llamacpp-summarize

main.py

from langchain_text_splitters import RecursiveCharacterTextSplitter
from llama_cpp import Llama, CreateChatCompletionResponse
from typing import cast
from pymupdf import pymupdf

print('[INFO] Loading text...')

doc = pymupdf.open('./report.pdf')
content = ''
for page in doc:
    content += page.get_text()

print('[INFO] Splitting text...')

splitter = RecursiveCharacterTextSplitter(chunk_size=2000, chunk_overlap=200)
splits = splitter.split_text(content)

llm = Llama(
    model_path="./models/Qwen3.5-4B-Q4_K_M.gguf",
    verbose=False,
    n_ctx=4098,
    n_threads=8,
    n_batch=512,
    n_ubatch=512,
    use_mmap=True,
    use_mlock=False,
    flash_attn=True,
    n_gpu_layers=0,
    seed=42,
    temperature=0
)

print('[INFO] Starting AI summarization...')

first_summary = ''
for split in splits:
    response = cast(
        CreateChatCompletionResponse,
        llm.create_chat_completion(
            messages=[
                {
                    "role": "system",
                    "content": """
                        você deve resumir este conteúdo.
                        mantenha o resumo conciso e traga \"o quê\" ao invés de \"como\".
                    """
                },
                {
                    "role": "user",
                    "content": f"""
                        <CONTEXTO>
                        {split}
                        </CONTEXTO>
                    """
                }
            ],
            temperature=0.1,
            stream=False
        )
    )
    result = response['choices'][0]['message']['content']
    print("=============================================================")
    print(result)
    first_summary += f"\n\n{result}"

print("======================== RESUMO 1 ==========================")
print(first_summary)


response = cast(
    CreateChatCompletionResponse,
    llm.create_chat_completion(
        messages=[
            {
                "role": "system",
                "content": """
                    você deve resumir este conteúdo.
                    mantenha o resumo conciso e traga \"o quê\" ao invés de \"como\".
                """
            },
            {
                "role": "user",
                "content": f"""
                    <CONTEXTO>
                    {first_summary}
                    </CONTEXTO>
                """
            }
        ],
        temperature=0.1,
        stream=False
    )
)

result = response['choices'][0]['message']['content']
print("======================== RESUMO 2 ==========================")
print(result)