Spaces:
Build error
Build error
olegperegudov
commited on
Commit
•
11f324c
1
Parent(s):
13b81ea
wip
Browse files- .gitignore +3 -0
- app.py +43 -2
- build_model.py +23 -0
- constants.py +45 -0
- utils.py +100 -0
.gitignore
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
data
|
2 |
+
env
|
3 |
+
model
|
app.py
CHANGED
@@ -1,4 +1,45 @@
|
|
1 |
import streamlit as st
|
2 |
|
3 |
-
|
4 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import streamlit as st
|
2 |
|
3 |
+
import utils
|
4 |
+
from build_model import load_model
|
5 |
+
|
6 |
+
st.title("Buzzbot")
|
7 |
+
|
8 |
+
# Initialize retriever and model
|
9 |
+
if "retriever" not in st.session_state:
|
10 |
+
st.session_state["retriever"] = utils.build_retriever()
|
11 |
+
if "model" not in st.session_state:
|
12 |
+
st.session_state["model"] = load_model()
|
13 |
+
if "conversation" not in st.session_state:
|
14 |
+
st.session_state["conversation"] = utils.Conversation()
|
15 |
+
|
16 |
+
# Initialize chat history
|
17 |
+
if "messages" not in st.session_state:
|
18 |
+
st.session_state.messages = []
|
19 |
+
|
20 |
+
# Display chat messages from history on app rerun
|
21 |
+
for message in st.session_state.messages:
|
22 |
+
with st.chat_message(message["role"]):
|
23 |
+
st.markdown(message["content"])
|
24 |
+
if message['role']=="assistant":
|
25 |
+
st.caption(message["source_docs"])
|
26 |
+
|
27 |
+
# Accept user input
|
28 |
+
if user_input := st.chat_input("What is up?"):
|
29 |
+
# Add user message to chat history
|
30 |
+
st.session_state.messages.append({"role": "user", "content": user_input, "source_docs": None})
|
31 |
+
# Display user message in chat message container
|
32 |
+
with st.chat_message("user"):
|
33 |
+
st.markdown(user_input)
|
34 |
+
|
35 |
+
# Display assistant response in chat message container
|
36 |
+
with st.chat_message("assistant"):
|
37 |
+
with st.spinner(""):
|
38 |
+
answer, source_docs = utils.ask_question(
|
39 |
+
user_input, st.session_state.conversation, st.session_state.model, st.session_state.retriever
|
40 |
+
)
|
41 |
+
st.write(answer)
|
42 |
+
# for source_doc in source_docs:
|
43 |
+
st.caption(source_docs)
|
44 |
+
|
45 |
+
st.session_state.messages.append({"role": "assistant", "content": answer, "source_docs": source_docs})
|
build_model.py
ADDED
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from langchain.callbacks.manager import CallbackManager
|
2 |
+
from langchain.callbacks.streaming_stdout import StreamingStdOutCallbackHandler
|
3 |
+
from langchain_community.llms import LlamaCpp
|
4 |
+
|
5 |
+
import constants
|
6 |
+
|
7 |
+
callback_manager = CallbackManager([StreamingStdOutCallbackHandler()])
|
8 |
+
|
9 |
+
|
10 |
+
def load_model():
|
11 |
+
return LlamaCpp(
|
12 |
+
model_path=constants.MODEL_SAVE_PATH,
|
13 |
+
temperature=constants.TEMPERATURE,
|
14 |
+
max_tokens=constants.MAX_TOKENS,
|
15 |
+
top_p=constants.TOP_P,
|
16 |
+
# callback_manager=callback_manager, # will stream to stdout, but wont attach to variable
|
17 |
+
verbose=False, # Verbose is required to pass to the callback manager
|
18 |
+
n_gpu_layers=constants.N_GPU_LAYERS,
|
19 |
+
n_batch=constants.N_BATCH,
|
20 |
+
n_ctx=constants.N_CTX,
|
21 |
+
repeat_penalty=constants.REPEAT_PENALTY,
|
22 |
+
streaming=False,
|
23 |
+
)
|
constants.py
ADDED
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
|
3 |
+
import torch
|
4 |
+
|
5 |
+
# model path
|
6 |
+
MODEL_NAME = "saiga_mistral_7b.Q4_K_M.gguf"
|
7 |
+
MODEL_URL = f"https://huggingface.co/TheBloke/saiga_mistral_7b-GGUF/blob/main/{MODEL_NAME}"
|
8 |
+
|
9 |
+
# FOR PRODUCTION
|
10 |
+
CWD = os.path.dirname(os.path.realpath(__file__))
|
11 |
+
DATA_PATH = os.path.join(CWD, "data")
|
12 |
+
DOCS_PATH = os.path.join(DATA_PATH, "docs")
|
13 |
+
MODEL_PATH = os.path.join(CWD, "model")
|
14 |
+
MODEL_SAVE_PATH = os.path.join(MODEL_PATH, MODEL_NAME)
|
15 |
+
|
16 |
+
# RAG params
|
17 |
+
N_GPU_LAYERS = (
|
18 |
+
-1 if torch.cuda.is_available() else 0
|
19 |
+
) # The number of layers to put on the GPU. The rest will be on the CPU (0 means all layers on the CPU).
|
20 |
+
N_BATCH = 1024 # Should be between 1 and n_ctx, consider the amount of VRAM in your GPU
|
21 |
+
|
22 |
+
TEMPERATURE = 0.1 # The temperature of the sampling. 0.1 is a good value for most cases
|
23 |
+
MAX_TOKENS = 1024 # The maximum number of tokens to generate
|
24 |
+
TOP_P = 2
|
25 |
+
N_CTX = 2048 # context len, up to a maximum of 32k
|
26 |
+
CHUNK_SIZE = 750 # max number of letters for each chunk during splitting
|
27 |
+
CHUNK_OVERLAP = 200 # overlap between chunks
|
28 |
+
SEARCH_TYPE = "mmr"
|
29 |
+
LAST_MESSAGES = 3 # The number of last messages in conversation history to include in the context
|
30 |
+
REPEAT_PENALTY = 1.1 # The penalty for repeating tokens in the output
|
31 |
+
DEVICE = "cuda" if N_GPU_LAYERS > 0 else "cpu"
|
32 |
+
|
33 |
+
EMBED_MODEL_NAME = "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"
|
34 |
+
VECTOR_STORE_PATH = os.path.join(DATA_PATH, "chroma_db")
|
35 |
+
|
36 |
+
# retriever config
|
37 |
+
SEARCH_KWARGS = {"k": 3, "score_threshold": 0.6}
|
38 |
+
|
39 |
+
DEFAULT_MESSAGE_TEMPLATE = "<s>{role}\n{content}</s>"
|
40 |
+
DEFAULT_RESPONSE_TEMPLATE = "<s>bot\n"
|
41 |
+
DEFAULT_SYSTEM_PROMPT = "Ты ассистент помощник, который отвечает на вопросы используя предоставленный контекст. \
|
42 |
+
В качестве контекста используются тексты из различных источников. \
|
43 |
+
Постарайся ответить на вопрос максимально точно. \
|
44 |
+
Для ответа используй только информацию из контекста и вопроса. Ничего не выдумывай. \
|
45 |
+
Если не можешь ответить на вопрос, напиши - 'Не хватает данных для ответа.' "
|
utils.py
ADDED
@@ -0,0 +1,100 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
|
3 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
4 |
+
from langchain_community.document_loaders import DirectoryLoader
|
5 |
+
from langchain_community.embeddings import HuggingFaceEmbeddings
|
6 |
+
from langchain_community.vectorstores import Chroma
|
7 |
+
|
8 |
+
import constants
|
9 |
+
|
10 |
+
|
11 |
+
class Conversation:
|
12 |
+
def __init__(
|
13 |
+
self,
|
14 |
+
message_template=constants.DEFAULT_MESSAGE_TEMPLATE,
|
15 |
+
system_prompt=constants.DEFAULT_SYSTEM_PROMPT,
|
16 |
+
response_template=constants.DEFAULT_RESPONSE_TEMPLATE,
|
17 |
+
):
|
18 |
+
self.message_template = message_template
|
19 |
+
self.response_template = response_template
|
20 |
+
self.messages = [{"role": "system", "content": system_prompt}]
|
21 |
+
|
22 |
+
def add_user_message(self, message):
|
23 |
+
self.messages.append({"role": "user", "content": message})
|
24 |
+
|
25 |
+
def add_bot_message(self, message):
|
26 |
+
self.messages.append({"role": "bot", "content": message})
|
27 |
+
|
28 |
+
def get_conversation_history(self):
|
29 |
+
final_text = ""
|
30 |
+
# 1st system message + last few messages (excluding system duplicate)
|
31 |
+
context_and_last_few_messages = [self.messages[0]] + self.messages[1:][-constants.LAST_MESSAGES :]
|
32 |
+
for message in context_and_last_few_messages:
|
33 |
+
message_text = self.message_template.format(**message)
|
34 |
+
final_text += message_text
|
35 |
+
return final_text.strip()
|
36 |
+
|
37 |
+
|
38 |
+
def source_documents(relevant_docs):
|
39 |
+
source_docs = set()
|
40 |
+
for doc in relevant_docs:
|
41 |
+
fname = doc.metadata["source"]
|
42 |
+
fname_base = os.path.splitext(os.path.basename(fname))[0]
|
43 |
+
source_docs.add(fname_base)
|
44 |
+
return list(source_docs)
|
45 |
+
|
46 |
+
|
47 |
+
def load_raw_documents():
|
48 |
+
return DirectoryLoader(constants.DOCS_PATH, glob="*.txt").load()
|
49 |
+
|
50 |
+
|
51 |
+
def build_nodes(raw_documents):
|
52 |
+
return RecursiveCharacterTextSplitter(
|
53 |
+
chunk_size=constants.CHUNK_SIZE,
|
54 |
+
chunk_overlap=constants.CHUNK_OVERLAP,
|
55 |
+
length_function=len,
|
56 |
+
is_separator_regex=False,
|
57 |
+
).split_documents(raw_documents)
|
58 |
+
|
59 |
+
|
60 |
+
def build_embeddings():
|
61 |
+
return HuggingFaceEmbeddings(model_name=constants.EMBED_MODEL_NAME, model_kwargs={"device": constants.DEVICE})
|
62 |
+
|
63 |
+
|
64 |
+
def build_db(nodes, embeddings):
|
65 |
+
return Chroma.from_documents(nodes, embeddings)
|
66 |
+
|
67 |
+
|
68 |
+
def build_retriever():
|
69 |
+
raw_documents = load_raw_documents()
|
70 |
+
nodes = build_nodes(raw_documents)
|
71 |
+
embeddings = build_embeddings()
|
72 |
+
db = build_db(nodes, embeddings)
|
73 |
+
return db.as_retriever(search_kwargs=constants.SEARCH_KWARGS, search_type=constants.SEARCH_TYPE)
|
74 |
+
|
75 |
+
|
76 |
+
def fetch_relevant_nodes(question, retriever):
|
77 |
+
relevant_docs = retriever.get_relevant_documents(question)
|
78 |
+
context = [doc.page_content for doc in relevant_docs]
|
79 |
+
source_docs = source_documents(relevant_docs)
|
80 |
+
context = list(set(context)) # remove duplicated strings from context
|
81 |
+
return context, source_docs
|
82 |
+
|
83 |
+
|
84 |
+
def ask_question(question, conversation, model, retriever):
|
85 |
+
|
86 |
+
context, source_docs = fetch_relevant_nodes(question, retriever)
|
87 |
+
|
88 |
+
# add user message to conversation's context
|
89 |
+
conversation.add_user_message(question)
|
90 |
+
conversation_history = conversation.get_conversation_history()
|
91 |
+
prompt = f"{conversation_history}\n\
|
92 |
+
{context}\n\
|
93 |
+
{constants.DEFAULT_RESPONSE_TEMPLATE}"
|
94 |
+
|
95 |
+
|
96 |
+
answer = model.invoke(prompt)
|
97 |
+
# add bot message to conversation's context
|
98 |
+
conversation.add_bot_message(answer)
|
99 |
+
|
100 |
+
return answer, source_docs
|