of the heavy lifting happens with the embedding generation and the construct of load_document should be able to optimize for it.

1 Like
3

Hi @Prakul_Agarwal Thanks for confirming the ‘for’ loop’s efficiency with the LangChain team. I appreciate your effort and will keep the loop as it doesn’t impact performance. Thanks a lot!

1 Like