Update BanglaRAG/bangla_rag_pipeline.py
Browse files- BanglaRAG/bangla_rag_pipeline.py +20 -98
BanglaRAG/bangla_rag_pipeline.py
CHANGED
@@ -4,7 +4,6 @@ from transformers import (
|
|
4 |
AutoTokenizer,
|
5 |
AutoModelForCausalLM,
|
6 |
pipeline,
|
7 |
-
GenerationConfig,
|
8 |
BitsAndBytesConfig,
|
9 |
)
|
10 |
from langchain_core.prompts import PromptTemplate
|
@@ -14,25 +13,12 @@ from langchain_community.vectorstores import Chroma
|
|
14 |
from langchain_community.llms.huggingface_pipeline import HuggingFacePipeline
|
15 |
from langchain_core.runnables import RunnableParallel, RunnablePassthrough
|
16 |
from langchain_core.output_parsers import StrOutputParser
|
17 |
-
from langchain_core.prompts import ChatPromptTemplate, HumanMessagePromptTemplate
|
18 |
-
from rich import print as rprint
|
19 |
-
from rich.panel import Panel
|
20 |
-
from tqdm import tqdm
|
21 |
import warnings
|
22 |
-
import re
|
23 |
|
24 |
warnings.filterwarnings("ignore")
|
25 |
|
26 |
class BanglaRAGChain:
|
27 |
-
"""
|
28 |
-
Bangla Retrieval-Augmented Generation (RAG) Chain for question answering.
|
29 |
-
This class uses a HuggingFace/local language model for text generation, a Chroma vector database for
|
30 |
-
document retrieval, and a custom prompt template to create a RAG chain that can generate
|
31 |
-
responses to user queries in Bengali.
|
32 |
-
"""
|
33 |
-
|
34 |
def __init__(self):
|
35 |
-
"""Initializes the BanglaRAGChain with default parameters."""
|
36 |
self._device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
37 |
self.chat_model_id = None
|
38 |
self.embed_model_id = None
|
@@ -71,22 +57,6 @@ class BanglaRAGChain:
|
|
71 |
chunk_overlap=150,
|
72 |
hf_token=None,
|
73 |
):
|
74 |
-
"""
|
75 |
-
Loads the required models and data for the RAG chain.
|
76 |
-
Args:
|
77 |
-
chat_model_id (str): The Hugging Face model ID for the chat model.
|
78 |
-
embed_model_id (str): The Hugging Face model ID for the embedding model.
|
79 |
-
text_path (str): Path to the text file to be indexed.
|
80 |
-
quantization (bool): Whether to quantize the model or not.
|
81 |
-
k (int): The number of documents to retrieve.
|
82 |
-
top_k (int): The top_k parameter for the generation configuration.
|
83 |
-
top_p (float): The top_p parameter for the generation configuration.
|
84 |
-
max_new_tokens (int): The maximum number of new tokens to generate.
|
85 |
-
temperature (float): The temperature parameter for the generation configuration.
|
86 |
-
chunk_size (int): The chunk size for text splitting.
|
87 |
-
chunk_overlap (int): The chunk overlap for text splitting.
|
88 |
-
hf_token (str): The Hugging Face token for authentication.
|
89 |
-
"""
|
90 |
self.chat_model_id = chat_model_id
|
91 |
self.embed_model_id = embed_model_id
|
92 |
self.k = k
|
@@ -103,26 +73,14 @@ class BanglaRAGChain:
|
|
103 |
if self.hf_token is not None:
|
104 |
os.environ["HF_TOKEN"] = str(self.hf_token)
|
105 |
|
106 |
-
rprint(Panel("[bold green]Loading chat models...", expand=False))
|
107 |
self._load_models()
|
108 |
-
|
109 |
-
rprint(Panel("[bold green]Creating document...", expand=False))
|
110 |
self._create_document()
|
111 |
-
|
112 |
-
rprint(Panel("[bold green]Updating Chroma database...", expand=False))
|
113 |
self._update_chroma_db()
|
114 |
-
|
115 |
-
rprint(Panel("[bold green]Initializing retriever...", expand=False))
|
116 |
self._get_retriever()
|
117 |
-
|
118 |
-
rprint(Panel("[bold green]Initializing LLM...", expand=False))
|
119 |
self._get_llm()
|
120 |
-
|
121 |
-
rprint(Panel("[bold green]Creating chain...", expand=False))
|
122 |
self._create_chain()
|
123 |
|
124 |
def _load_models(self):
|
125 |
-
"""Loads the chat model and tokenizer."""
|
126 |
try:
|
127 |
self.tokenizer = AutoTokenizer.from_pretrained(self.chat_model_id)
|
128 |
bnb_config = None
|
@@ -133,28 +91,23 @@ class BanglaRAGChain:
|
|
133 |
bnb_4bit_quant_type="nf4",
|
134 |
bnb_4bit_compute_dtype=torch.float16,
|
135 |
)
|
136 |
-
rprint(Panel("[bold green]Applying 4bit quantization...", expand=False))
|
137 |
self.chat_model = AutoModelForCausalLM.from_pretrained(
|
138 |
self.chat_model_id,
|
139 |
-
|
140 |
-
|
141 |
-
quantization_config=bnb_config,
|
142 |
device_map="auto",
|
|
|
143 |
)
|
144 |
-
rprint(Panel("[bold green]Applied 4bit quantization successfully", expand=False))
|
145 |
else:
|
146 |
self.chat_model = AutoModelForCausalLM.from_pretrained(
|
147 |
self.chat_model_id,
|
148 |
-
torch_dtype=torch.
|
149 |
-
low_cpu_mem_usage=True,
|
150 |
device_map="auto",
|
151 |
)
|
152 |
-
rprint(Panel("[bold green]Chat Model loaded successfully!", expand=False))
|
153 |
except Exception as e:
|
154 |
-
|
155 |
|
156 |
def _create_document(self):
|
157 |
-
"""Splits the input text into chunks using RecursiveCharacterTextSplitter."""
|
158 |
try:
|
159 |
with open(self.text_path, "r", encoding="utf-8") as file:
|
160 |
self._text_content = file.read()
|
@@ -163,44 +116,21 @@ class BanglaRAGChain:
|
|
163 |
chunk_size=self.chunk_size,
|
164 |
chunk_overlap=self.chunk_overlap,
|
165 |
)
|
166 |
-
self._documents =
|
167 |
-
tqdm(
|
168 |
-
character_splitter.split_text(self._text_content),
|
169 |
-
desc="Chunking text",
|
170 |
-
)
|
171 |
-
)
|
172 |
-
print(f"Number of chunks: {len(self._documents)}")
|
173 |
-
if False:
|
174 |
-
for i, chunk in enumerate(self._documents):
|
175 |
-
if i > 5:
|
176 |
-
break
|
177 |
-
print(f"Chunk {i}: {chunk}")
|
178 |
-
rprint(Panel("[bold green]Document created successfully!", expand=False))
|
179 |
except Exception as e:
|
180 |
-
|
181 |
|
182 |
def _update_chroma_db(self):
|
183 |
-
"""Updates the Chroma vector database with the text chunks."""
|
184 |
try:
|
185 |
-
|
186 |
-
|
187 |
-
|
188 |
-
embeddings = HuggingFaceEmbeddings(
|
189 |
-
model_name=self.embed_model_id, model_kwargs=model_kwargs
|
190 |
-
)
|
191 |
-
rprint(Panel(f"[bold green]Loaded embedding model successfully!", expand=False))
|
192 |
-
except Exception as e:
|
193 |
-
rprint(Panel(f"[red]embedding model loading failed: {e}", expand=False))
|
194 |
-
|
195 |
-
self._db = Chroma.from_texts(texts=self._documents, embedding=embeddings)
|
196 |
-
rprint(
|
197 |
-
Panel("[bold green]Chroma database updated successfully!", expand=False)
|
198 |
)
|
|
|
199 |
except Exception as e:
|
200 |
-
|
201 |
|
202 |
def _create_chain(self):
|
203 |
-
"""Creates the retrieval-augmented generation (RAG) chain."""
|
204 |
template = """Below is an instruction in Bengali language that describes a task, paired with an input also in Bengali language that provides further context. Write a response in Bengali that appropriately completes the request.
|
205 |
### Instruction:
|
206 |
{question}
|
@@ -242,22 +172,18 @@ class BanglaRAGChain:
|
|
242 |
).assign(answer=rag_chain_from_docs)
|
243 |
|
244 |
self._chain = rag_chain_with_source
|
245 |
-
rprint(Panel("[bold green]Chain created successfully!", expand=False))
|
246 |
except Exception as e:
|
247 |
-
|
248 |
|
249 |
def _get_retriever(self):
|
250 |
-
"""Creates a retriever for the vector database."""
|
251 |
try:
|
252 |
self._retriever = self._db.as_retriever(
|
253 |
search_type="similarity", search_kwargs={"k": self.k}
|
254 |
)
|
255 |
-
rprint(Panel("[bold green]Retriever created successfully!", expand=False))
|
256 |
except Exception as e:
|
257 |
-
|
258 |
|
259 |
def _get_llm(self):
|
260 |
-
"""Initializes the language model using the Hugging Face pipeline."""
|
261 |
try:
|
262 |
pipe = pipeline(
|
263 |
"text-generation",
|
@@ -271,26 +197,22 @@ class BanglaRAGChain:
|
|
271 |
top_p=self.top_p,
|
272 |
top_k=self.top_k,
|
273 |
repetition_penalty=1.2,
|
274 |
-
torch_dtype=torch.
|
275 |
)
|
276 |
|
277 |
self._llm = HuggingFacePipeline(pipeline=pipe)
|
278 |
-
rprint(Panel("[bold green]LLM initialized successfully!", expand=False))
|
279 |
except Exception as e:
|
280 |
-
|
281 |
-
self._llm = None
|
282 |
|
283 |
def __call__(self, query):
|
284 |
-
"""Runs the RAG chain on a user query and returns the generated answer."""
|
285 |
if not self._chain:
|
286 |
raise ValueError("The chain has not been initialized.")
|
287 |
-
|
288 |
-
|
289 |
-
return result["answer"], result["context"]
|
290 |
|
291 |
def _format_docs(self, docs):
|
292 |
-
"""Formats retrieved documents into a string format."""
|
293 |
context = ""
|
294 |
for i, doc in enumerate(docs):
|
295 |
context += f"\nDocument {i + 1}:\n{doc.page_content}\n\n"
|
296 |
-
return context
|
|
|
4 |
AutoTokenizer,
|
5 |
AutoModelForCausalLM,
|
6 |
pipeline,
|
|
|
7 |
BitsAndBytesConfig,
|
8 |
)
|
9 |
from langchain_core.prompts import PromptTemplate
|
|
|
13 |
from langchain_community.llms.huggingface_pipeline import HuggingFacePipeline
|
14 |
from langchain_core.runnables import RunnableParallel, RunnablePassthrough
|
15 |
from langchain_core.output_parsers import StrOutputParser
|
|
|
|
|
|
|
|
|
16 |
import warnings
|
|
|
17 |
|
18 |
warnings.filterwarnings("ignore")
|
19 |
|
20 |
class BanglaRAGChain:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
def __init__(self):
|
|
|
22 |
self._device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
23 |
self.chat_model_id = None
|
24 |
self.embed_model_id = None
|
|
|
57 |
chunk_overlap=150,
|
58 |
hf_token=None,
|
59 |
):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
60 |
self.chat_model_id = chat_model_id
|
61 |
self.embed_model_id = embed_model_id
|
62 |
self.k = k
|
|
|
73 |
if self.hf_token is not None:
|
74 |
os.environ["HF_TOKEN"] = str(self.hf_token)
|
75 |
|
|
|
76 |
self._load_models()
|
|
|
|
|
77 |
self._create_document()
|
|
|
|
|
78 |
self._update_chroma_db()
|
|
|
|
|
79 |
self._get_retriever()
|
|
|
|
|
80 |
self._get_llm()
|
|
|
|
|
81 |
self._create_chain()
|
82 |
|
83 |
def _load_models(self):
|
|
|
84 |
try:
|
85 |
self.tokenizer = AutoTokenizer.from_pretrained(self.chat_model_id)
|
86 |
bnb_config = None
|
|
|
91 |
bnb_4bit_quant_type="nf4",
|
92 |
bnb_4bit_compute_dtype=torch.float16,
|
93 |
)
|
|
|
94 |
self.chat_model = AutoModelForCausalLM.from_pretrained(
|
95 |
self.chat_model_id,
|
96 |
+
load_in_8bit=True,
|
97 |
+
torch_dtype=torch.bfloat16,
|
|
|
98 |
device_map="auto",
|
99 |
+
quantization_config=bnb_config,
|
100 |
)
|
|
|
101 |
else:
|
102 |
self.chat_model = AutoModelForCausalLM.from_pretrained(
|
103 |
self.chat_model_id,
|
104 |
+
torch_dtype=torch.bfloat16,
|
|
|
105 |
device_map="auto",
|
106 |
)
|
|
|
107 |
except Exception as e:
|
108 |
+
raise RuntimeError(f"Error loading chat model: {e}")
|
109 |
|
110 |
def _create_document(self):
|
|
|
111 |
try:
|
112 |
with open(self.text_path, "r", encoding="utf-8") as file:
|
113 |
self._text_content = file.read()
|
|
|
116 |
chunk_size=self.chunk_size,
|
117 |
chunk_overlap=self.chunk_overlap,
|
118 |
)
|
119 |
+
self._documents = character_splitter.split_text(self._text_content)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
120 |
except Exception as e:
|
121 |
+
raise RuntimeError(f"Chunking failed: {e}")
|
122 |
|
123 |
def _update_chroma_db(self):
|
|
|
124 |
try:
|
125 |
+
model_kwargs = {"device": self._device}
|
126 |
+
embeddings = HuggingFaceEmbeddings(
|
127 |
+
model_name=self.embed_model_id, model_kwargs=model_kwargs
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
128 |
)
|
129 |
+
self._db = Chroma.from_texts(texts=self._documents, embedding=embeddings)
|
130 |
except Exception as e:
|
131 |
+
raise RuntimeError(f"Vector DB initialization failed: {e}")
|
132 |
|
133 |
def _create_chain(self):
|
|
|
134 |
template = """Below is an instruction in Bengali language that describes a task, paired with an input also in Bengali language that provides further context. Write a response in Bengali that appropriately completes the request.
|
135 |
### Instruction:
|
136 |
{question}
|
|
|
172 |
).assign(answer=rag_chain_from_docs)
|
173 |
|
174 |
self._chain = rag_chain_with_source
|
|
|
175 |
except Exception as e:
|
176 |
+
raise RuntimeError(f"Chain creation failed: {e}")
|
177 |
|
178 |
def _get_retriever(self):
|
|
|
179 |
try:
|
180 |
self._retriever = self._db.as_retriever(
|
181 |
search_type="similarity", search_kwargs={"k": self.k}
|
182 |
)
|
|
|
183 |
except Exception as e:
|
184 |
+
raise RuntimeError(f"Retriever creation failed: {e}")
|
185 |
|
186 |
def _get_llm(self):
|
|
|
187 |
try:
|
188 |
pipe = pipeline(
|
189 |
"text-generation",
|
|
|
197 |
top_p=self.top_p,
|
198 |
top_k=self.top_k,
|
199 |
repetition_penalty=1.2,
|
200 |
+
torch_dtype=torch.bfloat16,
|
201 |
)
|
202 |
|
203 |
self._llm = HuggingFacePipeline(pipeline=pipe)
|
|
|
204 |
except Exception as e:
|
205 |
+
raise RuntimeError(f"LLM initialization failed: {e}")
|
206 |
+
self._llm = None
|
207 |
|
208 |
def __call__(self, query):
|
|
|
209 |
if not self._chain:
|
210 |
raise ValueError("The chain has not been initialized.")
|
211 |
+
result = self._chain.invoke({"question": query})
|
212 |
+
return result["answer"], result["context"]
|
|
|
213 |
|
214 |
def _format_docs(self, docs):
|
|
|
215 |
context = ""
|
216 |
for i, doc in enumerate(docs):
|
217 |
context += f"\nDocument {i + 1}:\n{doc.page_content}\n\n"
|
218 |
+
return context
|