Spaces:
Sleeping
Sleeping
Create RAGSample.py
Browse files- src/RAGSample.py +388 -0
src/RAGSample.py
ADDED
@@ -0,0 +1,388 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
from dotenv import load_dotenv
|
3 |
+
from langchain_community.document_loaders import WebBaseLoader
|
4 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
5 |
+
from langchain_community.vectorstores import Chroma
|
6 |
+
from langchain_ollama import ChatOllama
|
7 |
+
from langchain.prompts import PromptTemplate
|
8 |
+
from langchain_core.output_parsers import StrOutputParser
|
9 |
+
from langchain_core.retrievers import BaseRetriever
|
10 |
+
from langchain_core.runnables import Runnable
|
11 |
+
from langchain_core.documents import Document
|
12 |
+
from langchain_core.embeddings import Embeddings
|
13 |
+
import chromadb
|
14 |
+
import numpy as np
|
15 |
+
from sklearn.feature_extraction.text import TfidfVectorizer
|
16 |
+
from sklearn.metrics.pairwise import cosine_similarity
|
17 |
+
import pandas as pd
|
18 |
+
from typing import Optional, List
|
19 |
+
import re
|
20 |
+
|
21 |
+
# Disable ChromaDB telemetry to avoid the error
|
22 |
+
os.environ["ANONYMIZED_TELEMETRY"] = "False"
|
23 |
+
os.environ["CHROMA_SERVER_HOST"] = "localhost"
|
24 |
+
os.environ["CHROMA_SERVER_HTTP_PORT"] = "8000"
|
25 |
+
|
26 |
+
|
27 |
+
class ImprovedTFIDFEmbeddings(Embeddings):
|
28 |
+
"""Improved TF-IDF based embedding function with better preprocessing."""
|
29 |
+
|
30 |
+
def __init__(self):
|
31 |
+
self.vectorizer = TfidfVectorizer(
|
32 |
+
max_features=5000,
|
33 |
+
stop_words='english',
|
34 |
+
ngram_range=(1, 3),
|
35 |
+
min_df=1,
|
36 |
+
max_df=0.85,
|
37 |
+
lowercase=True,
|
38 |
+
strip_accents='unicode',
|
39 |
+
analyzer='word'
|
40 |
+
)
|
41 |
+
self.fitted = False
|
42 |
+
self.documents = []
|
43 |
+
|
44 |
+
def embed_documents(self, texts):
|
45 |
+
"""Create embeddings for a list of texts."""
|
46 |
+
if not self.fitted:
|
47 |
+
self.documents = texts
|
48 |
+
self.vectorizer.fit(texts)
|
49 |
+
self.fitted = True
|
50 |
+
|
51 |
+
# Transform texts to TF-IDF vectors
|
52 |
+
tfidf_matrix = self.vectorizer.transform(texts)
|
53 |
+
|
54 |
+
# Convert to dense arrays and normalize
|
55 |
+
embeddings = []
|
56 |
+
for i in range(tfidf_matrix.shape[0]):
|
57 |
+
embedding = tfidf_matrix[i].toarray().flatten()
|
58 |
+
# Normalize the embedding
|
59 |
+
norm = np.linalg.norm(embedding)
|
60 |
+
if norm > 0:
|
61 |
+
embedding = embedding / norm
|
62 |
+
# Pad or truncate to 512 dimensions
|
63 |
+
if len(embedding) < 512:
|
64 |
+
embedding = np.pad(embedding, (0, 512 - len(embedding)))
|
65 |
+
else:
|
66 |
+
embedding = embedding[:512]
|
67 |
+
embeddings.append(embedding.tolist())
|
68 |
+
|
69 |
+
return embeddings
|
70 |
+
|
71 |
+
def embed_query(self, text):
|
72 |
+
"""Create embedding for a single query text."""
|
73 |
+
if not self.fitted:
|
74 |
+
# If not fitted, fit with just this text
|
75 |
+
self.vectorizer.fit([text])
|
76 |
+
self.fitted = True
|
77 |
+
|
78 |
+
# Transform query to TF-IDF vector
|
79 |
+
tfidf_matrix = self.vectorizer.transform([text])
|
80 |
+
embedding = tfidf_matrix[0].toarray().flatten()
|
81 |
+
|
82 |
+
# Normalize the embedding
|
83 |
+
norm = np.linalg.norm(embedding)
|
84 |
+
if norm > 0:
|
85 |
+
embedding = embedding / norm
|
86 |
+
|
87 |
+
# Pad or truncate to 512 dimensions
|
88 |
+
if len(embedding) < 512:
|
89 |
+
embedding = np.pad(embedding, (0, 512 - len(embedding)))
|
90 |
+
else:
|
91 |
+
embedding = embedding[:512]
|
92 |
+
|
93 |
+
return embedding.tolist()
|
94 |
+
|
95 |
+
|
96 |
+
class SmartFAQRetriever(BaseRetriever):
|
97 |
+
"""Smart retriever optimized for FAQ datasets with semantic similarity."""
|
98 |
+
|
99 |
+
def __init__(self, documents: List[Document], k: int = 4):
|
100 |
+
super().__init__()
|
101 |
+
self._documents = documents
|
102 |
+
self._k = k
|
103 |
+
self._vectorizer = None # Use private attribute
|
104 |
+
|
105 |
+
@property
|
106 |
+
def documents(self):
|
107 |
+
return self._documents
|
108 |
+
|
109 |
+
@property
|
110 |
+
def k(self):
|
111 |
+
return self._k
|
112 |
+
|
113 |
+
def _get_relevant_documents(self, query: str) -> List[Document]:
|
114 |
+
"""Retrieve documents based on semantic similarity."""
|
115 |
+
# Ensure vectorizer is fitted
|
116 |
+
if not hasattr(self, '_vectorizer') or self._vectorizer is None or not hasattr(self._vectorizer, 'vocabulary_') or not self._vectorizer.vocabulary_:
|
117 |
+
print("[SmartFAQRetriever] Fitting vectorizer...")
|
118 |
+
self._vectorizer = TfidfVectorizer(
|
119 |
+
max_features=3000,
|
120 |
+
stop_words='english',
|
121 |
+
ngram_range=(1, 2),
|
122 |
+
min_df=1,
|
123 |
+
max_df=0.9
|
124 |
+
)
|
125 |
+
questions = []
|
126 |
+
for doc in self._documents:
|
127 |
+
if "QUESTION:" in doc.page_content:
|
128 |
+
question_part = doc.page_content.split("ANSWER:")[0]
|
129 |
+
question = question_part.replace("QUESTION:", "").strip()
|
130 |
+
questions.append(question)
|
131 |
+
else:
|
132 |
+
questions.append(doc.page_content)
|
133 |
+
self._vectorizer.fit(questions)
|
134 |
+
query_lower = query.lower().strip()
|
135 |
+
|
136 |
+
# Extract questions from documents
|
137 |
+
questions = []
|
138 |
+
for doc in self._documents:
|
139 |
+
if "QUESTION:" in doc.page_content:
|
140 |
+
question_part = doc.page_content.split("ANSWER:")[0]
|
141 |
+
question = question_part.replace("QUESTION:", "").strip()
|
142 |
+
questions.append(question)
|
143 |
+
else:
|
144 |
+
questions.append(doc.page_content)
|
145 |
+
|
146 |
+
# Transform query and questions to TF-IDF vectors
|
147 |
+
query_vector = self._vectorizer.transform([query_lower])
|
148 |
+
question_vectors = self._vectorizer.transform(questions)
|
149 |
+
|
150 |
+
# Calculate cosine similarities
|
151 |
+
similarities = cosine_similarity(query_vector, question_vectors).flatten()
|
152 |
+
|
153 |
+
# Get top k documents
|
154 |
+
top_indices = similarities.argsort()[-self._k:][::-1]
|
155 |
+
|
156 |
+
# Return documents with highest similarity scores
|
157 |
+
relevant_docs = [self._documents[i] for i in top_indices if similarities[i] > 0.1]
|
158 |
+
|
159 |
+
if not relevant_docs:
|
160 |
+
# Fallback to first k documents if no good matches
|
161 |
+
relevant_docs = self._documents[:self._k]
|
162 |
+
|
163 |
+
return relevant_docs
|
164 |
+
|
165 |
+
async def _aget_relevant_documents(self, query: str) -> List[Document]:
|
166 |
+
"""Async version of get_relevant_documents."""
|
167 |
+
return self._get_relevant_documents(query)
|
168 |
+
|
169 |
+
def setup_retriever(use_kaggle_data: bool = False, kaggle_dataset: Optional[str] = None,
|
170 |
+
kaggle_username: Optional[str] = None, kaggle_key: Optional[str] = None,
|
171 |
+
use_local_mental_health_data: bool = False) -> BaseRetriever:
|
172 |
+
"""
|
173 |
+
Creates a vector store with documents from test data, Kaggle datasets, or local mental health data.
|
174 |
+
|
175 |
+
Args:
|
176 |
+
use_kaggle_data: Whether to load Kaggle data instead of test documents
|
177 |
+
kaggle_dataset: Kaggle dataset name (e.g., 'username/dataset-name')
|
178 |
+
kaggle_username: Your Kaggle username (optional if using kaggle.json)
|
179 |
+
kaggle_key: Your Kaggle API key (optional if using kaggle.json)
|
180 |
+
use_local_mental_health_data: Whether to load local mental health FAQ data
|
181 |
+
"""
|
182 |
+
print("Setting up the retriever...")
|
183 |
+
|
184 |
+
if use_local_mental_health_data:
|
185 |
+
try:
|
186 |
+
print("Loading mental health FAQ data from local file...")
|
187 |
+
mental_health_file = "data/Mental_Health_FAQ.csv"
|
188 |
+
|
189 |
+
if not os.path.exists(mental_health_file):
|
190 |
+
print(f"Mental health FAQ file not found: {mental_health_file}")
|
191 |
+
use_local_mental_health_data = False
|
192 |
+
else:
|
193 |
+
# Load mental health FAQ data
|
194 |
+
df = pd.read_csv(mental_health_file)
|
195 |
+
documents = []
|
196 |
+
|
197 |
+
for _, row in df.iterrows():
|
198 |
+
question = row['Questions']
|
199 |
+
answer = row['Answers']
|
200 |
+
# Create document in FAQ format
|
201 |
+
content = f"QUESTION: {question}\nANSWER: {answer}"
|
202 |
+
documents.append(Document(page_content=content))
|
203 |
+
|
204 |
+
print(f"Loaded {len(documents)} mental health FAQ documents")
|
205 |
+
for i, doc in enumerate(documents[:3]):
|
206 |
+
print(f"Sample FAQ {i+1}: {doc.page_content[:200]}...")
|
207 |
+
|
208 |
+
except Exception as e:
|
209 |
+
print(f"Error loading mental health data: {e}")
|
210 |
+
use_local_mental_health_data = False
|
211 |
+
|
212 |
+
if use_kaggle_data and kaggle_dataset:
|
213 |
+
try:
|
214 |
+
from src.kaggle_loader import KaggleDataLoader
|
215 |
+
|
216 |
+
print(f"Loading Kaggle dataset: {kaggle_dataset}")
|
217 |
+
# Create loader without parameters - it will auto-load from kaggle.json
|
218 |
+
loader = KaggleDataLoader()
|
219 |
+
|
220 |
+
# Download the dataset
|
221 |
+
dataset_path = loader.download_dataset(kaggle_dataset)
|
222 |
+
|
223 |
+
# Load documents based on file type - only process files from this specific dataset
|
224 |
+
documents = []
|
225 |
+
|
226 |
+
# Get the dataset name to identify the correct files
|
227 |
+
dataset_name = kaggle_dataset.split('/')[-1]
|
228 |
+
print(f"Processing files in dataset directory: {dataset_path}")
|
229 |
+
|
230 |
+
for file in os.listdir(dataset_path):
|
231 |
+
file_path = os.path.join(dataset_path, file)
|
232 |
+
|
233 |
+
if file.endswith('.csv'):
|
234 |
+
print(f"Loading CSV file: {file}")
|
235 |
+
# For FAQ datasets, use the improved loading method
|
236 |
+
if 'faq' in file.lower() or 'mental' in file.lower():
|
237 |
+
documents.extend(loader.load_csv_dataset(file_path, [], chunk_size=50))
|
238 |
+
else:
|
239 |
+
# For other CSV files, use first few columns as text
|
240 |
+
df = pd.read_csv(file_path)
|
241 |
+
text_columns = df.columns[:3].tolist() # Use first 3 columns
|
242 |
+
documents.extend(loader.load_csv_dataset(file_path, text_columns, chunk_size=50))
|
243 |
+
|
244 |
+
elif file.endswith('.json'):
|
245 |
+
print(f"Loading JSON file: {file}")
|
246 |
+
documents.extend(loader.load_json_dataset(file_path))
|
247 |
+
|
248 |
+
elif file.endswith('.txt'):
|
249 |
+
print(f"Loading text file: {file}")
|
250 |
+
documents.extend(loader.load_text_dataset(file_path))
|
251 |
+
|
252 |
+
print(f"Loaded {len(documents)} documents from Kaggle dataset")
|
253 |
+
for i, doc in enumerate(documents[:3]):
|
254 |
+
print(f"Sample doc {i+1}: {doc.page_content[:200]}")
|
255 |
+
|
256 |
+
except Exception as e:
|
257 |
+
print(f"Error loading Kaggle data: {e}")
|
258 |
+
print("Falling back to test documents...")
|
259 |
+
use_kaggle_data = False
|
260 |
+
|
261 |
+
if not use_kaggle_data and not use_local_mental_health_data:
|
262 |
+
# No test documents - use mental health data as default
|
263 |
+
print("No specific data source specified, loading mental health FAQ data as default...")
|
264 |
+
try:
|
265 |
+
mental_health_file = "data/Mental_Health_FAQ.csv"
|
266 |
+
|
267 |
+
if not os.path.exists(mental_health_file):
|
268 |
+
raise FileNotFoundError(f"Mental health FAQ file not found: {mental_health_file}")
|
269 |
+
|
270 |
+
# Load mental health FAQ data
|
271 |
+
df = pd.read_csv(mental_health_file)
|
272 |
+
documents = []
|
273 |
+
|
274 |
+
for _, row in df.iterrows():
|
275 |
+
question = row['Questions']
|
276 |
+
answer = row['Answers']
|
277 |
+
# Create document in FAQ format
|
278 |
+
content = f"QUESTION: {question}\nANSWER: {answer}"
|
279 |
+
documents.append(Document(page_content=content))
|
280 |
+
|
281 |
+
print(f"Loaded {len(documents)} mental health FAQ documents")
|
282 |
+
for i, doc in enumerate(documents[:3]):
|
283 |
+
print(f"Sample FAQ {i+1}: {doc.page_content[:200]}...")
|
284 |
+
|
285 |
+
except Exception as e:
|
286 |
+
print(f"Error loading mental health data: {e}")
|
287 |
+
raise Exception("No valid data source available. Please ensure mental health FAQ data is present or provide Kaggle credentials.")
|
288 |
+
|
289 |
+
print("Creating TF-IDF embeddings...")
|
290 |
+
embeddings = ImprovedTFIDFEmbeddings()
|
291 |
+
|
292 |
+
print("Creating ChromaDB vector store...")
|
293 |
+
client = chromadb.PersistentClient(path="./src/chroma_db")
|
294 |
+
|
295 |
+
# Clear existing collections to prevent mixing old and new data
|
296 |
+
try:
|
297 |
+
collections = client.list_collections()
|
298 |
+
for collection in collections:
|
299 |
+
print(f"Deleting existing collection: {collection.name}")
|
300 |
+
client.delete_collection(collection.name)
|
301 |
+
except Exception as e:
|
302 |
+
print(f"Warning: Could not clear existing collections: {e}")
|
303 |
+
|
304 |
+
print(f"Processing {len(documents)} documents...")
|
305 |
+
|
306 |
+
# Check if this is a FAQ dataset and use smart retriever
|
307 |
+
if any("QUESTION:" in doc.page_content for doc in documents):
|
308 |
+
print("Using SmartFAQRetriever for better semantic matching...")
|
309 |
+
return SmartFAQRetriever(documents, k=4)
|
310 |
+
else:
|
311 |
+
# Use vector store for non-FAQ datasets
|
312 |
+
vectorstore = Chroma.from_documents(
|
313 |
+
documents=documents,
|
314 |
+
embedding=embeddings,
|
315 |
+
client=client
|
316 |
+
)
|
317 |
+
print("Retriever setup complete.")
|
318 |
+
return vectorstore.as_retriever(k=4)
|
319 |
+
|
320 |
+
def setup_rag_chain() -> Runnable:
|
321 |
+
"""Sets up the RAG chain with a prompt template and an LLM."""
|
322 |
+
# Define the prompt template for the LLM
|
323 |
+
prompt = PromptTemplate(
|
324 |
+
template="""You are an assistant for question-answering tasks.
|
325 |
+
Use the following documents to answer the question.
|
326 |
+
If you don't know the answer, just say that you don't know.
|
327 |
+
Use three sentences maximum and keep the answer concise:
|
328 |
+
Question: {question}
|
329 |
+
Documents: {documents}
|
330 |
+
Answer:
|
331 |
+
""",
|
332 |
+
input_variables=["question", "documents"],
|
333 |
+
)
|
334 |
+
|
335 |
+
# Initialize the LLM with dolphin-llama3:8b model
|
336 |
+
# Note: This requires the Ollama server to be running with the specified model
|
337 |
+
llm = ChatOllama(
|
338 |
+
model="dolphin-llama3:8b",
|
339 |
+
temperature=0,
|
340 |
+
)
|
341 |
+
|
342 |
+
# Create a chain combining the prompt template and LLM
|
343 |
+
return prompt | llm | StrOutputParser()
|
344 |
+
|
345 |
+
|
346 |
+
# Define the RAG application class
|
347 |
+
class RAGApplication:
|
348 |
+
def __init__(self, retriever: BaseRetriever, rag_chain: Runnable):
|
349 |
+
self.retriever = retriever
|
350 |
+
self.rag_chain = rag_chain
|
351 |
+
|
352 |
+
def run(self, question: str) -> str:
|
353 |
+
"""Runs the RAG pipeline for a given question."""
|
354 |
+
# Retrieve relevant documents
|
355 |
+
documents = self.retriever.invoke(question)
|
356 |
+
|
357 |
+
# Debug: Print retrieved documents
|
358 |
+
print(f"\nDEBUG: Retrieved {len(documents)} documents for question: '{question}'")
|
359 |
+
for i, doc in enumerate(documents):
|
360 |
+
print(f"DEBUG: Document {i+1}: {doc.page_content[:200]}...")
|
361 |
+
|
362 |
+
# Extract content from retrieved documents
|
363 |
+
doc_texts = "\n\n".join([doc.page_content for doc in documents])
|
364 |
+
|
365 |
+
# Debug: Print the combined document text
|
366 |
+
print(f"DEBUG: Combined document text: {doc_texts[:300]}...")
|
367 |
+
|
368 |
+
# Get the answer from the language model
|
369 |
+
answer = self.rag_chain.invoke({"question": question, "documents": doc_texts})
|
370 |
+
return answer
|
371 |
+
|
372 |
+
# Main execution block
|
373 |
+
if __name__ == "__main__":
|
374 |
+
load_dotenv()
|
375 |
+
|
376 |
+
# 1. Setup the components
|
377 |
+
retriever = setup_retriever()
|
378 |
+
rag_chain = setup_rag_chain()
|
379 |
+
|
380 |
+
# 2. Initialize the RAG application
|
381 |
+
rag_application = RAGApplication(retriever, rag_chain)
|
382 |
+
|
383 |
+
# 3. Run an example query
|
384 |
+
question = "What is prompt engineering"
|
385 |
+
print("\n--- Running RAG Application ---")
|
386 |
+
print(f"Question: {question}")
|
387 |
+
answer = rag_application.run(question)
|
388 |
+
print(f"Answer: {answer}")
|