Upload 4 files

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+Regimento.pdf filter=lfs diff=lfs merge=lfs -text

Dockerfile

@@ -0,0 +1,32 @@
+FROM python:3.10-slim
+
+
+WORKDIR /app
+
+
+# Evita issues de encoding
+ENV PYTHONIOENCODING=utf-8
+
+
+# Instala dependências do sistema
+RUN apt-get update && apt-get install -y \
+build-essential \
+&& rm -rf /var/lib/apt/lists/*
+
+
+# Copia requirements
+COPY requirements.txt .
+RUN pip install --no-cache-dir -r requirements.txt
+
+
+# Copia app e PDF
+COPY app.py .
+COPY Regimento.pdf ./Regimento.pdf
+
+
+# Porta padrão do Gradio
+EXPOSE 7860
+
+
+# Comando para iniciar
+CMD ["python", "app.py"]

Regimento.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b987432f9f88e8de1384c7f27c94f75b02af946bede2d39c0779135f90ea7c8a
+size 1639996

app.py

@@ -0,0 +1,218 @@
+# -*- coding: utf-8 -*-
+"""app.ipynb
+
+Automatically generated by Colab.
+
+Original file is located at
+    https://colab.research.google.com/drive/18vjUd8TiNpmeTVPmcgYEYrMyQlozkkzT
+"""
+
+import os
+import torch
+import gradio as gr
+
+from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
+from langchain.prompts import PromptTemplate
+from langchain_community.embeddings import HuggingFaceEmbeddings
+from langchain_community.vectorstores import Chroma
+from pypdf import PdfReader
+from langchain_text_splitters import RecursiveCharacterTextSplitter
+from langchain_core.documents import Document
+
+
+CAMINHO_DB = "db"
+
+
+# =====================================================
+#                 PREPARAÇÃO DO PDF
+# =====================================================
+
+def carregar_pdf(caminho):
+    reader = PdfReader(caminho)
+    textos = []
+    for i, pagina in enumerate(reader.pages):
+        texto = pagina.extract_text()
+        if texto:
+            textos.append(
+                Document(page_content=texto, metadata={"page": i + 1})
+            )
+    return textos
+
+
+def dividir_em_chunks(documentos):
+    splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=200)
+    return splitter.split_documents(documentos)
+
+
+def vetorizar_chunks(chunks):
+    embeddings = HuggingFaceEmbeddings(
+        model_name="intfloat/multilingual-e5-small",
+        model_kwargs={"device": "cpu"},
+        encode_kwargs={"batch_size": 32},
+    )
+
+    db = Chroma(
+        embedding_function=embeddings,
+        persist_directory=CAMINHO_DB,
+    )
+
+    db.add_documents(chunks)
+    db.persist()
+    return db
+
+
+def criar_db(caminho_pdf="Regimento.pdf"):
+    documentos = carregar_pdf(caminho_pdf)
+    chunks = dividir_em_chunks(documentos)
+    return vetorizar_chunks(chunks)
+
+
+# =====================================================
+#     CRIAR DB AUTOMATICAMENTE SE NÃO EXISTIR
+# =====================================================
+
+if not os.path.exists(CAMINHO_DB):
+    print("🔧 Criando base vetorial...")
+    criar_db()
+else:
+    print("📚 Base vetorial encontrada. Carregando...")
+
+
+# =====================================================
+#                   PROMPT TEMPLATE
+# =====================================================
+
+prompt_template = """
+Primeiramente, inicie a resposta sempre com "Oi, querido!"
+
+E depois responda a pergunta:
+{pergunta}
+
+Com base nessas informações:
+{base_conhecimento}
+
+Se não houver resposta na base, diga apenas: não sei te dizer isso.
+"""
+
+_prompt = PromptTemplate(
+    template=prompt_template,
+    input_variables=["pergunta", "base_conhecimento"],
+)
+
+
+# =====================================================
+#                   EMBEDDINGS + CHROMA
+# =====================================================
+
+emb = HuggingFaceEmbeddings(model_name="intfloat/multilingual-e5-small")
+db = Chroma(persist_directory=CAMINHO_DB, embedding_function=emb)
+
+
+# =====================================================
+#                  CARREGAR MODELO LLM
+# =====================================================
+
+def carregar_modelo():
+    MODEL = "Qwen/Qwen2.5-1.5B-Instruct"
+
+    tok = AutoTokenizer.from_pretrained(MODEL)
+
+    mdl = AutoModelForCausalLM.from_pretrained(
+        MODEL,
+        device_map="auto",
+        torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
+    )
+
+    gen = pipeline(
+        "text-generation",
+        model=mdl,
+        tokenizer=tok,
+        max_new_tokens=1000,
+        temperature=0.2,
+        do_sample=False,
+        pad_token_id=tok.eos_token_id,
+        return_full_text=False,
+    )
+
+    return tok, gen
+
+
+tok, generator = carregar_modelo()
+
+
+# =====================================================
+#                    FUNÇÕES DO RAG
+# =====================================================
+
+def listar_fontes(resultados):
+    pags = []
+    for doc, score in resultados:
+        p = doc.metadata.get("page")
+        if p and p not in pags:
+            pags.append(p)
+    return ", ".join([f"p.{p}" for p in pags])
+
+
+def rag_chat(user_msg, history):
+    resultados = db.similarity_search_with_relevance_scores(user_msg, k=3)
+
+    if not resultados or resultados[0][1] < 0.7:
+        resp = "Oi, querido! Não consegui encontrar algo relevante na base para responder com segurança."
+        return history + [(user_msg, resp)]
+
+    textos = [
+        f"(p.{doc.metadata.get('page')}) {doc.page_content}"
+        for doc, score in resultados
+    ]
+
+    base_conhecimento = "\n\n----\n\n".join(textos)
+
+    mensagem = _prompt.format(
+        pergunta=user_msg,
+        base_conhecimento=base_conhecimento,
+    )
+
+    messages = [
+        {"role": "system", "content": "Você responde em PT-BR de forma objetiva e educada."},
+        {"role": "user", "content": mensagem},
+    ]
+
+    prompt_chat = tok.apply_chat_template(
+        messages, tokenize=False, add_generation_prompt=True
+    )
+
+    out = generator(prompt_chat, return_full_text=False)[0]["generated_text"].strip()
+
+    fontes = listar_fontes(resultados)
+    if fontes:
+        out += f"\n\nFontes: {fontes}"
+
+    return history + [(user_msg, out)]
+
+
+# =====================================================
+#                   INTERFACE GRADIO
+# =====================================================
+
+with gr.Blocks(title="CHAT IEPG") as demo:
+
+    gr.Markdown("<h1 style='text-align:center;'>CHAT IEPG</h1>")
+    gr.Markdown(
+        "<p style='text-align:center;'>Faça perguntas sobre o Regimento. "
+        "O chatbot usa RAG (Chroma + E5) com o modelo Qwen.</p>"
+    )
+
+    chat = gr.Chatbot(height=450)
+    txt = gr.Textbox(label="Pergunta", placeholder="Digite sua pergunta...")
+    btn = gr.Button("Enviar")
+    clear = gr.Button("Limpar")
+
+    def responder(user_msg, history):
+        return rag_chat(user_msg, history), ""
+
+    txt.submit(responder, [txt, chat], [chat, txt])
+    btn.click(responder, [txt, chat], [chat, txt])
+    clear.click(lambda: ([], ""), None, [chat, txt])
+
+if __name__ == "__main__":
+    demo.launch(server_name="0.0.0.0", server_port=int(os.getenv("PORT", 7860)))

requirements.txt

@@ -0,0 +1,14 @@
+pypdf
+sentence-transformers>=3.0.1
+langchain-community>=0.2.0
+chromadb
+langchain-text-splitters
+langchain-core
+transformers
+accelerate
+huggingface_hub
+torch
+gradio
+langchain
+streamlit
+streamlit-chat