Spaces:
Runtime error
Runtime error
Update main.py
Browse files
main.py
CHANGED
@@ -1,4 +1,3 @@
|
|
1 |
-
import json # to work with JSON
|
2 |
import threading # to allow streaming response
|
3 |
import time # to pave the deliver of the message
|
4 |
|
@@ -6,46 +5,50 @@ import datasets # for loading RAG database
|
|
6 |
import faiss # to create a search index
|
7 |
import gradio # for the interface
|
8 |
import numpy # to work with vectors
|
9 |
-
import pandas # to work with pandas
|
10 |
import sentence_transformers # to load an embedding model
|
11 |
import spaces # for GPU
|
12 |
import transformers # to load an LLM
|
13 |
|
14 |
-
#
|
15 |
GREETING = (
|
16 |
"Howdy! I'm an AI agent that uses [retrieval-augmented generation](https://en.wikipedia.org/wiki/Retrieval-augmented_generation) "
|
17 |
"to answer questions about research published at [ASME IDETC](https://asmedigitalcollection.asme.org/IDETC-CIE) within the last 10 years or so. "
|
18 |
"I always try to cite my sources, but sometimes things get a little weird. "
|
19 |
"What can I tell you about today?"
|
20 |
)
|
|
|
|
|
21 |
EXAMPLE_QUERIES = [
|
22 |
"What's the difference between a markov chain and a hidden markov model?",
|
23 |
"What can you tell me about analytical target cascading?",
|
24 |
"What is known about different modes for human-AI teaming?",
|
25 |
]
|
|
|
|
|
26 |
EMBEDDING_MODEL_NAME = "all-MiniLM-L6-v2"
|
|
|
|
|
27 |
LLM_MODEL_NAME = "Qwen/Qwen2-7B-Instruct"
|
28 |
|
29 |
# Load the dataset and convert to pandas
|
30 |
data = datasets.load_dataset("ccm/rag-idetc")["train"].to_pandas()
|
31 |
|
32 |
# Load the model for later use in embeddings
|
33 |
-
|
34 |
|
35 |
# Create an LLM pipeline that we can send queries to
|
36 |
tokenizer = transformers.AutoTokenizer.from_pretrained(LLM_MODEL_NAME)
|
37 |
streamer = transformers.TextIteratorStreamer(
|
38 |
tokenizer, skip_prompt=True, skip_special_tokens=True
|
39 |
)
|
40 |
-
|
41 |
LLM_MODEL_NAME, torch_dtype="auto", device_map="auto"
|
42 |
)
|
43 |
|
44 |
# Create a FAISS index for fast similarity search
|
45 |
-
|
46 |
-
vectors = numpy.stack(data["embedding"].tolist(), axis=0).astype('float32')
|
47 |
index = faiss.IndexFlatL2(len(data["embedding"][0]))
|
48 |
-
index.metric_type =
|
49 |
faiss.normalize_L2(vectors)
|
50 |
index.train(vectors)
|
51 |
index.add(vectors)
|
@@ -60,7 +63,7 @@ def preprocess(query: str, k: int) -> tuple[str, str]:
|
|
60 |
Returns:
|
61 |
tuple[str, str]: A tuple containing the prompt and references
|
62 |
"""
|
63 |
-
encoded_query = numpy.expand_dims(
|
64 |
faiss.normalize_L2(encoded_query)
|
65 |
D, I = index.search(encoded_query, k)
|
66 |
top_five = data.loc[I[0]]
|
@@ -68,16 +71,16 @@ def preprocess(query: str, k: int) -> tuple[str, str]:
|
|
68 |
print(top_five["text"].values)
|
69 |
|
70 |
prompt = (
|
71 |
-
"You are an AI assistant who delights in helping people learn about research from the IDETC Conference."
|
72 |
"Your main task is to provide an ANSWER to the USER_QUERY based on the RESEARCH_EXCERPTS."
|
73 |
"Your ANSWER should be concise.\n\n"
|
74 |
-
"RESEARCH_EXCERPTS:\n{{
|
75 |
"USER_GUERY:\n{{QUERY_GOES_HERE}}\n\n"
|
76 |
"ANSWER:\n"
|
77 |
)
|
78 |
|
79 |
references = {}
|
80 |
-
|
81 |
|
82 |
for i in range(k):
|
83 |
title = top_five["title"].values[i]
|
@@ -86,21 +89,32 @@ def preprocess(query: str, k: int) -> tuple[str, str]:
|
|
86 |
path = top_five["path"].values[i]
|
87 |
text = top_five["text"].values[i]
|
88 |
|
89 |
-
|
|
|
|
|
90 |
header = "[" + title.title() + "](" + url + ")\n"
|
91 |
|
92 |
if header not in references.keys():
|
93 |
references[header] = []
|
94 |
|
95 |
references[header].append(text)
|
96 |
-
|
97 |
-
prompt = prompt.replace("{{
|
98 |
prompt = prompt.replace("{{QUERY_GOES_HERE}}", query)
|
99 |
|
100 |
print(references)
|
101 |
-
|
102 |
-
return prompt, "\n\n### References\n\n"
|
103 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
104 |
def postprocess(response: str, bypass_from_preprocessing: str) -> str:
|
105 |
"""
|
106 |
Applies a postprocessing step to the LLM's response before the user receives it
|
@@ -142,7 +156,7 @@ def reply(message: str, history: list[str]) -> str:
|
|
142 |
model_inputs = tokenizer([text], return_tensors="pt").to("cuda:0")
|
143 |
|
144 |
generate_kwargs = dict(model_inputs, streamer=streamer, max_new_tokens=512)
|
145 |
-
t = threading.Thread(target=
|
146 |
t.start()
|
147 |
|
148 |
partial_message = ""
|
@@ -160,7 +174,10 @@ gradio.ChatInterface(
|
|
160 |
reply,
|
161 |
examples=EXAMPLE_QUERIES,
|
162 |
chatbot=gradio.Chatbot(
|
163 |
-
avatar_images=
|
|
|
|
|
|
|
164 |
show_label=False,
|
165 |
show_share_button=False,
|
166 |
show_copy_button=False,
|
@@ -172,5 +189,3 @@ gradio.ChatInterface(
|
|
172 |
undo_btn=None,
|
173 |
clear_btn=None,
|
174 |
).launch(debug=True)
|
175 |
-
|
176 |
-
|
|
|
|
|
1 |
import threading # to allow streaming response
|
2 |
import time # to pave the deliver of the message
|
3 |
|
|
|
5 |
import faiss # to create a search index
|
6 |
import gradio # for the interface
|
7 |
import numpy # to work with vectors
|
|
|
8 |
import sentence_transformers # to load an embedding model
|
9 |
import spaces # for GPU
|
10 |
import transformers # to load an LLM
|
11 |
|
12 |
+
# The greeting supplied by the agent when it starts
|
13 |
GREETING = (
|
14 |
"Howdy! I'm an AI agent that uses [retrieval-augmented generation](https://en.wikipedia.org/wiki/Retrieval-augmented_generation) "
|
15 |
"to answer questions about research published at [ASME IDETC](https://asmedigitalcollection.asme.org/IDETC-CIE) within the last 10 years or so. "
|
16 |
"I always try to cite my sources, but sometimes things get a little weird. "
|
17 |
"What can I tell you about today?"
|
18 |
)
|
19 |
+
|
20 |
+
# Example queries supplied in the interface
|
21 |
EXAMPLE_QUERIES = [
|
22 |
"What's the difference between a markov chain and a hidden markov model?",
|
23 |
"What can you tell me about analytical target cascading?",
|
24 |
"What is known about different modes for human-AI teaming?",
|
25 |
]
|
26 |
+
|
27 |
+
# The embedding model used
|
28 |
EMBEDDING_MODEL_NAME = "all-MiniLM-L6-v2"
|
29 |
+
|
30 |
+
# The conversational model used
|
31 |
LLM_MODEL_NAME = "Qwen/Qwen2-7B-Instruct"
|
32 |
|
33 |
# Load the dataset and convert to pandas
|
34 |
data = datasets.load_dataset("ccm/rag-idetc")["train"].to_pandas()
|
35 |
|
36 |
# Load the model for later use in embeddings
|
37 |
+
embedding_model = sentence_transformers.SentenceTransformer(EMBEDDING_MODEL_NAME)
|
38 |
|
39 |
# Create an LLM pipeline that we can send queries to
|
40 |
tokenizer = transformers.AutoTokenizer.from_pretrained(LLM_MODEL_NAME)
|
41 |
streamer = transformers.TextIteratorStreamer(
|
42 |
tokenizer, skip_prompt=True, skip_special_tokens=True
|
43 |
)
|
44 |
+
chat_model = transformers.AutoModelForCausalLM.from_pretrained(
|
45 |
LLM_MODEL_NAME, torch_dtype="auto", device_map="auto"
|
46 |
)
|
47 |
|
48 |
# Create a FAISS index for fast similarity search
|
49 |
+
vectors = numpy.stack(data["embedding"].tolist(), axis=0).astype("float32")
|
|
|
50 |
index = faiss.IndexFlatL2(len(data["embedding"][0]))
|
51 |
+
index.metric_type = faiss.METRIC_INNER_PRODUCT
|
52 |
faiss.normalize_L2(vectors)
|
53 |
index.train(vectors)
|
54 |
index.add(vectors)
|
|
|
63 |
Returns:
|
64 |
tuple[str, str]: A tuple containing the prompt and references
|
65 |
"""
|
66 |
+
encoded_query = numpy.expand_dims(embedding_model.encode(query), axis=0)
|
67 |
faiss.normalize_L2(encoded_query)
|
68 |
D, I = index.search(encoded_query, k)
|
69 |
top_five = data.loc[I[0]]
|
|
|
71 |
print(top_five["text"].values)
|
72 |
|
73 |
prompt = (
|
74 |
+
"You are an AI assistant who delights in helping people learn about research from the IDETC Conference."
|
75 |
"Your main task is to provide an ANSWER to the USER_QUERY based on the RESEARCH_EXCERPTS."
|
76 |
"Your ANSWER should be concise.\n\n"
|
77 |
+
"RESEARCH_EXCERPTS:\n{{EXCERPTS_GO_HERE}}\n\n"
|
78 |
"USER_GUERY:\n{{QUERY_GOES_HERE}}\n\n"
|
79 |
"ANSWER:\n"
|
80 |
)
|
81 |
|
82 |
references = {}
|
83 |
+
research_excerpts = ""
|
84 |
|
85 |
for i in range(k):
|
86 |
title = top_five["title"].values[i]
|
|
|
89 |
path = top_five["path"].values[i]
|
90 |
text = top_five["text"].values[i]
|
91 |
|
92 |
+
research_excerpts += (
|
93 |
+
str(i + i) + ". This excerpt is from: '" + title + "':\n" + text + "\n"
|
94 |
+
)
|
95 |
header = "[" + title.title() + "](" + url + ")\n"
|
96 |
|
97 |
if header not in references.keys():
|
98 |
references[header] = []
|
99 |
|
100 |
references[header].append(text)
|
101 |
+
|
102 |
+
prompt = prompt.replace("{{EXCERPTS_GO_HERE}}", research_excerpts)
|
103 |
prompt = prompt.replace("{{QUERY_GOES_HERE}}", query)
|
104 |
|
105 |
print(references)
|
106 |
+
|
107 |
+
return prompt, "\n\n### References\n\n" + "\n".join(
|
108 |
+
[
|
109 |
+
str(i + 1)
|
110 |
+
+ ". "
|
111 |
+
+ ref
|
112 |
+
+ "\n - ".join(["", *['"...' + x + '..."' for x in references[ref]]])
|
113 |
+
for i, ref in enumerate(references.keys())
|
114 |
+
]
|
115 |
+
)
|
116 |
+
|
117 |
+
|
118 |
def postprocess(response: str, bypass_from_preprocessing: str) -> str:
|
119 |
"""
|
120 |
Applies a postprocessing step to the LLM's response before the user receives it
|
|
|
156 |
model_inputs = tokenizer([text], return_tensors="pt").to("cuda:0")
|
157 |
|
158 |
generate_kwargs = dict(model_inputs, streamer=streamer, max_new_tokens=512)
|
159 |
+
t = threading.Thread(target=chat_model.generate, kwargs=generate_kwargs)
|
160 |
t.start()
|
161 |
|
162 |
partial_message = ""
|
|
|
174 |
reply,
|
175 |
examples=EXAMPLE_QUERIES,
|
176 |
chatbot=gradio.Chatbot(
|
177 |
+
avatar_images=(
|
178 |
+
None,
|
179 |
+
"https://event.asme.org/Events/media/library/images/IDETC-CIE/IDETC-Logo-Announcements.png?ext=.png",
|
180 |
+
),
|
181 |
show_label=False,
|
182 |
show_share_button=False,
|
183 |
show_copy_button=False,
|
|
|
189 |
undo_btn=None,
|
190 |
clear_btn=None,
|
191 |
).launch(debug=True)
|
|
|
|