Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,150 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
from litellm import completion
|
3 |
+
import os
|
4 |
+
|
5 |
+
os.environ['GROQ_API_KEY'] = "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"
|
6 |
+
response = completion(
|
7 |
+
model="groq/llama3-8b-8192",
|
8 |
+
messages=[
|
9 |
+
{"role": "user", "content": "hello from litellm"}
|
10 |
+
],
|
11 |
+
)
|
12 |
+
from datasets import load_dataset
|
13 |
+
|
14 |
+
dataset = load_dataset("hugginglearners/russia-ukraine-conflict-articles")
|
15 |
+
|
16 |
+
|
17 |
+
docs = [item['articles'] for item in dataset['train'].select(range(10))]
|
18 |
+
def chunk_document(doc: str, doc_id: int, desired_chunk_size: int = 100, max_chunk_size: int = 3000):
|
19 |
+
chunk = ''
|
20 |
+
chunk_number = 0
|
21 |
+
for line in doc.splitlines():
|
22 |
+
chunk += line + '\n'
|
23 |
+
if len(chunk) >= desired_chunk_size:
|
24 |
+
yield (doc_id, chunk_number, chunk[:max_chunk_size])
|
25 |
+
chunk = ''
|
26 |
+
chunk_number += 1
|
27 |
+
if chunk:
|
28 |
+
yield (doc_id, chunk_number, chunk)
|
29 |
+
|
30 |
+
def chunk_documents(docs: List[str], desired_chunk_size: int = 100, max_chunk_size: int = 3000):
|
31 |
+
chunks = []
|
32 |
+
for doc_id, doc in enumerate(docs):
|
33 |
+
chunks.extend(chunk_document(doc, doc_id, desired_chunk_size, max_chunk_size))
|
34 |
+
return chunks
|
35 |
+
|
36 |
+
from typing import List
|
37 |
+
import numpy as np
|
38 |
+
from rank_bm25 import BM25Okapi
|
39 |
+
from sentence_transformers import SentenceTransformer
|
40 |
+
import torch
|
41 |
+
class Retriever:
|
42 |
+
def __init__(self, docs: List[str]):
|
43 |
+
|
44 |
+
self.chunks = chunk_documents(docs)
|
45 |
+
self.docs = [chunk[2] for chunk in self.chunks]
|
46 |
+
tokenized_docs = [doc.lower().split(" ") for doc in self.docs]
|
47 |
+
self.bm25 = BM25Okapi(tokenized_docs)
|
48 |
+
self.sbert = SentenceTransformer('sentence-transformers/all-distilroberta-v1')
|
49 |
+
self.doc_embeddings = self.sbert.encode(self.docs)
|
50 |
+
|
51 |
+
def get_docs(self, query, method="bm25", n=3):
|
52 |
+
if method == "bm25":
|
53 |
+
scores = self._get_bm25_scores(query)
|
54 |
+
elif method == "sbert":
|
55 |
+
scores = self._get_semantic_scores(query)
|
56 |
+
elif method == "hybrid":
|
57 |
+
bm25_scores = self._get_bm25_scores(query)
|
58 |
+
semantic_scores = self._get_semantic_scores(query)
|
59 |
+
scores = 0.3 * bm25_scores + 0.7 * semantic_scores
|
60 |
+
else:
|
61 |
+
raise ValueError("Invalid method. Choose 'bm25', 'sbert', or 'hybrid'.")
|
62 |
+
|
63 |
+
sorted_indices = np.argsort(scores)[::-1]
|
64 |
+
# Повертаємо перші n документів із інформацією про джерело
|
65 |
+
return [(self.chunks[i][0], self.chunks[i][1], self.docs[i]) for i in sorted_indices[:n]]
|
66 |
+
|
67 |
+
def _get_bm25_scores(self, query):
|
68 |
+
tokenized_query = query.lower().split(" ")
|
69 |
+
return self.bm25.get_scores(tokenized_query)
|
70 |
+
|
71 |
+
def _get_semantic_scores(self, query):
|
72 |
+
query_embedding = self.sbert.encode(query)
|
73 |
+
scores = torch.cosine_similarity(
|
74 |
+
torch.tensor(query_embedding).unsqueeze(0),
|
75 |
+
torch.tensor(self.doc_embeddings),
|
76 |
+
dim=1
|
77 |
+
)
|
78 |
+
return scores.numpy()
|
79 |
+
class QuestionAnsweringBot:
|
80 |
+
PROMPT = '''\
|
81 |
+
You are a helpful assistant that can answer questions.
|
82 |
+
|
83 |
+
Rules:
|
84 |
+
-Reply with the answer only and nothing but the answer.
|
85 |
+
-Say 'I don't know(((' if you don't know the answer.
|
86 |
+
-Use the provided context.
|
87 |
+
'''
|
88 |
+
|
89 |
+
def __init__(self, docs):
|
90 |
+
self.retriever = Retriever(docs)
|
91 |
+
|
92 |
+
def answer_question(self, question: str, method: str = "bm25") -> str:
|
93 |
+
context_with_indices = self.retriever.get_docs(question, method=method)
|
94 |
+
if not context_with_indices:
|
95 |
+
return "I don't know((("
|
96 |
+
|
97 |
+
# контекст для моделі
|
98 |
+
context = "\n".join([f"Doc {doc_id}, Chunk {chunk_id}: {text}" for doc_id, chunk_id, text in context_with_indices])
|
99 |
+
|
100 |
+
messages = [
|
101 |
+
{"role": "system", "content": self.PROMPT},
|
102 |
+
{"role": "user", "content": f"Context: {context}\nQuestion: {question}"}
|
103 |
+
]
|
104 |
+
|
105 |
+
try:
|
106 |
+
|
107 |
+
completionn = completion(
|
108 |
+
model="groq/llama3-8b-8192",
|
109 |
+
messages=messages,
|
110 |
+
)
|
111 |
+
# Відповідь
|
112 |
+
answer = completionn['choices'][0]['message']['content']
|
113 |
+
|
114 |
+
# джерела
|
115 |
+
sources = [f"Doc {doc_id}: Chunk {chunk_id}; " for doc_id, chunk_id, _ in context_with_indices]
|
116 |
+
return f"{answer} [{', '.join(sources)}]"
|
117 |
+
except Exception as e:
|
118 |
+
return f"Error: {str(e)}"
|
119 |
+
|
120 |
+
|
121 |
+
# question = "Tell about war"
|
122 |
+
docs = docs
|
123 |
+
# bot = QuestionAnsweringBot(docs)
|
124 |
+
# answer = bot.answer_question(question)
|
125 |
+
|
126 |
+
# print(f'Q: {question}')
|
127 |
+
# print(f'A: {answer}')
|
128 |
+
import gradio as gr
|
129 |
+
|
130 |
+
def answer_question_with_method(query, method):
|
131 |
+
bot = QuestionAnsweringBot(docs)
|
132 |
+
return bot.answer_question(query, method=method)
|
133 |
+
|
134 |
+
|
135 |
+
# Створення інтерфейсу
|
136 |
+
demo = gr.Interface(
|
137 |
+
fn=answer_question_with_method,
|
138 |
+
inputs=[
|
139 |
+
gr.Textbox(label="Your Question"),
|
140 |
+
gr.Dropdown(
|
141 |
+
choices=["bm25", "sbert", "hybrid"],
|
142 |
+
value="hybrid",
|
143 |
+
label="Select Retrieval Method"
|
144 |
+
)
|
145 |
+
],
|
146 |
+
outputs="text"
|
147 |
+
)
|
148 |
+
|
149 |
+
demo.launch()
|
150 |
+
|