File size: 13,472 Bytes
132f0a2
57cf1ec
132f0a2
 
 
 
08a44c9
 
 
 
5f9b6f1
 
 
 
132f0a2
 
 
 
 
08a44c9
 
 
 
132f0a2
 
 
 
 
08a44c9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
132f0a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5f9b6f1
132f0a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5f9b6f1
132f0a2
 
5f9b6f1
132f0a2
 
 
 
 
 
 
 
 
08a44c9
132f0a2
 
 
 
08a44c9
132f0a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
08a44c9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
132f0a2
5f9b6f1
132f0a2
 
5f9b6f1
132f0a2
 
 
08a44c9
132f0a2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
08a44c9
 
132f0a2
 
 
 
 
 
 
 
 
 
 
 
 
 
08a44c9
132f0a2
 
 
08a44c9
 
 
 
132f0a2
 
 
08a44c9
132f0a2
 
 
 
 
 
08a44c9
 
132f0a2
 
5f9b6f1
132f0a2
08a44c9
 
 
 
 
 
 
 
 
 
 
 
132f0a2
08a44c9
132f0a2
 
5f9b6f1
132f0a2
 
 
 
 
 
 
 
5f9b6f1
57cf1ec
132f0a2
57cf1ec
132f0a2
8a7cf31
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
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
# app.py
import gradio as gr
import os
import re
import shutil
import torch
import pickle # For saving/loading Python objects

# LangChain imports
from langchain_community.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain_community.vectorstores import FAISS
from langchain.chains import RetrievalQA
from langchain_community.llms import HuggingFacePipeline
from transformers import pipeline, AutoTokenizer, AutoModelForSeq2SeqLM

# --- Configuration ---
ARXIV_DIR = "./arxiv_papers" # Directory to save downloaded papers
KB_STORAGE_DIR = "./knowledge_base_storage" # Directory to save/load KB
FAISS_INDEX_PATH = os.path.join(KB_STORAGE_DIR, "faiss_index.bin")
CHUNKS_PATH = os.path.join(KB_STORAGE_DIR, "knowledge_base_chunks.pkl")

CHUNK_SIZE = 500 # Characters per chunk
CHUNK_OVERLAP = 50 # Overlap between chunks
EMBEDDING_MODEL_NAME = 'all-MiniLM-L6-v2'
LLM_MODEL_NAME = "google/flan-t5-small"

# Ensure KB storage directory exists
os.makedirs(KB_STORAGE_DIR, exist_ok=True)

# --- Helper Functions for arXiv and PDF Processing ---

def clean_text(text: str) -> str:
    """Basic text cleaning: replaces multiple spaces/newlines with single space and strips whitespace."""
    text = re.sub(r'\s+', ' ', text)
    text = text.strip()
    return text

def get_arxiv_papers(query: str, max_papers: int = 5) -> list[str]:
    """
    Searches arXiv for papers, downloads their PDFs, and returns a list of file paths.
    Clears the ARXIV_DIR before downloading new papers.
    """
    # Clear existing papers before downloading new ones
    if os.path.exists(ARXIV_DIR):
        shutil.rmtree(ARXIV_DIR)
    os.makedirs(ARXIV_DIR, exist_ok=True)

    print(f"Searching arXiv for '{query}' and downloading up to {max_papers} papers...")
    import arxiv # Import here to ensure it's available when this function is called
    search_results = arxiv.Search(
        query=query,
        max_results=max_papers,
        sort_by=arxiv.SortCriterion.Relevance,
        sort_order=arxiv.SortOrder.Descending
    )
    downloaded_files = []
    for i, result in enumerate(search_results.results()):
        try:
            # Create a safe filename
            safe_title = re.sub(r'[\\/:*?"<>|]', '', result.title) # Remove invalid characters
            filename = f"{ARXIV_DIR}/{safe_title[:100]}_{result.arxiv_id}.pdf" # Limit title length
            print(f"Downloading paper {i+1}/{max_papers}: {result.title}")
            result.download_pdf(filename=filename)
            downloaded_files.append(filename)
        except Exception as e:
            print(f"Could not download {result.title}: {e}")
    return downloaded_files

# --- RAGAgent Class ---

class RAGAgent:
    def __init__(self):
        self.embedding_model = None
        self.llm = None
        self.vectorstore = None
        self.qa_chain = None
        self.is_initialized = False

    def _load_models(self):
        """Loads the embedding and generation models if not already loaded."""
        if self.embedding_model is None:
            print(f"Loading Embedding Model: {EMBEDDING_MODEL_NAME}...")
            self.embedding_model = HuggingFaceEmbeddings(model_name=EMBEDDING_MODEL_NAME)
        
        if self.llm is None:
            print(f"Loading LLM Model: {LLM_MODEL_NAME}...")
            tokenizer = AutoTokenizer.from_pretrained(LLM_MODEL_NAME)
            model = AutoModelForSeq2SeqLM.from_pretrained(LLM_MODEL_NAME)
            
            # Determine device for pipeline
            device = 0 if torch.cuda.is_available() else -1

            # Create a Hugging Face pipeline for text generation
            text_generation_pipeline = pipeline(
                "text2text-generation",
                model=model,
                tokenizer=tokenizer,
                max_new_tokens=150, # Set a default max_new_tokens for the pipeline
                min_length=20,
                num_beams=5,
                early_stopping=True,
                device=device
            )
            self.llm = HuggingFacePipeline(pipeline=text_generation_pipeline)
        
        self.is_initialized = True

    def initialize_knowledge_base(self, arxiv_query: str, max_papers: int = 5) -> str:
        """
        Initializes the knowledge base by downloading, extracting, and chunking
        arXiv papers using LangChain components, then building a FAISS vectorstore.
        """
        self._load_models() # Ensure models are loaded first

        # Clear existing papers before downloading new ones
        if os.path.exists(ARXIV_DIR):
            shutil.rmtree(ARXIV_DIR)
        os.makedirs(ARXIV_DIR, exist_ok=True)

        self.vectorstore = None
        self.qa_chain = None
        self.knowledge_base_chunks = [] # Reset chunks

        print(f"Searching arXiv for '{arxiv_query}' and downloading up to {max_papers} papers...")
        try:
            # Manual download using arxiv library (as it offers more control over filenames)
            pdf_paths = get_arxiv_papers(arxiv_query, max_papers) # Call the helper function
            
            if not pdf_paths:
                return "No papers found or downloaded for the given query. Please try a different query."

            # Load documents from downloaded PDFs using PyPDFLoader
            all_documents = []
            for pdf_path in pdf_paths:
                try:
                    loader = PyPDFLoader(pdf_path)
                    all_documents.extend(loader.load())
                except Exception as e:
                    print(f"Error loading PDF {pdf_path}: {e}")

            if not all_documents:
                return "Could not load any documents from downloaded PDFs. Please try a different query or fewer papers."

            print(f"Loaded {len(all_documents)} raw documents from PDFs.")

            # Split documents into chunks using RecursiveCharacterTextSplitter
            text_splitter = RecursiveCharacterTextSplitter(
                chunk_size=CHUNK_SIZE,
                chunk_overlap=CHUNK_OVERLAP,
                length_function=len,
                is_separator_regex=False,
            )
            self.knowledge_base_chunks = text_splitter.split_documents(all_documents)
            
            if not self.knowledge_base_chunks:
                return "No meaningful text chunks could be created from the papers after splitting."

            print(f"Total chunks created: {len(self.knowledge_base_chunks)}")

            # Create FAISS vectorstore from chunks and embeddings
            print("Creating FAISS vectorstore from chunks...")
            self.vectorstore = FAISS.from_documents(self.knowledge_base_chunks, self.embedding_model)
            print(f"FAISS vectorstore created with {len(self.knowledge_base_chunks)} documents.")

            # Create RetrievalQA chain
            self.qa_chain = RetrievalQA.from_chain_type(
                llm=self.llm,
                chain_type="stuff", # "stuff" puts all retrieved docs into one prompt
                retriever=self.vectorstore.as_retriever(search_kwargs={"k": 3}), # Retrieve top 3 docs
                return_source_documents=False # Set to True if you want to return source docs
            )
            
            return f"Knowledge base loaded with {len(self.knowledge_base_chunks)} chunks from {len(pdf_paths)} arXiv papers on '{arxiv_query}'."

        except Exception as e:
            print(f"Error during knowledge base initialization: {e}")
            return f"An error occurred during knowledge base initialization: {e}"

    def save_knowledge_base(self) -> str:
        """Saves the current FAISS vectorstore and knowledge base chunks to disk."""
        if not self.vectorstore or not self.knowledge_base_chunks:
            return "No knowledge base to save. Please load one first."
        
        try:
            # Save FAISS index
            self.vectorstore.save_local(KB_STORAGE_DIR, index_name="faiss_index")
            # Save chunks (metadata for FAISS, or for re-building if needed)
            with open(CHUNKS_PATH, 'wb') as f:
                pickle.dump(self.knowledge_base_chunks, f)
            print(f"Knowledge base saved to {KB_STORAGE_DIR}")
            return f"Knowledge base saved successfully to {KB_STORAGE_DIR}."
        except Exception as e:
            print(f"Error saving knowledge base: {e}")
            return f"Error saving knowledge base: {e}"

    def load_knowledge_base(self) -> str:
        """Loads the FAISS vectorstore and knowledge base chunks from disk."""
        self._load_models() # Ensure models are loaded before loading KB
        
        if not os.path.exists(FAISS_INDEX_PATH) or not os.path.exists(CHUNKS_PATH):
            return "Saved knowledge base not found. Please load or create one first."
        
        try:
            # Load FAISS index
            self.vectorstore = FAISS.load_local(KB_STORAGE_DIR, self.embedding_model, index_name="faiss_index", allow_dangerous_deserialization=True)
            # Load chunks
            with open(CHUNKS_PATH, 'rb') as f:
                self.knowledge_base_chunks = pickle.load(f)
            
            # Re-create RetrievalQA chain after loading vectorstore
            self.qa_chain = RetrievalQA.from_chain_type(
                llm=self.llm,
                chain_type="stuff",
                retriever=self.vectorstore.as_retriever(search_kwargs={"k": 3}),
                return_source_documents=False
            )
            
            print(f"Knowledge base loaded from {KB_STORAGE_DIR}")
            return f"Knowledge base loaded successfully from {KB_STORAGE_DIR} with {len(self.knowledge_base_chunks)} chunks."
        except Exception as e:
            print(f"Error loading knowledge base: {e}")
            self.vectorstore = None
            self.qa_chain = None
            self.knowledge_base_chunks = []
            return f"Error loading knowledge base: {e}"

    def query_agent(self, query: str) -> str:
        """
        Retrieves relevant information from the knowledge base and generates an answer
         using the LangChain RetrievalQA chain.
        """
        if not query.strip():
            return "Please enter a question."
        if not self.is_initialized or self.qa_chain is None:
            return "Knowledge base not loaded. Please initialize it by providing an arXiv query or loading from disk."

        print(f"\n--- Querying LLM with LangChain QA Chain ---\nQuestion: {query}\n----------------------")

        try:
            # Use the RetrievalQA chain to get the answer
            result = self.qa_chain.invoke({"query": query})
            answer = result["result"].strip()
        except Exception as e:
            print(f"Error during generation: {e}")
            answer = "I apologize, but I encountered an error while generating the answer. Please try again or rephrase your question."

        return answer

# --- Gradio Interface ---

# Instantiate the RAGAgent
rag_agent_instance = RAGAgent()

print("Setting up Gradio interface...")

with gr.Blocks() as demo:
    gr.Markdown("# 📚 Educational RAG Agent with Persistent Knowledge Base")
    gr.Markdown("First, load a knowledge base from arXiv, then you can save it or load a previously saved one. Finally, ask questions!")

    with gr.Row():
        arxiv_input = gr.Textbox(
            label="arXiv Search Query (e.g., 'Large Language Models', 'Reinforcement Learning')",
            placeholder="Enter a topic to search for papers on arXiv...",
            lines=1
        )
        max_papers_slider = gr.Slider(
            minimum=1,
            maximum=10,
            step=1,
            value=3,
            label="Max Papers to Download"
        )
        load_kb_from_arxiv_button = gr.Button("Load KB from arXiv")
    
    kb_status_output = gr.Textbox(label="Knowledge Base Status", interactive=False)

    with gr.Row():
        save_kb_button = gr.Button("Save Knowledge Base to Disk")
        load_kb_from_disk_button = gr.Button("Load Knowledge Base from Disk")

    with gr.Row():
        question_input = gr.Textbox(
            lines=3,
            placeholder="Ask a question based on the loaded knowledge base...",
            label="Your Question"
        )
        answer_output = gr.Textbox(label="Answer", lines=7, interactive=False)
    
    submit_button = gr.Button("Get Answer")

    load_kb_from_arxiv_button.click(
        fn=rag_agent_instance.initialize_knowledge_base,
        inputs=[arxiv_input, max_papers_slider],
        outputs=kb_status_output
    )

    save_kb_button.click(
        fn=rag_agent_instance.save_knowledge_base,
        inputs=[],
        outputs=kb_status_output
    )

    load_kb_from_disk_button.click(
        fn=rag_agent_instance.load_knowledge_base,
        inputs=[],
        outputs=kb_status_output
    )

    submit_button.click(
        fn=rag_agent_instance.query_agent,
        inputs=question_input,
        outputs=answer_output
    )

    gr.Examples(
        examples=[
            ["What is the transformer architecture?"],
            ["Explain attention mechanisms in deep learning."],
            ["What are the challenges in reinforcement learning?"],
        ],
        inputs=question_input
    )

# Launch the Gradio app
if __name__ == "__main__":
    print("Launching Gradio app...")
    demo.launch(share=False)