Making a Natural Language based LLM Recommendation System

Making a use case of Anime recommender system using llm and Langchain/LlamaIndex. Using graph database and normal Json query

GitHub

Github project

Overall Process for Development

First step was to figure out what to do first so, I used 3 tier architecture to figure out how my recomender system will work.

So the plan is to work in following order:

• Data scrapping

• Database creation

• Backend setup

  • Rag stack

  • model selection

  • performance comparison

  • User auth

• UI (Still figuring out)

Data Scrapping

So to get data on anime, I used the following api Jikan API It has details regarding anime. Now we have to get the data of most anime that is available.

So let write a srcipt:

import requests

import re

import time

import tqdm

import json

base_url = "https://api.jikan.moe/v4/anime"

def replace_special_characters(input_string):

    # Replace spaces and special characters with underscores

    result = re.sub(r'[^a-zA-Z0-9]', '_', input_string)

    return result

for i in tqdm.tqdm(range(1,26302)):

    response = requests.request("GET", f"{base_url}/{i}/full", headers={}, data={})

    if(response.status_code == 200):

        res_json = response.json()

        mal_id = res_json["data"]["mal_id"]

        anime_name =replace_special_characters(res_json["data"]["titles"][0]["title"])

        json.dump(res_json,open(f"data/json_data/{mal_id}_{anime_name}.json","w"))

        time.sleep(1)

This will get us around ~26k anime data.

Database

Now that we have ourselves the data let’s start on making a database and a graph database.

First let’s Have a look at our data for a single anime.(I choose one of my fav😉).

So for normal Json based index we are done.

But for Graph database creation i made following data model for neo4j based on above data:

I’ll be using neomodel python package to create and interact with Graph database (Neo4j cloud database).

Here’s the data model:

from neomodel import (config, StructuredNode, StringProperty, IntegerProperty,

     ArrayProperty ,UniqueIdProperty, RelationshipTo, FloatProperty, DateProperty)

from neomodel.contrib import  SemiStructuredNode

class Date(StructuredNode):

    Date = DateProperty()

class Synonyms(StructuredNode):#1

    Synonyms_array = ArrayProperty(StringProperty())

class Genre(StructuredNode):#2

    genre = StringProperty()

    mal_id = IntegerProperty(unique_index=True)

class Producer(StructuredNode):#3

    name = StringProperty()

    url = StringProperty()

    type = StringProperty()

    mal_id = IntegerProperty(unique_index=True)

class Themes(StructuredNode):

    name = StringProperty()

    url = StringProperty()

    type = StringProperty()

    mal_id = IntegerProperty(unique_index=True)

class Demographics(StructuredNode):

    name = StringProperty()

    url = StringProperty()

    type = StringProperty()

    mal_id = IntegerProperty(unique_index=True)

class Manga(StructuredNode):

    name = StringProperty()

    url = StringProperty()

    type = StringProperty()

    mal_id = IntegerProperty(unique_index=True)

class Studios(StructuredNode):

    name = StringProperty()

    url = StringProperty()

    type = StringProperty()

    mal_id = IntegerProperty(unique_index=True)

#%%

class Anime(SemiStructuredNode):

    mal_id = IntegerProperty(unique_index=True)#

    url = StringProperty()#

    title = StringProperty()#

    type = StringProperty()#

    source = StringProperty()#

    episodes = IntegerProperty()#

    status = StringProperty()#

    airing = StringProperty()#

    duration = StringProperty()#

    rating = StringProperty()#

    score = FloatProperty()#

    scored_by = IntegerProperty()#

    rank = IntegerProperty()#

    popularity = IntegerProperty()#

    members = IntegerProperty()#

    favorites = IntegerProperty()#

    synopsis = StringProperty()#

    background = StringProperty()#

    season = StringProperty()#

    year = DateProperty()

    # Relationships

    aired_from = RelationshipTo(Date, 'aired_from')

    aired_till = RelationshipTo(Date, 'aired_till')

    broadcasted = RelationshipTo(Date, 'Broadcasted_on')

    genres = RelationshipTo(Genre, 'is_of_genre')

    producers = RelationshipTo(Producer, 'produced')

    studios = RelationshipTo(Studios, 'made')

    themes = RelationshipTo(Themes, 'is_of_theme')

    demographics = RelationshipTo(Demographics, 'is_of_demographic')

    manga = RelationshipTo(Manga, 'is adapted from')

    synonyms= RelationshipTo(Synonyms,'is aslo called')

Graph DB visualization

Recommender System

Let’s first try out the json based inder and recomendation ai.

JSON Based Indexer

• We have around 26k json data. And will use this as our knowledge base.

• Let’s test out Json_querry_engine from llama Index

Graph Based Indexer

• Will update

Comparisions

• Will update

---