File size: 8,931 Bytes
43add07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
69a10b1
 
 
 
43add07
 
 
 
188bee1
43add07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
69a10b1
 
 
 
 
 
43add07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
69a10b1
43add07
 
69a10b1
43add07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
fb160c8
 
 
 
43add07
 
 
 
 
 
 
fb160c8
 
43add07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
69a10b1
 
 
 
 
 
43add07
69a10b1
 
 
 
 
 
 
 
 
 
 
43add07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
69a10b1
43add07
 
 
 
 
 
 
 
 
69a10b1
43add07
fb160c8
 
43add07
 
fb160c8
 
43add07
fb160c8
 
43add07
fb160c8
 
43add07
fb160c8
 
 
 
 
43add07
 
 
fb160c8
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
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
# utils.py
"""
Financial Chatbot Utilities
Core functionality for RAG-based financial chatbot
"""

import os
import re
import nltk
from nltk.corpus import stopwords
from collections import deque
from typing import Tuple
import torch

# LangChain components
from langchain_community.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.vectorstores import Chroma
from langchain_huggingface import HuggingFaceEmbeddings

# Models and ML
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
from rank_bm25 import BM25Okapi
from sentence_transformers import CrossEncoder
from sklearn.metrics.pairwise import cosine_similarity

# Initialize NLTK stopwords
# nltk.download('stopwords')
# stop_words = set(stopwords.words('english'))
nltk.data.path.append('./nltk_data')  # Point to local NLTK data
stop_words = set(nltk.corpus.stopwords.words('english'))

# mount
import sys
sys.path.append('/mount/src/gen_ai_dev')

# Configuration
DATA_PATH = "./Infy financial report/"
DATA_FILES = ["INFY_2022_2023.pdf", "INFY_2023_2024.pdf"]
EMBEDDING_MODEL = "sentence-transformers/all-MiniLM-L6-v2"
LLM_MODEL = "gpt2" # "HuggingFaceH4/zephyr-7b-beta"  # "microsoft/phi-2"

# Environment settings
os.environ["TOKENIZERS_PARALLELISM"] = "false"
os.environ["CHROMA_DISABLE_TELEMETRY"] = "true"

# Suppress specific warnings
import warnings

warnings.filterwarnings("ignore", message=".*oneDNN custom operations.*")
warnings.filterwarnings("ignore", message=".*cuBLAS factory.*")


# ------------------------------
# Load and Chunk Documents
# ------------------------------
def load_and_chunk_documents():
    """Load and split PDF documents into manageable chunks"""
    text_splitter = RecursiveCharacterTextSplitter(
        chunk_size=500,
        chunk_overlap=100,
        separators=["\n\n", "\n", ".", " ", ""]
    )

    all_chunks = []
    for file in DATA_FILES:
        try:
            loader = PyPDFLoader(os.path.join(DATA_PATH, file))
            pages = loader.load()
            all_chunks.extend(text_splitter.split_documents(pages))
        except Exception as e:
            print(f"Error loading {file}: {e}")

    return all_chunks


# ------------------------------
# Vector Store and Search Setup
# ------------------------------
text_chunks = load_and_chunk_documents()
embeddings = HuggingFaceEmbeddings(model_name=EMBEDDING_MODEL)

vector_db = Chroma.from_documents(
    documents=text_chunks,
    embedding=embeddings,
    persist_directory="./chroma_db"
)
vector_db.persist()

# BM25 setup
bm25_corpus = [chunk.page_content for chunk in text_chunks]
bm25_tokenized = [doc.split() for doc in bm25_corpus]
bm25 = BM25Okapi(bm25_tokenized)

# Cross-encoder for re-ranking
cross_encoder = CrossEncoder('cross-encoder/ms-marco-MiniLM-L-6-v2')


# ------------------------------
# Conversation Memory
# ------------------------------
class ConversationMemory:
    """Stores recent conversation context"""

    def __init__(self, max_size=5):
        self.buffer = deque(maxlen=max_size)

    def add_interaction(self, query: str, response: str) -> None:
        self.buffer.append((query, response))

    def get_context(self) -> str:
        return "\n".join(
            [f"Previous Q: {q}\nPrevious A: {r}" for q, r in self.buffer]
        )


memory = ConversationMemory(max_size=3)


# ------------------------------
# Hybrid Retrieval System
# ------------------------------
def hybrid_retrieval(query: str, top_k: int = 5) -> str:
    try:
        # Semantic search
        semantic_results = vector_db.similarity_search(query, k=top_k * 2)
        print(f"\n\n[For Debug Only] Semantic Results: {semantic_results}")

        # Keyword search
        keyword_results = bm25.get_top_n(query.split(), bm25_corpus, n=top_k * 2)
        print(f"\n\n[For Debug Only] Keyword Results: {keyword_results}\n\n")

        # Combine and deduplicate results
        combined = []
        seen = set()

        for doc in semantic_results:
            content = doc.page_content
            if content not in seen:
                combined.append((content, "semantic"))
                seen.add(content)

        for doc in keyword_results:
            if doc not in seen:
                combined.append((doc, "keyword"))
                seen.add(doc)

        # Re-rank results using cross-encoder
        pairs = [(query, content) for content, _ in combined]
        scores = cross_encoder.predict(pairs)

        # Sort by scores
        sorted_results = sorted(
            zip(combined, scores),
            key=lambda x: x[1],
            reverse=True
        )

        final_results = [f"[{source}] {content}" for (content, source), _ in sorted_results[:top_k]]

        memory_context = memory.get_context()
        if memory_context:
            final_results.append(f"[memory] {memory_context}")

        return "\n\n".join(final_results)

    except Exception as e:
        print(f"Retrieval error: {e}")
        return ""


# ------------------------------
# Safety Guardrails
# ------------------------------
class SafetyGuard:
    """Validates input and filters output"""

    def __init__(self):
        # self.financial_terms = {
        #     'revenue', 'profit', 'ebitda', 'balance', 'cash',
        #     'income', 'fiscal', 'growth', 'margin', 'expense'
        # }
        self.blocked_topics = {
            'politics', 'sports', 'entertainment', 'religion',
            'medical', 'hypothetical', 'opinion', 'personal'
        }

    def validate_input(self, query: str) -> Tuple[bool, str]:
        query_lower = query.lower()
        # if not any(term in query_lower for term in self.financial_terms):
        #     return False, "Please ask financial questions."
        if any(topic in query_lower for topic in self.blocked_topics):
            return False, "I only discuss financial topics."
        return True, ""

    def filter_output(self, response: str) -> str:
        phrases_to_remove = {
            "I'm not sure", "I don't know", "maybe",
            "possibly", "could be", "uncertain", "perhaps"
        }
        for phrase in phrases_to_remove:
            response = response.replace(phrase, "")

        sentences = re.split(r'[.!?]', response)
        if len(sentences) > 2:
            response = '. '.join(sentences[:2]) + '.'

        return response.strip()


guard = SafetyGuard()

# ------------------------------
# LLM Initialization
# ------------------------------
try:
    tokenizer = AutoTokenizer.from_pretrained(LLM_MODEL)
    model = AutoModelForCausalLM.from_pretrained(
        LLM_MODEL,
        device_map="cpu",
        torch_dtype=torch.float32
    )

    generator = pipeline(
        "text-generation",
        model=model,
        tokenizer=tokenizer,
        max_new_tokens=400,
        do_sample=True,
        temperature=0.3,
        top_k=30,
        top_p=0.9,
        repetition_penalty=1.2
    )
except Exception as e:
    print(f"Error loading model: {e}")
    raise


# ------------------------------
# Response Generation
# ------------------------------
def extract_final_response(full_response: str) -> str:
    parts = full_response.split("<|im_start|>assistant")
    if len(parts) > 1:
        response = parts[-1].split("<|im_end|>")[0]
        return re.sub(r'\s+', ' ', response).strip()
    return full_response


def generate_answer(query: str) -> Tuple[str, float]:
    try:
        is_valid, msg = guard.validate_input(query)
        if not is_valid:
            return msg, 0.0

        context = hybrid_retrieval(query)
        vector_db.persist()

        prompt = f"""<|im_start|>system
You are a financial analyst. Provide a brief answer using the context.
Context: {context}<|im_end|>
<|im_start|>user
{query}<|im_end|>
<|im_start|>assistant
Answer:"""

        print(f"\n\n[For Debug Only] Prompt: {prompt}\n\n")
        response = generator(prompt)[0]['generated_text']
        print(f"\n\n[For Debug Only] response: {response}\n\n")
        
        clean_response = extract_final_response(response)
        clean_response = guard.filter_output(clean_response)
        print(f"\n\n[For Debug Only] clean_response: {clean_response}\n\n")
        
        query_embed = embeddings.embed_query(query)
        print(f"\n\n[For Debug Only] query_embed: {query_embed}\n\n")
        
        response_embed = embeddings.embed_query(clean_response)
        print(f"\n\n[For Debug Only] response_embed: {response_embed}\n\n")
        
        confidence = cosine_similarity([query_embed], [response_embed])[0][0]
        print(f"\n\n[For Debug Only] confidence: {confidence}\n\n")
        
        memory.add_interaction(query, clean_response)       
        
        print(f"\n\n[For Debug Only] I'm Done \n\n")
        return clean_response, round(confidence, 2)

    except Exception as e:
        return f"Error processing request: {e}", 0.0