Update app.py
Browse files
app.py
CHANGED
@@ -1,15 +1,3 @@
|
|
1 |
-
# -*- coding: utf-8 -*-
|
2 |
-
"""LLM Comparison
|
3 |
-
|
4 |
-
Automatically generated by Colab.
|
5 |
-
|
6 |
-
Original file is located at
|
7 |
-
https://colab.research.google.com/drive/156SKaX3DY6jwOhcpwZVM5AiLscOAbNNJ
|
8 |
-
"""
|
9 |
-
|
10 |
-
# Commented out IPython magic to ensure Python compatibility.
|
11 |
-
# %pip install -qU pixeltable gradio sentence-transformers tiktoken openai openpyxl
|
12 |
-
|
13 |
import gradio as gr
|
14 |
import pandas as pd
|
15 |
import pixeltable as pxt
|
@@ -50,22 +38,29 @@ def create_prompt(top_k_list: list[dict], question: str) -> str:
|
|
50 |
|
51 |
{question}'''
|
52 |
|
|
|
53 |
def process_files(ground_truth_file, pdf_files):
|
54 |
-
#
|
|
|
|
|
|
|
|
|
|
|
55 |
if ground_truth_file.name.endswith('.csv'):
|
56 |
queries_t = pxt.io.import_csv('rag_demo.queries', ground_truth_file.name)
|
57 |
else:
|
58 |
queries_t = pxt.io.import_excel('rag_demo.queries', ground_truth_file.name)
|
59 |
|
60 |
-
#
|
61 |
documents_t = pxt.create_table(
|
62 |
'rag_demo.documents',
|
63 |
{'document': pxt.DocumentType()}
|
64 |
)
|
65 |
|
|
|
66 |
documents_t.insert({'document': file.name} for file in pdf_files if file.name.endswith('.pdf'))
|
67 |
|
68 |
-
|
69 |
chunks_t = pxt.create_view(
|
70 |
'rag_demo.chunks',
|
71 |
documents_t,
|
@@ -76,10 +71,10 @@ def process_files(ground_truth_file, pdf_files):
|
|
76 |
)
|
77 |
)
|
78 |
|
79 |
-
# Add embedding index
|
80 |
chunks_t.add_embedding_index('text', string_embed=e5_embed)
|
81 |
|
82 |
-
#
|
83 |
@chunks_t.query
|
84 |
def top_k(query_text: str):
|
85 |
sim = chunks_t.text.similarity(query_text)
|
@@ -89,13 +84,13 @@ def process_files(ground_truth_file, pdf_files):
|
|
89 |
.limit(5)
|
90 |
)
|
91 |
|
92 |
-
# Add computed columns to
|
93 |
queries_t['question_context'] = chunks_t.top_k(queries_t.Question)
|
94 |
queries_t['prompt'] = create_prompt(
|
95 |
queries_t.question_context, queries_t.Question
|
96 |
)
|
97 |
|
98 |
-
# Prepare messages for OpenAI
|
99 |
messages = [
|
100 |
{
|
101 |
'role': 'system',
|
@@ -109,17 +104,18 @@ def process_files(ground_truth_file, pdf_files):
|
|
109 |
|
110 |
# Add OpenAI response column
|
111 |
queries_t['response'] = openai.chat_completions(
|
112 |
-
model='gpt-4o-mini-2024-07-18
|
113 |
)
|
114 |
|
115 |
-
|
|
|
116 |
|
|
|
117 |
df_output = queries_t.select(queries_t.Question, queries_t.correct_answer, queries_t.answer).collect().to_pandas()
|
118 |
|
119 |
try:
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
except Exception as e:
|
124 |
return f"An error occurred: {str(e)}", None
|
125 |
|
@@ -127,18 +123,21 @@ def process_files(ground_truth_file, pdf_files):
|
|
127 |
with gr.Blocks() as demo:
|
128 |
gr.Markdown("# RAG Demo App")
|
129 |
|
|
|
130 |
with gr.Row():
|
131 |
ground_truth_file = gr.File(label="Upload Ground Truth (CSV or XLSX)", file_count="single")
|
132 |
pdf_files = gr.File(label="Upload PDF Documents", file_count="multiple")
|
133 |
|
134 |
-
|
|
|
135 |
|
|
|
136 |
df_output = gr.DataFrame(label="Pixeltable Table")
|
137 |
|
|
|
138 |
#question_input = gr.Textbox(label="Enter your question")
|
139 |
#query_button = gr.Button("Query LLM")
|
140 |
-
|
141 |
-
process_button.click(process_files, inputs=[ground_truth_file, pdf_files], outputs=df_output)
|
142 |
#query_button.click(query_llm, inputs=question_input, outputs=output_dataframe)
|
143 |
|
144 |
if __name__ == "__main__":
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
import pandas as pd
|
3 |
import pixeltable as pxt
|
|
|
38 |
|
39 |
{question}'''
|
40 |
|
41 |
+
# Gradio Application
|
42 |
def process_files(ground_truth_file, pdf_files):
|
43 |
+
# Ensure a clean slate for the demo by removing and recreating the 'rag_demo' directory
|
44 |
+
pxt.drop_dir('rag_demo', force=True)
|
45 |
+
pxt.create_dir('rag_demo')
|
46 |
+
|
47 |
+
# Process the ground truth file, which contains questions and correct answers
|
48 |
+
# Import as CSV or Excel depending on the file extension
|
49 |
if ground_truth_file.name.endswith('.csv'):
|
50 |
queries_t = pxt.io.import_csv('rag_demo.queries', ground_truth_file.name)
|
51 |
else:
|
52 |
queries_t = pxt.io.import_excel('rag_demo.queries', ground_truth_file.name)
|
53 |
|
54 |
+
# Create a table to store the uploaded PDF documents
|
55 |
documents_t = pxt.create_table(
|
56 |
'rag_demo.documents',
|
57 |
{'document': pxt.DocumentType()}
|
58 |
)
|
59 |
|
60 |
+
# Insert the PDF files into the documents table
|
61 |
documents_t.insert({'document': file.name} for file in pdf_files if file.name.endswith('.pdf'))
|
62 |
|
63 |
+
# Create a view that splits the documents into smaller chunks
|
64 |
chunks_t = pxt.create_view(
|
65 |
'rag_demo.chunks',
|
66 |
documents_t,
|
|
|
71 |
)
|
72 |
)
|
73 |
|
74 |
+
# Add an embedding index to the chunks for similarity search
|
75 |
chunks_t.add_embedding_index('text', string_embed=e5_embed)
|
76 |
|
77 |
+
# Define a query function to retrieve the top-k most similar chunks for a given question
|
78 |
@chunks_t.query
|
79 |
def top_k(query_text: str):
|
80 |
sim = chunks_t.text.similarity(query_text)
|
|
|
84 |
.limit(5)
|
85 |
)
|
86 |
|
87 |
+
# Add computed columns to the queries table for context retrieval and prompt creation
|
88 |
queries_t['question_context'] = chunks_t.top_k(queries_t.Question)
|
89 |
queries_t['prompt'] = create_prompt(
|
90 |
queries_t.question_context, queries_t.Question
|
91 |
)
|
92 |
|
93 |
+
# Prepare messages for the OpenAI API, including system instructions and user prompt
|
94 |
messages = [
|
95 |
{
|
96 |
'role': 'system',
|
|
|
104 |
|
105 |
# Add OpenAI response column
|
106 |
queries_t['response'] = openai.chat_completions(
|
107 |
+
model='gpt-4o-mini-2024-07-18, messages=messages
|
108 |
)
|
109 |
|
110 |
+
# Extract the answer text from the API response
|
111 |
+
queries_t['answer'] = queries_t.response.choices[0].message.content.astype(pxt.StringType())
|
112 |
|
113 |
+
# Prepare the output dataframe with questions, correct answers, and model-generated answers
|
114 |
df_output = queries_t.select(queries_t.Question, queries_t.correct_answer, queries_t.answer).collect().to_pandas()
|
115 |
|
116 |
try:
|
117 |
+
# Return the output dataframe for display
|
118 |
+
return df_output
|
|
|
119 |
except Exception as e:
|
120 |
return f"An error occurred: {str(e)}", None
|
121 |
|
|
|
123 |
with gr.Blocks() as demo:
|
124 |
gr.Markdown("# RAG Demo App")
|
125 |
|
126 |
+
# File upload components for ground truth and PDF documents
|
127 |
with gr.Row():
|
128 |
ground_truth_file = gr.File(label="Upload Ground Truth (CSV or XLSX)", file_count="single")
|
129 |
pdf_files = gr.File(label="Upload PDF Documents", file_count="multiple")
|
130 |
|
131 |
+
# Button to trigger file processing
|
132 |
+
process_button = gr.Button("Process Files and Generate Outputs")
|
133 |
|
134 |
+
# Output component to display the results
|
135 |
df_output = gr.DataFrame(label="Pixeltable Table")
|
136 |
|
137 |
+
process_button.click(process_files, inputs=[ground_truth_file, pdf_files], outputs=df_output)
|
138 |
#question_input = gr.Textbox(label="Enter your question")
|
139 |
#query_button = gr.Button("Query LLM")
|
140 |
+
|
|
|
141 |
#query_button.click(query_llm, inputs=question_input, outputs=output_dataframe)
|
142 |
|
143 |
if __name__ == "__main__":
|