Spaces:
Sleeping
Sleeping
drakosfire
commited on
Commit
•
9bade29
1
Parent(s):
c21541c
Basic Logging, timestamp, simple token estimate, source files
Browse files- SRD_embeddings.csv +2 -2
- app.py +54 -12
SRD_embeddings.csv
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:0ffdfe9de524d440d57d359270fe8a774009188b528946a79a55f8dd7294e5fe
|
3 |
+
size 51272010
|
app.py
CHANGED
@@ -5,6 +5,7 @@ from sentence_transformers import util, SentenceTransformer
|
|
5 |
import torch
|
6 |
import time
|
7 |
from time import perf_counter as timer
|
|
|
8 |
import textwrap
|
9 |
import json
|
10 |
import textwrap
|
@@ -15,10 +16,7 @@ print("Launching")
|
|
15 |
|
16 |
client = OpenAI()
|
17 |
|
18 |
-
|
19 |
-
def print_wrapped(text, wrap_length=80):
|
20 |
-
wrapped_text = textwrap.fill(text, wrap_length)
|
21 |
-
print(wrapped_text)
|
22 |
|
23 |
# Import saved file and view
|
24 |
embeddings_df_save_path = "./SRD_embeddings.csv"
|
@@ -39,6 +37,28 @@ pages_and_chunks = text_chunks_and_embedding_df_load.to_dict(orient="records")
|
|
39 |
# Convert embeddings to torch tensor and send to device (note: NumPy arrays are float64, torch tensors are float32 by default)
|
40 |
embeddings = torch.tensor(np.array(text_chunks_and_embedding_df_load["embedding"].tolist()), dtype=torch.float32).to('cpu')
|
41 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
def retrieve_relevant_resources(query: str,
|
43 |
embeddings: torch.tensor,
|
44 |
model: SentenceTransformer=embedding_model,
|
@@ -84,18 +104,23 @@ def print_top_results_and_scores(query: str,
|
|
84 |
# Loop through zipped together scores and indicies
|
85 |
for score, index in zip(scores, indices):
|
86 |
print(f"Score: {score:.4f}")
|
|
|
87 |
# Print relevant sentence chunk (since the scores are in descending order, the most relevant chunk will be first)
|
88 |
print_wrapped(pages_and_chunks[index]["sentence_chunk"])
|
89 |
# Print the page number too so we can reference the textbook further and check the results
|
90 |
print(f"File of Origin: {pages_and_chunks[index]['file_path']}")
|
91 |
print("\n")
|
92 |
|
|
|
|
|
93 |
def prompt_formatter(query: str,
|
94 |
context_items: list[dict]) -> str:
|
95 |
-
"""
|
96 |
-
Augments query with text-based context from context_items.
|
97 |
-
"""
|
98 |
# Join context items into one dotted paragraph
|
|
|
|
|
|
|
|
|
|
|
99 |
context = "- " + "\n- ".join([item["sentence_chunk"] for item in context_items])
|
100 |
|
101 |
# Create a base prompt with examples to help the model
|
@@ -139,23 +164,36 @@ Use the context provided to answer the user's query concisely. """
|
|
139 |
|
140 |
|
141 |
with gr.Blocks() as RulesLawyer:
|
|
|
|
|
|
|
142 |
chatbot = gr.Chatbot()
|
143 |
msg = gr.Textbox()
|
144 |
clear = gr.ClearButton([msg, chatbot])
|
145 |
|
|
|
|
|
|
|
|
|
|
|
146 |
def respond(message, chat_history):
|
|
|
|
|
|
|
|
|
147 |
|
148 |
# Get relevant resources
|
149 |
-
scores, indices =
|
150 |
embeddings=embeddings)
|
151 |
-
|
152 |
# Create a list of context items
|
153 |
context_items = [pages_and_chunks[i] for i in indices]
|
|
|
154 |
|
155 |
# Format prompt with context items
|
156 |
-
prompt = prompt_formatter(query=message,
|
157 |
context_items=context_items)
|
158 |
-
|
159 |
bot_message = client.chat.completions.create(
|
160 |
model="gpt-4",
|
161 |
messages=[
|
@@ -171,9 +209,13 @@ with gr.Blocks() as RulesLawyer:
|
|
171 |
presence_penalty=0
|
172 |
)
|
173 |
chat_history.append((message, bot_message.choices[0].message.content))
|
|
|
|
|
174 |
time.sleep(2)
|
175 |
return "", chat_history
|
176 |
-
msg.
|
|
|
|
|
177 |
|
178 |
if __name__ == "__main__":
|
179 |
RulesLawyer.launch()
|
|
|
5 |
import torch
|
6 |
import time
|
7 |
from time import perf_counter as timer
|
8 |
+
from datetime import datetime
|
9 |
import textwrap
|
10 |
import json
|
11 |
import textwrap
|
|
|
16 |
|
17 |
client = OpenAI()
|
18 |
|
19 |
+
|
|
|
|
|
|
|
20 |
|
21 |
# Import saved file and view
|
22 |
embeddings_df_save_path = "./SRD_embeddings.csv"
|
|
|
37 |
# Convert embeddings to torch tensor and send to device (note: NumPy arrays are float64, torch tensors are float32 by default)
|
38 |
embeddings = torch.tensor(np.array(text_chunks_and_embedding_df_load["embedding"].tolist()), dtype=torch.float32).to('cpu')
|
39 |
|
40 |
+
# Define helper function to print wrapped text
|
41 |
+
def print_wrapped(text, wrap_length=80):
|
42 |
+
wrapped_text = textwrap.fill(text, wrap_length)
|
43 |
+
print(wrapped_text)
|
44 |
+
|
45 |
+
def hybrid_estimate_tokens(text: str)-> float:
|
46 |
+
# Part 1: Estimate based on spaces and punctuation
|
47 |
+
estimated_words = text.count(' ') + 1 # Counting words by spaces
|
48 |
+
punctuation_count = sum(1 for char in text if char in ',.!?;:') # Counting punctuation as potential separate tokens
|
49 |
+
estimate1 = estimated_words + punctuation_count
|
50 |
+
|
51 |
+
# Part 2: Estimate based on total characters divided by average token length
|
52 |
+
average_token_length = 4
|
53 |
+
total_characters = len(text)
|
54 |
+
estimate2 = (total_characters // average_token_length) + punctuation_count
|
55 |
+
|
56 |
+
# Average the two estimates
|
57 |
+
estimated_tokens = (estimate1 + estimate2) / 2
|
58 |
+
|
59 |
+
return estimated_tokens
|
60 |
+
|
61 |
+
|
62 |
def retrieve_relevant_resources(query: str,
|
63 |
embeddings: torch.tensor,
|
64 |
model: SentenceTransformer=embedding_model,
|
|
|
104 |
# Loop through zipped together scores and indicies
|
105 |
for score, index in zip(scores, indices):
|
106 |
print(f"Score: {score:.4f}")
|
107 |
+
print(f"Token Count : {pages_and_chunks[index]['chunk_token_count']}")
|
108 |
# Print relevant sentence chunk (since the scores are in descending order, the most relevant chunk will be first)
|
109 |
print_wrapped(pages_and_chunks[index]["sentence_chunk"])
|
110 |
# Print the page number too so we can reference the textbook further and check the results
|
111 |
print(f"File of Origin: {pages_and_chunks[index]['file_path']}")
|
112 |
print("\n")
|
113 |
|
114 |
+
return scores, indices
|
115 |
+
|
116 |
def prompt_formatter(query: str,
|
117 |
context_items: list[dict]) -> str:
|
|
|
|
|
|
|
118 |
# Join context items into one dotted paragraph
|
119 |
+
# print(context_items[0])
|
120 |
+
|
121 |
+
# Alternate print method
|
122 |
+
# print("\n".join([item["file_path"] + "\n" + str(item['chunk_token_count']) + "\n" + item["sentence_chunk"] for item in context_items]))
|
123 |
+
|
124 |
context = "- " + "\n- ".join([item["sentence_chunk"] for item in context_items])
|
125 |
|
126 |
# Create a base prompt with examples to help the model
|
|
|
164 |
|
165 |
|
166 |
with gr.Blocks() as RulesLawyer:
|
167 |
+
|
168 |
+
message_state = gr.State()
|
169 |
+
chatbot_state = gr.State([])
|
170 |
chatbot = gr.Chatbot()
|
171 |
msg = gr.Textbox()
|
172 |
clear = gr.ClearButton([msg, chatbot])
|
173 |
|
174 |
+
def store_message(message):
|
175 |
+
|
176 |
+
return message
|
177 |
+
|
178 |
+
|
179 |
def respond(message, chat_history):
|
180 |
+
print(datetime.now())
|
181 |
+
print(f"User Input : {message}")
|
182 |
+
print(f"Chat History: {chat_history}")
|
183 |
+
print(f"""Token Estimate: {hybrid_estimate_tokens(f"{message} {chat_history}")}""")
|
184 |
|
185 |
# Get relevant resources
|
186 |
+
scores, indices = print_top_results_and_scores(query=message,
|
187 |
embeddings=embeddings)
|
188 |
+
|
189 |
# Create a list of context items
|
190 |
context_items = [pages_and_chunks[i] for i in indices]
|
191 |
+
|
192 |
|
193 |
# Format prompt with context items
|
194 |
+
prompt = prompt_formatter(query=f"Chat History : {chat_history} + {message}",
|
195 |
context_items=context_items)
|
196 |
+
|
197 |
bot_message = client.chat.completions.create(
|
198 |
model="gpt-4",
|
199 |
messages=[
|
|
|
209 |
presence_penalty=0
|
210 |
)
|
211 |
chat_history.append((message, bot_message.choices[0].message.content))
|
212 |
+
print(f"Response : {bot_message.choices[0].message.content}")
|
213 |
+
|
214 |
time.sleep(2)
|
215 |
return "", chat_history
|
216 |
+
msg.change(store_message, inputs = [msg], outputs = [message_state])
|
217 |
+
chatbot.change(store_message, [chatbot], [chatbot_state])
|
218 |
+
msg.submit(respond, [message_state, chatbot_state], [msg, chatbot])
|
219 |
|
220 |
if __name__ == "__main__":
|
221 |
RulesLawyer.launch()
|