Spaces:
Build error
Build error
Vlad Severin
commited on
Commit
•
7550f47
1
Parent(s):
75ac2c1
add app.py
Browse files- .gitignore +1 -0
- app.py +130 -0
- clean_code_processed.csv +0 -0
- document_embeddings_clean_code.pickle +3 -0
- requirements.txt +4 -0
.gitignore
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
__pycache__
|
app.py
ADDED
@@ -0,0 +1,130 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import json
|
2 |
+
import openai
|
3 |
+
import tiktoken
|
4 |
+
import numpy as np
|
5 |
+
import pandas as pd
|
6 |
+
import gradio as gr
|
7 |
+
import pickle
|
8 |
+
|
9 |
+
COMPLETIONS_MODEL = "text-davinci-003"
|
10 |
+
EMBEDDING_MODEL = "text-embedding-ada-002"
|
11 |
+
|
12 |
+
with open("document_embeddings_clean_code.pickle", "rb") as document_embeddings_clean_code:
|
13 |
+
document_embeddings = pickle.load(document_embeddings_clean_code)
|
14 |
+
|
15 |
+
df = pd.read_csv('./clean_code_processed.csv')
|
16 |
+
df = df.set_index(["title", "section"])
|
17 |
+
df = df[df.tokens > 40]
|
18 |
+
df_for_embeddings = df
|
19 |
+
|
20 |
+
|
21 |
+
def get_embedding(text, model=EMBEDDING_MODEL):
|
22 |
+
result = openai.Embedding.create(
|
23 |
+
model=model,
|
24 |
+
input=text
|
25 |
+
)
|
26 |
+
return result["data"][0]["embedding"]
|
27 |
+
|
28 |
+
def vector_similarity(x, y):
|
29 |
+
"""
|
30 |
+
Returns the similarity between two vectors.
|
31 |
+
|
32 |
+
Because OpenAI Embeddings are normalized to length 1, the cosine similarity is the same as the dot product.
|
33 |
+
"""
|
34 |
+
return np.dot(np.array(x), np.array(y))
|
35 |
+
|
36 |
+
def order_document_sections_by_query_similarity(query, contexts):
|
37 |
+
"""
|
38 |
+
Find the query embedding for the supplied query, and compare it against all of the pre-calculated document embeddings
|
39 |
+
to find the most relevant sections.
|
40 |
+
|
41 |
+
Return the list of document sections, sorted by relevance in descending order.
|
42 |
+
"""
|
43 |
+
query_embedding = get_embedding(query)
|
44 |
+
|
45 |
+
document_similarities = sorted([
|
46 |
+
(vector_similarity(query_embedding, doc_embedding), doc_index) for doc_index, doc_embedding in contexts.items()
|
47 |
+
], reverse=True)
|
48 |
+
|
49 |
+
return document_similarities
|
50 |
+
|
51 |
+
def construct_prompt(question, context_embeddings, df):
|
52 |
+
"""
|
53 |
+
Fetch relevant
|
54 |
+
"""
|
55 |
+
most_relevant_document_sections = order_document_sections_by_query_similarity(question, context_embeddings)
|
56 |
+
|
57 |
+
chosen_sections = []
|
58 |
+
chosen_sections_len = 0
|
59 |
+
chosen_sections_indexes = []
|
60 |
+
|
61 |
+
for _, section_index in most_relevant_document_sections:
|
62 |
+
# Add contexts until we run out of space.
|
63 |
+
tokens = df._get_value(section_index, "tokens")
|
64 |
+
if type(tokens) != np.int64:
|
65 |
+
continue
|
66 |
+
|
67 |
+
chosen_sections_len += df._get_value(section_index, "tokens") + separator_len
|
68 |
+
if chosen_sections_len > MAX_SECTION_LEN:
|
69 |
+
break
|
70 |
+
|
71 |
+
chosen_sections.append(SEPARATOR + df._get_value(section_index, "content").replace("\n", " "))
|
72 |
+
chosen_sections_indexes.append(section_index)
|
73 |
+
|
74 |
+
# Useful diagnostic information
|
75 |
+
print(f"Selected {len(chosen_sections)} document sections:")
|
76 |
+
print("\n".join(str(index) for index in chosen_sections_indexes))
|
77 |
+
|
78 |
+
header = """Answer the question as truthfully as possible using the provided context, and if the answer is not contained within the text below, say "I don't know."\n\nContext:\n"""
|
79 |
+
|
80 |
+
return (header + "".join(chosen_sections) + "\n\n Q: " + question + "\n A:", chosen_sections_indexes)
|
81 |
+
|
82 |
+
|
83 |
+
|
84 |
+
MAX_SECTION_LEN = 2000
|
85 |
+
SEPARATOR = "\n* "
|
86 |
+
ENCODING = "gpt2" # encoding for text-davinci-003
|
87 |
+
|
88 |
+
encoding = tiktoken.get_encoding(ENCODING)
|
89 |
+
separator_len = len(encoding.encode(SEPARATOR))
|
90 |
+
|
91 |
+
COMPLETIONS_API_PARAMS = {
|
92 |
+
# We use temperature of 0.0 because it gives the most predictable, factual answer.
|
93 |
+
"temperature": 0.0,
|
94 |
+
"max_tokens": 1500,
|
95 |
+
"model": COMPLETIONS_MODEL,
|
96 |
+
}
|
97 |
+
|
98 |
+
def answer_query_with_context(
|
99 |
+
query,
|
100 |
+
df,
|
101 |
+
document_embeddings
|
102 |
+
):
|
103 |
+
prompt, chosen_sections_indexes = construct_prompt(
|
104 |
+
query,
|
105 |
+
document_embeddings,
|
106 |
+
df
|
107 |
+
)
|
108 |
+
|
109 |
+
for i in range(len(chosen_sections_indexes)):
|
110 |
+
chosen_sections_indexes[i] = chosen_sections_indexes[i][0]
|
111 |
+
|
112 |
+
response = openai.Completion.create(
|
113 |
+
prompt=prompt,
|
114 |
+
**COMPLETIONS_API_PARAMS
|
115 |
+
)
|
116 |
+
|
117 |
+
return (response["choices"][0]["text"].strip(" \n"), chosen_sections_indexes)
|
118 |
+
|
119 |
+
def handle(question):
|
120 |
+
answer, related_documents = answer_query_with_context(question, df_for_embeddings, document_embeddings)
|
121 |
+
return answer + "\n\nRelated chapters:\n" + "\n".join(related_documents)
|
122 |
+
|
123 |
+
demo = gr.Interface(
|
124 |
+
fn=handle,
|
125 |
+
inputs="text",
|
126 |
+
outputs="text",
|
127 |
+
cache_examples=False,
|
128 |
+
)
|
129 |
+
|
130 |
+
demo.launch()
|
clean_code_processed.csv
ADDED
The diff for this file is too large to render.
See raw diff
|
|
document_embeddings_clean_code.pickle
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:600168ec8f03d7f4c56ab5493d7ed71cb283cababcc0e24121b85b9512a41914
|
3 |
+
size 2216273
|
requirements.txt
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
openai
|
2 |
+
tiktoken
|
3 |
+
numpy
|
4 |
+
pandas
|