Custom models

All custom models are run with ChunksTabularExtractor, if you want custom extractor you need to initialize it with model of your choice.

Note that small local models tend to trim long outputs and could require more careful tuning of data description.

Run custom model

You can instantiate Parsera with any chat model supported by LangChain, for example, to run gpt-4o-mini from OpenAI API:

import os
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="gpt-4o-mini",
    temperature=0.0,
    timeout=120,
)

url = "https://github.com/raznem/parsera"
elements = {
    "Stars": "Number of stars",
    "Fork": "Number of forks",
}
scrapper = Parsera(model=llm)
result = scrapper.run(url=url, elements=elements)

Run local model with Ollama

First, you should install and run ollama in your local environment: official installation guide. Additionally, you need to install langchain_ollama with:

pip install langchain-ollama

The next step is pulling the model. For example, to pull Qwen2.5 14B run:

ollama pull qwen2.5:14b

After all the setup simply run:

from parsera import Parsera
from langchain_ollama import ChatOllama

url = "https://github.com/raznem/parsera"
elements = {
    "Stars": "Number of stars",
    "Fork": "Number of forks",
}

llm = ChatOllama(
    model="qwen2.5:14b",
    temperature=0,
)

scrapper = Parsera(model=llm)
result = await scrapper.arun(url=url, elements=elements)

Run local model with Trasformers

Currently, we only support models that include a system token

You should install Transformers with either pytorch (recommended) or TensorFlow 2.0

Transformers Installation Guide

Example:

from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
from parsera.engine.model import HuggingFaceModel
from parsera import Parsera

# Define the URL and elements to scrape
url = "https://github.com/raznem/parsera"
elements = {
    "Stars": "Number of stars",
    "Fork": "Number of forks",
}

# Initialize model with transformers pipeline
tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-128k-instruct", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("microsoft/Phi-3-mini-128k-instruct", trust_remote_code=True)
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, max_new_tokens=5000)

# Initialize HuggingFaceModel
llm = HuggingFaceModel(pipeline=pipe)

# Scrapper with HuggingFace model
scrapper = Parsera(model=llm)
result = scrapper.run(url=url, elements=elements)