Let’s take a look at the v2 of the NuExtract model we previously visited here. It’s possible alternative to extract knowledge graphs from text to feed into knwl or other knowledge graph building tools.
The UV pyproject.toml file should look like this:
[project]name="nuex"version="0.1.0"description="Add your description here"readme="README.md"requires-python=">=3.12"dependencies=["accelerate>=1.11.0","haystack-ai>=2.19.0","jupyter>=1.1.1","networkx>=3.5","pillow>=12.0.0","pyvis>=0.3.2","qwen-vl-utils>=0.0.14","torch>=2.9.0","torchvision>=0.24.0","trafilatura>=2.0.0","transformers>=4.57.1",]
There are three models (2B,4B, 8B) available on Hugging Face and from a little experimentation the 2B model seems to be giving the best results. What makes this model so interesting is the speed at which it can extract structured data from text. If you follow the quick-start of Knwl you will see that it takes a lot longer to extract entities and relationships from text compared to NuExtract v2. NuExtract does the job in seconds, compared to minutes when using big models like GPT-5 or Claude 4.
huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...
To disable this warning, you can either:
- Avoid using `tokenizers` before the fork if possible
- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)
We take a little paragraph from the Wikipedia article on Charles Dickens and provide the LLM with some examples of the kind of structured data we want to extract. The Transformers parameters are particular interest here and after some experimentation we found that greedy decoding (do_sample=False, num_beams=1) seems to work best. That is, the greedy approach of picking the most likely next token at each step seems to give better results than sampling from a distribution of possible next tokens.
template ="""{ "output": { "entities": ["string"], "relationships": [["string", "string", "string"]] }}"""document ="""Charles John Huffam Dickens (7 February 1812 – 9 June 1870) was an English novelist, journalist, short story writer and social critic. He created some of literature's best-known fictional characters, and is regarded by many as the greatest novelist of the Victorian era. His works enjoyed unprecedented popularity during his lifetime and, by the 20th century, critics and scholars had recognised him as a literary genius. His novels and short stories are widely read today.Born in Portsmouth, Dickens left school at age 12 to work in a boot-blacking factory when his father John was incarcerated in a debtors' prison. After three years, he returned to school before beginning his literary career as a journalist. Dickens edited a weekly journal for 20 years; wrote 15 novels, five novellas, hundreds of short stories and nonfiction articles; lectured and performed readings extensively; was a tireless letter writer; and campaigned vigorously for children's rights, education and other social reforms.Dickens's literary success began with the 1836 serial publication of The Pickwick Papers, a publishing phenomenon—thanks largely to the introduction of the character Sam Weller in the fourth episode—that sparked Pickwick merchandise and spin-offs. Within a few years, Dickens had become an international literary celebrity, famous for his humour, satire and keen observation of character and society. His novels, most of them published in monthly or weekly instalments, pioneered the serial publication of narrative fiction, which became the dominant Victorian mode for novel publication. Cliffhanger endings in his serial publications kept readers in suspense. The instalment format allowed Dickens to evaluate his audience's reaction, and he often modified his plot and character development based on such feedback. For example, when his wife's chiropodist expressed distress at the way Miss Mowcher in David Copperfield seemed to reflect her own disabilities, Dickens improved the character with positive features. His plots were carefully constructed and he often wove elements from topical events into his narratives. Masses of the illiterate poor would individually pay a halfpenny to have each new monthly episode read to them, opening up and inspiring a new class of readers.His 1843 novella A Christmas Carol remains especially popular and continues to inspire adaptations in every creative medium. Oliver Twist and Great Expectations are also frequently adapted and, like many of his novels, evoke images of early Victorian London. His 1853 novel Bleak House, a satire on the judicial system, helped support a reformist movement that culminated in the 1870s legal reform in England. A Tale of Two Cities (1859; set in London and Paris) is regarded as his best-known work of historical fiction. The most famous celebrity of his era, he undertook, in response to public demand, a series of public reading tours in the later part of his career. The term Dickensian is used to describe something that is reminiscent of Dickens and his writings, such as poor social or working conditions, or comically repulsive characters."""examples = [ {"input": "Stephen is the manager of Anna. Stephen works in Belgium.","output": """{ "entities": ["Stephen", "Anna", "Belgium"], "relationships": [["Stephen", "manager_of", "Anna"], ["Stephen", "works_in", "Belgium"]] }""" }, {"input": "Google was founded by Larry Page and Sergey Brin while they were Ph.D. students at Stanford University.","output": """{ "entities": ["Google", "Larry Page", "Sergey Brin", "Stanford University"], "relationships": [["Larry Page", "co_founder_of", "Google"], ["Sergey Brin", "co_founder_of", "Google"], ["Larry Page", "student_at", "Stanford University"], ["Sergey Brin", "student_at", "Stanford University"]] }""" }]messages = [{"role": "user", "content": document}]text = processor.tokenizer.apply_chat_template( messages, template=template, examples=examples, # examples provided here tokenize=False, add_generation_prompt=True,)inputs = processor( text=[text], images=None, padding=True, return_tensors="pt",)inputs = inputs.to(model.device)start_time = time.time()# https://huggingface.co/docs/transformers/main_classes/text_generationgenerated_ids = model.generate(**inputs,**{"do_sample": False, # False means greedy decoding"num_beams": 1, "max_new_tokens": 2*4096, "temperature": None, "top_p": None, "top_k": None,# "max_time": 120,})end_time = time.time()generated_ids_trimmed = [ out_ids[len(in_ids) :] for in_ids, out_ids inzip(inputs.input_ids, generated_ids)]result = processor.batch_decode( generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False)try: g = json.loads(result[0])print(json.dumps(g, indent=4))print(f"Found {len(g['entities']) if'entities'in g else0} entities and {len(g['relationships']) if'relationships'in g else0} relationships in {(end_time - start_time):.1f} seconds.")except json.JSONDecodeError as e:print("Bad format JSON output:", result[0])
{
"entities": [
"Charles John Huffam Dickens",
"Sam Weller",
"Miss Mowcher",
"David Copperfield",
"A Christmas Carol",
"Oliver Twist",
"Great Expectations",
"Bleak House",
"A Tale of Two Cities",
"Victorian era",
"Victorian London",
"legal reform in England"
],
"relationships": [
[
"Charles John Huffam Dickens",
"born_in",
"Portsmouth"
],
[
"Charles John Huffam Dickens",
"created",
"Sam Weller"
],
[
"Charles John Huffam Dickens",
"wrote",
"A Christmas Carol"
],
[
"Charles John Huffam Dickens",
"adapted",
"Oliver Twist"
],
[
"Charles John Huffam Dickens",
"adapted",
"Great Expectations"
],
[
"Charles John Huffam Dickens",
"adapted",
"Bleak House"
],
[
"Charles John Huffam Dickens",
"adapted",
"A Tale of Two Cities"
],
[
"Charles John Huffam Dickens",
"published",
"The Pickwick Papers"
],
[
"Charles John Huffam Dickens",
"lectured",
"public reading tours"
]
]
}
Found 12 entities and 9 relationships in 16.8 seconds.
Doing this in less than 17s on a Mac M4 Pro is quite impressive. A bit of NetworkX and PyVis magic at the end to visualize the extracted knowledge graph.
import networkx as nx# Create a new graphG = nx.Graph()# Add nodes and edgesfor e in g["entities"]: G.add_node(e, label=e)for r in g["relationships"]: from_node = r[0] relation_type = r[1] to_node = r[2]if from_node in G.nodes() and to_node in G.nodes(): G.add_edge(from_node, to_node, type=relation_type)
The v2 is definitely a step up from the previous version and a very interesting model to extract structured data from text very quickly. When the structure becomes more complex however, it seems to struggle a bit more. But for simple entity and relationship extraction tasks, it’s a great tool to have in your toolbox.
