File size: 5,397 Bytes
ce77033
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
import json
import os
from dotenv import load_dotenv
from openai import OpenAI
from tqdm import tqdm
from scripts.regulatory_change_foundation import CONTEXT_CATEGORIES
from scripts.utility_functions import call_nlp_service, render_prompt


# Load environment variables from .env file
load_dotenv()

api_key = os.getenv("OPENAI_API_KEY")
openai_client = OpenAI(api_key=api_key)


def preprocess_text_with_nlp(text, max_chunk_size=512, overlap=50):
    """Enhanced NLP preprocessing identical to your first experiment using PyMuPDF text extraction"""
    return call_nlp_service({"text": text}, "preprocess_text_with_nlp_pymupdf")


def create_prompt_with_nlp(chunk, preprocessed_data):
    return render_prompt(chunk, include_nlp=True, preprocessed_data=preprocessed_data)


def classify_changes_with_nlp(text_content, location_info):
    """Classify changes with NLP preprocessing."""
    # Apply NLP preprocessing
    preprocessed_data = preprocess_text_with_nlp(text_content)

    # Split into chunks (using the same method as your first experiment)
    result = call_nlp_service({"text": text_content}, "recursive_character_text_splitter")
    chunks = result["chunks"]

    results = []
    for chunk in chunks:
        response = openai_client.chat.completions.create(
            model="gpt-4o-mini",
            messages=[
                {
                    "role": "system",
                    "content": "You are a legal expert analyzing German regulatory changes. Return only JSON.",
                },
                {
                    "role": "user",
                    "content": create_prompt_with_nlp(chunk, preprocessed_data),
                },
            ],
            temperature=0.7,
            max_tokens=1024,
        )

        try:
            result = json.loads(response.choices[0].message.content)
            if result.get("changes_detected", False):
                result["location"] = location_info
                result["source_text"] = chunk
                results.append(result)
        except json.JSONDecodeError:
            continue

    return results if results else None


def extract_hierarchical_text(block):
    """Extract text from a block including its parent and grandparent contexts."""
    text_parts = []

    # Check if block has a grandparent
    if (
        "parent" in block
        and block["parent"] is not None
        and "parent" in block["parent"]
        and block["parent"]["parent"] is not None
    ):
        text_parts.append(block["parent"]["parent"]["text"])

    # Check if block has a parent
    if "parent" in block and block["parent"] is not None:
        text_parts.append(block["parent"]["text"])

    # Add the current block's text
    text_parts.append(block["text"])

    # Join all text parts with newlines between them
    return "\n\n".join(text_parts)


def traverse_blocks_with_nlp(blocks, parent=None, results=None, is_top_level=True):
    """Traverse hierarchy with NLP-enhanced analysis."""
    if results is None:
        results = []

    iterable = (
        tqdm(blocks, desc="Processing Text blocks with NLP") if is_top_level else blocks
    )

    for block in iterable:
        block["parent"] = parent

        if "children" in block and not block["children"]:  # Leaf node
            text_content = extract_hierarchical_text(block)
            location_info = {
                "page_number": block["page_number"],
                "block_text": block["text"],
            }

            changes = classify_changes_with_nlp(text_content, location_info)
            if changes:
                for change in changes:
                    change["full_text"] = text_content
                    results.append(change)
        else:
            traverse_blocks_with_nlp(
                block["children"], block, results, is_top_level=False
            )

    return results


def pymupdf_regulatory_change_detector_with_nlp_insights(hierarchical_structure):
    """Main function with NLP integration."""
    if not hierarchical_structure:
        return {"error": "No structure provided"}, []

    analysis_summary = {
        "total_changes_detected": 0,
        "changes_by_type": {"addition": 0, "deletion": 0, "modification": 0},
    }
    changes_by_page = {}

    results = traverse_blocks_with_nlp(hierarchical_structure["blocks"])

    for change in results:
        analysis_summary["total_changes_detected"] += len(change["classifications"])
        for classification in change["classifications"]:
            analysis_summary["changes_by_type"][classification["change_type"]] += 1

            change_subtype = (
                "context" if classification["change"] in CONTEXT_CATEGORIES else "scope"
            )
            page_num = change["location"]["page_number"]
            changes_by_page.setdefault(page_num, []).append(
                {
                    "change": classification["change"],
                    "change_type": classification["change_type"],
                    "change_subtype": change_subtype,
                    "relevant_text": classification["relevant_text"],
                    "explanation": classification["explanation"],
                    "nlp_evidence": classification["evidence"],
                }
            )

    return {
        "analysis_summary": analysis_summary,
        "changes_by_page": changes_by_page,
    }, results