Spaces:
Sleeping
Sleeping
manjunathshiva
commited on
Commit
•
fe0e9f4
1
Parent(s):
18cc60c
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,167 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from llama_index.core import (
|
2 |
+
VectorStoreIndex
|
3 |
+
)
|
4 |
+
from llama_index.core import Settings
|
5 |
+
from llama_index.embeddings.huggingface import HuggingFaceEmbedding
|
6 |
+
from llama_index.vector_stores.qdrant import QdrantVectorStore
|
7 |
+
from qdrant_client import QdrantClient
|
8 |
+
from typing import Any, List, Tuple
|
9 |
+
import torch
|
10 |
+
from transformers import AutoTokenizer, AutoModelForMaskedLM
|
11 |
+
import streamlit as st
|
12 |
+
from llama_index.llms.huggingface import (
|
13 |
+
HuggingFaceInferenceAPI
|
14 |
+
)
|
15 |
+
import os
|
16 |
+
HUGGINGFACEHUB_API_TOKEN = os.environ.get("HUGGINGFACEHUB_API_TOKEN")
|
17 |
+
Q_END_POINT = os.environ.get("Q_END_POINT")
|
18 |
+
Q_API_KEY = os.environ.get("Q_API_KEY")
|
19 |
+
|
20 |
+
|
21 |
+
#DOC
|
22 |
+
#https://docs.llamaindex.ai/en/stable/examples/vector_stores/qdrant_hybrid.html
|
23 |
+
|
24 |
+
doc_tokenizer = AutoTokenizer.from_pretrained(
|
25 |
+
"naver/efficient-splade-VI-BT-large-doc"
|
26 |
+
)
|
27 |
+
doc_model = AutoModelForMaskedLM.from_pretrained(
|
28 |
+
"naver/efficient-splade-VI-BT-large-doc"
|
29 |
+
)
|
30 |
+
|
31 |
+
query_tokenizer = AutoTokenizer.from_pretrained(
|
32 |
+
"naver/efficient-splade-VI-BT-large-query"
|
33 |
+
)
|
34 |
+
query_model = AutoModelForMaskedLM.from_pretrained(
|
35 |
+
"naver/efficient-splade-VI-BT-large-query"
|
36 |
+
)
|
37 |
+
|
38 |
+
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
39 |
+
|
40 |
+
doc_model = doc_model.to(device)
|
41 |
+
query_model = query_model.to(device)
|
42 |
+
|
43 |
+
|
44 |
+
def sparse_doc_vectors(
|
45 |
+
texts: List[str],
|
46 |
+
) -> Tuple[List[List[int]], List[List[float]]]:
|
47 |
+
"""
|
48 |
+
Computes vectors from logits and attention mask using ReLU, log, and max operations.
|
49 |
+
"""
|
50 |
+
tokens = doc_tokenizer(
|
51 |
+
texts, truncation=True, padding=True, return_tensors="pt"
|
52 |
+
)
|
53 |
+
if torch.cuda.is_available():
|
54 |
+
tokens = tokens.to("cuda:1")
|
55 |
+
|
56 |
+
output = doc_model(**tokens)
|
57 |
+
logits, attention_mask = output.logits, tokens.attention_mask
|
58 |
+
relu_log = torch.log(1 + torch.relu(logits))
|
59 |
+
weighted_log = relu_log * attention_mask.unsqueeze(-1)
|
60 |
+
tvecs, _ = torch.max(weighted_log, dim=1)
|
61 |
+
|
62 |
+
# extract the vectors that are non-zero and their indices
|
63 |
+
indices = []
|
64 |
+
vecs = []
|
65 |
+
for batch in tvecs:
|
66 |
+
indices.append(batch.nonzero(as_tuple=True)[0].tolist())
|
67 |
+
vecs.append(batch[indices[-1]].tolist())
|
68 |
+
|
69 |
+
return indices, vecs
|
70 |
+
|
71 |
+
|
72 |
+
def sparse_query_vectors(
|
73 |
+
texts: List[str],
|
74 |
+
) -> Tuple[List[List[int]], List[List[float]]]:
|
75 |
+
"""
|
76 |
+
Computes vectors from logits and attention mask using ReLU, log, and max operations.
|
77 |
+
"""
|
78 |
+
# TODO: compute sparse vectors in batches if max length is exceeded
|
79 |
+
tokens = query_tokenizer(
|
80 |
+
texts, truncation=True, padding=True, return_tensors="pt"
|
81 |
+
)
|
82 |
+
if torch.cuda.is_available():
|
83 |
+
tokens = tokens.to("cuda:1")
|
84 |
+
|
85 |
+
|
86 |
+
output = query_model(**tokens)
|
87 |
+
logits, attention_mask = output.logits, tokens.attention_mask
|
88 |
+
relu_log = torch.log(1 + torch.relu(logits))
|
89 |
+
weighted_log = relu_log * attention_mask.unsqueeze(-1)
|
90 |
+
tvecs, _ = torch.max(weighted_log, dim=1)
|
91 |
+
|
92 |
+
# extract the vectors that are non-zero and their indices
|
93 |
+
indices = []
|
94 |
+
vecs = []
|
95 |
+
for batch in tvecs:
|
96 |
+
indices.append(batch.nonzero(as_tuple=True)[0].tolist())
|
97 |
+
vecs.append(batch[indices[-1]].tolist())
|
98 |
+
|
99 |
+
return indices, vecs
|
100 |
+
|
101 |
+
st.header("Chat with the Bhagavad Gita docs 💬 📚"")
|
102 |
+
|
103 |
+
if "messages" not in st.session_state.keys(): # Initialize the chat message history
|
104 |
+
st.session_state.messages = [
|
105 |
+
{"role": "assistant", "content": "Ask me a question about Gita!"}
|
106 |
+
]
|
107 |
+
|
108 |
+
|
109 |
+
# creates a persistant index to disk
|
110 |
+
client = QdrantClient(
|
111 |
+
Q_END_POINT,
|
112 |
+
api_key=Q_API_KEY,
|
113 |
+
)
|
114 |
+
# create our vector store with hybrid indexing enabled
|
115 |
+
# batch_size controls how many nodes are encoded with sparse vectors at once
|
116 |
+
vector_store = QdrantVectorStore(
|
117 |
+
"bhagavad_gita", client=client, enable_hybrid=True, batch_size=20,force_disable_check_same_thread=True,
|
118 |
+
sparse_doc_fn=sparse_doc_vectors,
|
119 |
+
sparse_query_fn=sparse_query_vectors,
|
120 |
+
)
|
121 |
+
|
122 |
+
|
123 |
+
llm = HuggingFaceInferenceAPI(
|
124 |
+
model_name="mistralai/Mistral-7B-Instruct-v0.2",
|
125 |
+
token=HUGGINGFACEHUB_API_TOKEN,
|
126 |
+
context_window=8096,
|
127 |
+
)
|
128 |
+
Settings.llm = llm
|
129 |
+
Settings.tokenzier = AutoTokenizer.from_pretrained(
|
130 |
+
"mistralai/Mistral-7B-Instruct-v0.2"
|
131 |
+
)
|
132 |
+
|
133 |
+
embed_model = HuggingFaceEmbedding(model_name="BAAI/bge-base-en-v1.5", device="cpu")
|
134 |
+
Settings.embed_model = embed_model
|
135 |
+
|
136 |
+
index = VectorStoreIndex.from_vector_store(vector_store=vector_store,embed_model=embed_model)
|
137 |
+
|
138 |
+
from llama_index.core.memory import ChatMemoryBuffer
|
139 |
+
memory = ChatMemoryBuffer.from_defaults(token_limit=1500)
|
140 |
+
|
141 |
+
chat_engine = index.as_chat_engine(chat_mode="condense_question",
|
142 |
+
verbose=True,
|
143 |
+
memory=memory,
|
144 |
+
sparse_top_k=10,
|
145 |
+
vector_store_query_mode="hybrid",
|
146 |
+
similarity_top_k=3,
|
147 |
+
)
|
148 |
+
|
149 |
+
if prompt := st.chat_input("Your question"): # Prompt for user input and save to chat history
|
150 |
+
st.session_state.messages.append({"role": "user", "content": prompt})
|
151 |
+
|
152 |
+
for message in st.session_state.messages: # Display the prior chat messages
|
153 |
+
with st.chat_message(message["role"]):
|
154 |
+
st.write(message["content"])
|
155 |
+
|
156 |
+
# If last message is not from assistant, generate a new response
|
157 |
+
if st.session_state.messages[-1]["role"] != "assistant":
|
158 |
+
with st.chat_message("assistant"):
|
159 |
+
with st.spinner("Thinking..."):
|
160 |
+
response = chat_engine.chat(prompt)
|
161 |
+
st.write(response.response)
|
162 |
+
message = {"role": "assistant", "content": response.response}
|
163 |
+
st.session_state.messages.append(message) # Add response to message history
|
164 |
+
|
165 |
+
|
166 |
+
|
167 |
+
|