search.py
CHANGED
@@ -1,12 +1,25 @@
|
|
1 |
-
from transformers import
|
2 |
-
import pandas as pd
|
3 |
-
from pdfminer.high_level import extract_text
|
4 |
from docx import Document
|
|
|
|
|
5 |
from dataclasses import dataclass
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
|
11 |
@dataclass
|
12 |
class Paragraph:
|
@@ -14,19 +27,109 @@ class Paragraph:
|
|
14 |
paragraph_num: int
|
15 |
content: str
|
16 |
|
17 |
-
def read_pdf_pdfminer(file_path) ->
|
18 |
text = extract_text(file_path).replace('\n', ' ').strip()
|
19 |
-
paragraphs =
|
20 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
|
22 |
-
def read_docx(file) ->
|
23 |
doc = Document(file)
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
32 |
return answer
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
|
|
|
|
2 |
from docx import Document
|
3 |
+
from pdfminer.high_level import extract_text
|
4 |
+
from transformers import GPT2Tokenizer
|
5 |
from dataclasses import dataclass
|
6 |
+
from typing import List
|
7 |
+
from tqdm import tqdm
|
8 |
+
import os
|
9 |
+
import pandas as pd
|
10 |
+
import re
|
11 |
+
from sklearn.feature_extraction.text import TfidfVectorizer
|
12 |
+
import numpy as np
|
13 |
+
|
14 |
+
tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-2", trust_remote_code=True)
|
15 |
+
model = AutoModelForCausalLM.from_pretrained("microsoft/phi-2", trust_remote_code=True)
|
16 |
+
|
17 |
+
EMBEDDING_SEG_LEN = 1500
|
18 |
+
EMBEDDING_MODEL = "gpt-4"
|
19 |
|
20 |
+
EMBEDDING_CTX_LENGTH = 8191
|
21 |
+
EMBEDDING_ENCODING = "cl100k_base"
|
22 |
+
ENCODING = "gpt2"
|
23 |
|
24 |
@dataclass
|
25 |
class Paragraph:
|
|
|
27 |
paragraph_num: int
|
28 |
content: str
|
29 |
|
30 |
+
def read_pdf_pdfminer(file_path) -> List[Paragraph]:
|
31 |
text = extract_text(file_path).replace('\n', ' ').strip()
|
32 |
+
paragraphs = batched(text, EMBEDDING_SEG_LEN)
|
33 |
+
paragraphs_objs = []
|
34 |
+
paragraph_num = 1
|
35 |
+
for p in paragraphs:
|
36 |
+
para = Paragraph(0, paragraph_num, p)
|
37 |
+
paragraphs_objs.append(para)
|
38 |
+
paragraph_num += 1
|
39 |
+
return paragraphs_objs
|
40 |
|
41 |
+
def read_docx(file) -> List[Paragraph]:
|
42 |
doc = Document(file)
|
43 |
+
paragraphs = []
|
44 |
+
for paragraph_num, paragraph in enumerate(doc.paragraphs, start=1):
|
45 |
+
content = paragraph.text.strip()
|
46 |
+
if content:
|
47 |
+
para = Paragraph(1, paragraph_num, content)
|
48 |
+
paragraphs.append(para)
|
49 |
+
return paragraphs
|
50 |
+
|
51 |
+
def count_tokens(text, tokenizer):
|
52 |
+
return len(tokenizer.encode(text))
|
53 |
+
|
54 |
+
def batched(iterable, n):
|
55 |
+
l = len(iterable)
|
56 |
+
for ndx in range(0, l, n):
|
57 |
+
yield iterable[ndx : min(ndx + n, l)]
|
58 |
+
|
59 |
+
def compute_doc_embeddings(df, tokenizer):
|
60 |
+
embeddings = {}
|
61 |
+
for index, row in tqdm(df.iterrows(), total=df.shape[0]):
|
62 |
+
doc = row["content"]
|
63 |
+
doc_embedding = get_embedding(doc, tokenizer)
|
64 |
+
embeddings[index] = doc_embedding
|
65 |
+
return embeddings
|
66 |
+
|
67 |
+
def enhanced_context_extraction(document, keywords, vectorizer, tfidf_scores, top_n=5):
|
68 |
+
paragraphs = [para for para in document.split("\n") if para]
|
69 |
+
scores = [sum([para.lower().count(keyword) * tfidf_scores[vectorizer.vocabulary_[keyword]] for keyword in keywords if keyword in para.lower()]) for para in paragraphs]
|
70 |
+
|
71 |
+
top_indices = sorted(range(len(scores)), key=lambda i: scores[i], reverse=True)[:top_n]
|
72 |
+
relevant_paragraphs = [paragraphs[i] for i in top_indices]
|
73 |
+
|
74 |
+
return " ".join(relevant_paragraphs)
|
75 |
+
|
76 |
+
def targeted_context_extraction(document, keywords, vectorizer, tfidf_scores, top_n=5):
|
77 |
+
paragraphs = [para for para in document.split("\n") if para]
|
78 |
+
scores = [sum([para.lower().count(keyword) * tfidf_scores[vectorizer.vocabulary_[keyword]] for keyword in keywords]) for para in paragraphs]
|
79 |
+
|
80 |
+
top_indices = sorted(range(len(scores)), key=lambda i: scores[i], reverse=True)[:top_n]
|
81 |
+
relevant_paragraphs = [paragraphs[i] for i in top_indices]
|
82 |
+
|
83 |
+
return " ".join(relevant_paragraphs)
|
84 |
+
|
85 |
+
|
86 |
+
def extract_page_and_clause_references(paragraph: str) -> str:
|
87 |
+
page_matches = re.findall(r'Page (\d+)', paragraph)
|
88 |
+
clause_matches = re.findall(r'Clause (\d+\.\d+)', paragraph)
|
89 |
+
|
90 |
+
page_ref = f"Page {page_matches[0]}" if page_matches else ""
|
91 |
+
clause_ref = f"Clause {clause_matches[0]}" if clause_matches else ""
|
92 |
+
|
93 |
+
return f"({page_ref}, {clause_ref})".strip(", ")
|
94 |
+
|
95 |
+
def refine_answer_based_on_question(question: str, answer: str) -> str:
|
96 |
+
if "Does the agreement contain" in question:
|
97 |
+
if "not" in answer or "No" in answer:
|
98 |
+
refined_answer = f"No, the agreement does not contain {answer}"
|
99 |
+
else:
|
100 |
+
refined_answer = f"Yes, the agreement contains {answer}"
|
101 |
+
else:
|
102 |
+
refined_answer = answer
|
103 |
+
|
104 |
+
return refined_answer
|
105 |
+
|
106 |
+
def answer_query_with_context(question: str, df: pd.DataFrame, tokenizer, model, top_n_paragraphs: int = 5) -> str:
|
107 |
+
question_words = set(question.split())
|
108 |
+
|
109 |
+
priority_keywords = ["duration", "term", "period", "month", "year", "day", "week", "agreement", "obligation", "effective date"]
|
110 |
+
|
111 |
+
df['relevance_score'] = df['content'].apply(lambda x: len(question_words.intersection(set(x.split()))) + sum([x.lower().count(pk) for pk in priority_keywords]))
|
112 |
+
|
113 |
+
most_relevant_paragraphs = df.sort_values(by='relevance_score', ascending=False).iloc[:top_n_paragraphs]['content'].tolist()
|
114 |
+
|
115 |
+
context = "\n\n".join(most_relevant_paragraphs)
|
116 |
+
prompt = f"Question: {question}\n\nContext: {context}\n\nAnswer:"
|
117 |
+
|
118 |
+
inputs = tokenizer.encode(prompt, return_tensors="pt", max_length=512, truncation=True)
|
119 |
+
outputs = model.generate(inputs, max_length=200)
|
120 |
+
answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
121 |
+
|
122 |
+
references = extract_page_and_clause_references(context)
|
123 |
+
answer = refine_answer_based_on_question(question, answer) + " " + references
|
124 |
+
|
125 |
return answer
|
126 |
+
|
127 |
+
def get_embedding(text, tokenizer):
|
128 |
+
try:
|
129 |
+
inputs = tokenizer(text, return_tensors="pt", max_length=512, truncation=True)
|
130 |
+
outputs = model(**inputs)
|
131 |
+
embedding = outputs.last_hidden_state
|
132 |
+
except Exception as e:
|
133 |
+
print("Error obtaining embedding:", e)
|
134 |
+
embedding = []
|
135 |
+
return embedding
|