Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -10,20 +10,20 @@ from langchain.memory import ConversationBufferMemory
|
|
| 10 |
from langchain.agents import Tool, initialize_agent
|
| 11 |
from langchain.agents.agent_types import AgentType
|
| 12 |
|
| 13 |
-
#
|
| 14 |
pipe = pipeline(
|
| 15 |
"text2text-generation",
|
| 16 |
-
model="
|
| 17 |
max_new_tokens=100,
|
| 18 |
truncation=True,
|
| 19 |
max_length=512,
|
| 20 |
)
|
| 21 |
llm = HuggingFacePipeline(pipeline=pipe)
|
| 22 |
|
| 23 |
-
# Mem贸ria limitada para
|
| 24 |
memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True, k=3)
|
| 25 |
|
| 26 |
-
#
|
| 27 |
def carregar_documento_txt(path):
|
| 28 |
loader = TextLoader(path, encoding="latin-1")
|
| 29 |
documentos = loader.load()
|
|
@@ -35,14 +35,14 @@ def carregar_documento_txt(path):
|
|
| 35 |
|
| 36 |
vectorstore = carregar_documento_txt("meu_documento.txt")
|
| 37 |
|
| 38 |
-
# Cadeia de
|
| 39 |
chat_chain = ConversationalRetrievalChain.from_llm(
|
| 40 |
llm=llm,
|
| 41 |
retriever=vectorstore.as_retriever(),
|
| 42 |
memory=memory,
|
| 43 |
)
|
| 44 |
|
| 45 |
-
# Ferramentas
|
| 46 |
def responder_documento(pergunta):
|
| 47 |
return chat_chain.run({"question": pergunta})
|
| 48 |
|
|
@@ -53,15 +53,16 @@ tools = [
|
|
| 53 |
Tool(
|
| 54 |
name="Responder com Documento",
|
| 55 |
func=responder_documento,
|
| 56 |
-
description="
|
| 57 |
),
|
| 58 |
Tool(
|
| 59 |
name="Responder Geral",
|
| 60 |
func=responder_geral,
|
| 61 |
-
description="
|
| 62 |
)
|
| 63 |
]
|
| 64 |
|
|
|
|
| 65 |
agent = initialize_agent(
|
| 66 |
tools=tools,
|
| 67 |
llm=llm,
|
|
@@ -81,9 +82,9 @@ with gr.Blocks() as demo:
|
|
| 81 |
def interagir(msg, history):
|
| 82 |
try:
|
| 83 |
resposta = agent.invoke(msg)
|
| 84 |
-
#
|
| 85 |
if isinstance(resposta, dict):
|
| 86 |
-
resposta = str(resposta)
|
| 87 |
elif not isinstance(resposta, str):
|
| 88 |
resposta = str(resposta)
|
| 89 |
except Exception as e:
|
|
@@ -92,10 +93,8 @@ with gr.Blocks() as demo:
|
|
| 92 |
return "", history
|
| 93 |
|
| 94 |
entrada.submit(interagir, [entrada, estado_chat], [entrada, chatbot])
|
| 95 |
-
|
| 96 |
-
enviar_btn.click(interagir, [entrada, estado_chat], [entrada, chatbot])
|
| 97 |
|
| 98 |
demo.launch(server_name="0.0.0.0", server_port=7860, show_api=False)
|
| 99 |
|
| 100 |
|
| 101 |
-
|
|
|
|
| 10 |
from langchain.agents import Tool, initialize_agent
|
| 11 |
from langchain.agents.agent_types import AgentType
|
| 12 |
|
| 13 |
+
# Modelo com melhor performance para agentes
|
| 14 |
pipe = pipeline(
|
| 15 |
"text2text-generation",
|
| 16 |
+
model="declare-lab/flan-alpaca-base",
|
| 17 |
max_new_tokens=100,
|
| 18 |
truncation=True,
|
| 19 |
max_length=512,
|
| 20 |
)
|
| 21 |
llm = HuggingFacePipeline(pipeline=pipe)
|
| 22 |
|
| 23 |
+
# Mem贸ria limitada para evitar prompts muito longos
|
| 24 |
memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True, k=3)
|
| 25 |
|
| 26 |
+
# Carrega e divide documento em partes menores
|
| 27 |
def carregar_documento_txt(path):
|
| 28 |
loader = TextLoader(path, encoding="latin-1")
|
| 29 |
documentos = loader.load()
|
|
|
|
| 35 |
|
| 36 |
vectorstore = carregar_documento_txt("meu_documento.txt")
|
| 37 |
|
| 38 |
+
# Cadeia de busca com mem贸ria
|
| 39 |
chat_chain = ConversationalRetrievalChain.from_llm(
|
| 40 |
llm=llm,
|
| 41 |
retriever=vectorstore.as_retriever(),
|
| 42 |
memory=memory,
|
| 43 |
)
|
| 44 |
|
| 45 |
+
# Ferramentas do agente
|
| 46 |
def responder_documento(pergunta):
|
| 47 |
return chat_chain.run({"question": pergunta})
|
| 48 |
|
|
|
|
| 53 |
Tool(
|
| 54 |
name="Responder com Documento",
|
| 55 |
func=responder_documento,
|
| 56 |
+
description="Use para responder perguntas espec铆ficas sobre o documento."
|
| 57 |
),
|
| 58 |
Tool(
|
| 59 |
name="Responder Geral",
|
| 60 |
func=responder_geral,
|
| 61 |
+
description="Use para perguntas gerais ou bate-papo."
|
| 62 |
)
|
| 63 |
]
|
| 64 |
|
| 65 |
+
# Agente com tratamento de parsing
|
| 66 |
agent = initialize_agent(
|
| 67 |
tools=tools,
|
| 68 |
llm=llm,
|
|
|
|
| 82 |
def interagir(msg, history):
|
| 83 |
try:
|
| 84 |
resposta = agent.invoke(msg)
|
| 85 |
+
# Extrai texto se for dict
|
| 86 |
if isinstance(resposta, dict):
|
| 87 |
+
resposta = resposta.get('output', str(resposta))
|
| 88 |
elif not isinstance(resposta, str):
|
| 89 |
resposta = str(resposta)
|
| 90 |
except Exception as e:
|
|
|
|
| 93 |
return "", history
|
| 94 |
|
| 95 |
entrada.submit(interagir, [entrada, estado_chat], [entrada, chatbot])
|
| 96 |
+
gr.Button("Enviar").click(interagir, [entrada, estado_chat], [entrada, chatbot])
|
|
|
|
| 97 |
|
| 98 |
demo.launch(server_name="0.0.0.0", server_port=7860, show_api=False)
|
| 99 |
|
| 100 |
|
|
|