Spaces:
Sleeping
Sleeping
arjunanand13
commited on
Commit
•
184e87b
1
Parent(s):
e1175ed
Update app3.py
Browse files
app3.py
CHANGED
@@ -16,16 +16,152 @@ from langchain_community.llms import HuggingFaceEndpoint
|
|
16 |
from langchain.memory import ConversationBufferMemory
|
17 |
from sentence_transformers import SentenceTransformer, util
|
18 |
import torch
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
|
20 |
# Constants and setup
|
21 |
list_llm = ["meta-llama/Meta-Llama-3-8B-Instruct", "mistralai/Mistral-7B-Instruct-v0.2"]
|
22 |
list_llm_simple = [os.path.basename(llm) for llm in list_llm]
|
23 |
api_token = os.getenv("HF_TOKEN")
|
24 |
|
|
|
|
|
|
|
|
|
|
|
25 |
# Initialize sentence transformer for evaluation
|
26 |
sentence_model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
|
27 |
|
28 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
def get_text_splitter(strategy: str, chunk_size: int = 1024, chunk_overlap: int = 64):
|
30 |
splitters = {
|
31 |
"recursive": RecursiveCharacterTextSplitter(
|
@@ -43,105 +179,38 @@ def get_text_splitter(strategy: str, chunk_size: int = 1024, chunk_overlap: int
|
|
43 |
}
|
44 |
return splitters.get(strategy)
|
45 |
|
46 |
-
|
47 |
-
|
48 |
-
embeddings1 = sentence_model.encode([text1], convert_to_tensor=True)
|
49 |
-
embeddings2 = sentence_model.encode([text2], convert_to_tensor=True)
|
50 |
-
similarity = util.pytorch_cos_sim(embeddings1, embeddings2)
|
51 |
-
return float(similarity[0][0])
|
52 |
-
|
53 |
-
def evaluate_response(question: str, answer: str, ground_truth: str, contexts: List[str]) -> Dict[str, float]:
|
54 |
-
# Answer similarity with ground truth
|
55 |
-
answer_similarity = calculate_semantic_similarity(answer, ground_truth)
|
56 |
-
|
57 |
-
# Context relevance - average similarity between question and contexts
|
58 |
-
context_scores = [calculate_semantic_similarity(question, ctx) for ctx in contexts]
|
59 |
-
context_relevance = np.mean(context_scores)
|
60 |
-
|
61 |
-
# Answer relevance - similarity between question and answer
|
62 |
-
answer_relevance = calculate_semantic_similarity(question, answer)
|
63 |
-
|
64 |
-
return {
|
65 |
-
"answer_similarity": answer_similarity,
|
66 |
-
"context_relevance": context_relevance,
|
67 |
-
"answer_relevance": answer_relevance,
|
68 |
-
"average_score": np.mean([answer_similarity, context_relevance, answer_relevance])
|
69 |
-
}
|
70 |
-
|
71 |
-
# Load and split PDF document
|
72 |
-
def load_doc(list_file_path: List[str], splitting_strategy: str = "recursive"):
|
73 |
loaders = [PyPDFLoader(x) for x in list_file_path]
|
74 |
pages = []
|
75 |
for loader in loaders:
|
76 |
pages.extend(loader.load())
|
77 |
|
78 |
-
text_splitter = get_text_splitter(splitting_strategy)
|
79 |
doc_splits = text_splitter.split_documents(pages)
|
80 |
return doc_splits
|
81 |
|
82 |
-
# Vector database creation functions
|
83 |
-
def create_faiss_db(splits, embeddings):
|
84 |
-
return FAISS.from_documents(splits, embeddings)
|
85 |
-
|
86 |
-
def create_chroma_db(splits, embeddings):
|
87 |
-
return Chroma.from_documents(splits, embeddings)
|
88 |
-
|
89 |
-
def create_qdrant_db(splits, embeddings):
|
90 |
-
return Qdrant.from_documents(
|
91 |
-
splits,
|
92 |
-
embeddings,
|
93 |
-
location=":memory:",
|
94 |
-
collection_name="pdf_docs"
|
95 |
-
)
|
96 |
-
|
97 |
def create_db(splits, db_choice: str = "faiss"):
|
98 |
embeddings = HuggingFaceEmbeddings()
|
99 |
db_creators = {
|
100 |
-
"faiss":
|
101 |
-
"chroma":
|
102 |
-
"qdrant":
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
dataset = load_dataset("explodinggradients/fiqa", split="test", trust_remote_code=True)
|
108 |
-
return dataset
|
109 |
-
|
110 |
-
def evaluate_rag_pipeline(qa_chain, dataset):
|
111 |
-
# Sample a few examples for evaluation
|
112 |
-
eval_samples = dataset.select(range(5))
|
113 |
-
|
114 |
-
results = []
|
115 |
-
for sample in eval_samples:
|
116 |
-
question = sample["question"]
|
117 |
-
|
118 |
-
# Get response from the chain
|
119 |
-
response = qa_chain.invoke({
|
120 |
-
"question": question,
|
121 |
-
"chat_history": []
|
122 |
-
})
|
123 |
-
|
124 |
-
# Evaluate response
|
125 |
-
eval_result = evaluate_response(
|
126 |
-
question=question,
|
127 |
-
answer=response["answer"],
|
128 |
-
ground_truth=sample["answer"],
|
129 |
-
contexts=[doc.page_content for doc in response["source_documents"]]
|
130 |
)
|
131 |
-
|
132 |
-
results.append(eval_result)
|
133 |
-
|
134 |
-
# Calculate average scores across all samples
|
135 |
-
avg_results = {
|
136 |
-
metric: float(np.mean([r[metric] for r in results]))
|
137 |
-
for metric in results[0].keys()
|
138 |
}
|
139 |
-
|
140 |
-
|
|
|
|
|
|
|
|
|
|
|
141 |
|
142 |
-
# Initialize langchain LLM chain
|
143 |
def initialize_llmchain(llm_choice, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
|
144 |
-
# Get the full model name from the index
|
145 |
llm_model = list_llm[llm_choice]
|
146 |
|
147 |
llm = HuggingFaceEndpoint(
|
@@ -149,8 +218,7 @@ def initialize_llmchain(llm_choice, temperature, max_tokens, top_k, vector_db, p
|
|
149 |
huggingfacehub_api_token=api_token,
|
150 |
temperature=temperature,
|
151 |
max_new_tokens=max_tokens,
|
152 |
-
top_k=top_k
|
153 |
-
model=llm_model # Add model parameter
|
154 |
)
|
155 |
|
156 |
memory = ConversationBufferMemory(
|
@@ -163,162 +231,278 @@ def initialize_llmchain(llm_choice, temperature, max_tokens, top_k, vector_db, p
|
|
163 |
qa_chain = ConversationalRetrievalChain.from_llm(
|
164 |
llm,
|
165 |
retriever=retriever,
|
166 |
-
chain_type="stuff",
|
167 |
memory=memory,
|
168 |
-
return_source_documents=True
|
169 |
-
verbose=False,
|
170 |
)
|
171 |
return qa_chain, "LLM initialized successfully!"
|
172 |
|
173 |
-
def initialize_database(list_file_obj, splitting_strategy, db_choice, progress=gr.Progress()):
|
174 |
-
list_file_path = [x.name for x in list_file_obj if x is not None]
|
175 |
-
doc_splits = load_doc(list_file_path, splitting_strategy)
|
176 |
-
vector_db = create_db(doc_splits, db_choice)
|
177 |
-
return vector_db, f"Database created using {splitting_strategy} splitting and {db_choice} vector database!"
|
178 |
-
|
179 |
-
def format_chat_history(message, chat_history):
|
180 |
-
formatted_chat_history = []
|
181 |
-
for user_message, bot_message in chat_history:
|
182 |
-
formatted_chat_history.append(f"User: {user_message}")
|
183 |
-
formatted_chat_history.append(f"Assistant: {bot_message}")
|
184 |
-
return formatted_chat_history
|
185 |
-
|
186 |
def conversation(qa_chain, message, history):
|
187 |
-
|
188 |
response = qa_chain.invoke({
|
189 |
"question": message,
|
190 |
-
"chat_history":
|
191 |
})
|
192 |
|
193 |
response_answer = response["answer"]
|
194 |
-
if
|
195 |
response_answer = response_answer.split("Helpful Answer:")[-1]
|
196 |
|
197 |
-
|
198 |
-
|
199 |
-
|
200 |
-
|
201 |
|
202 |
-
|
203 |
-
|
204 |
-
|
|
|
205 |
|
206 |
-
|
|
|
|
|
|
|
207 |
|
208 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
209 |
|
210 |
def demo():
|
|
|
|
|
211 |
with gr.Blocks(theme=gr.themes.Default(primary_hue="red", secondary_hue="pink", neutral_hue="sky")) as demo:
|
212 |
vector_db = gr.State()
|
213 |
qa_chain = gr.State()
|
214 |
|
215 |
-
gr.HTML("<center><h1>Enhanced RAG PDF Chatbot</h1></center>")
|
216 |
-
gr.Markdown("""<b>Query your PDF documents with advanced RAG capabilities!</b>""")
|
217 |
|
218 |
-
with gr.
|
219 |
-
|
220 |
-
|
221 |
with gr.Row():
|
222 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
223 |
|
224 |
with gr.Row():
|
225 |
-
|
|
|
|
|
|
|
226 |
["recursive", "fixed", "token"],
|
227 |
label="Text Splitting Strategy",
|
228 |
value="recursive"
|
229 |
)
|
230 |
-
|
231 |
-
["
|
232 |
-
label="
|
233 |
-
value="
|
234 |
)
|
235 |
|
236 |
with gr.Row():
|
237 |
-
|
238 |
-
|
239 |
-
|
240 |
-
with gr.Row():
|
241 |
-
db_progress = gr.Textbox(value="Not initialized", show_label=False)
|
242 |
-
evaluation_results = gr.JSON(label="Evaluation Results")
|
243 |
-
|
244 |
-
gr.Markdown("<b>Select Large Language Model (LLM) and input parameters</b>")
|
245 |
-
with gr.Row():
|
246 |
-
llm_btn = gr.Radio(list_llm_simple, label="Available LLMs", value=list_llm_simple[0], type="index")
|
247 |
-
|
248 |
-
with gr.Row():
|
249 |
-
with gr.Accordion("LLM input parameters", open=False):
|
250 |
-
slider_temperature = gr.Slider(minimum=0.01, maximum=1.0, value=0.5, step=0.1, label="Temperature")
|
251 |
-
slider_maxtokens = gr.Slider(minimum=128, maximum=9192, value=4096, step=128, label="Max New Tokens")
|
252 |
-
slider_topk = gr.Slider(minimum=1, maximum=10, value=3, step=1, label="top-k")
|
253 |
-
|
254 |
-
with gr.Row():
|
255 |
-
qachain_btn = gr.Button("Initialize Question Answering Chatbot")
|
256 |
-
llm_progress = gr.Textbox(value="Not initialized", show_label=False)
|
257 |
-
|
258 |
-
with gr.Column(scale=200):
|
259 |
-
gr.Markdown("<b>Step 2 - Chat with your Document</b>")
|
260 |
-
chatbot = gr.Chatbot(height=505)
|
261 |
-
|
262 |
-
with gr.Accordion("Relevant context from the source document", open=False):
|
263 |
-
with gr.Row():
|
264 |
-
doc_source1 = gr.Textbox(label="Reference 1", lines=2, container=True, scale=20)
|
265 |
-
source1_page = gr.Number(label="Page", scale=1)
|
266 |
-
with gr.Row():
|
267 |
-
doc_source2 = gr.Textbox(label="Reference 2", lines=2, container=True, scale=20)
|
268 |
-
source2_page = gr.Number(label="Page", scale=1)
|
269 |
-
with gr.Row():
|
270 |
-
doc_source3 = gr.Textbox(label="Reference 3", lines=2, container=True, scale=20)
|
271 |
-
source3_page = gr.Number(label="Page", scale=1)
|
272 |
-
|
273 |
-
with gr.Row():
|
274 |
-
msg = gr.Textbox(placeholder="Ask a question", container=True)
|
275 |
-
with gr.Row():
|
276 |
-
submit_btn = gr.Button("Submit")
|
277 |
-
clear_btn = gr.ClearButton([msg, chatbot], value="Clear")
|
278 |
-
|
279 |
# Event handlers
|
280 |
db_btn.click(
|
281 |
initialize_database,
|
282 |
-
inputs=[document, splitting_strategy, db_choice],
|
283 |
outputs=[vector_db, db_progress]
|
284 |
)
|
285 |
|
286 |
-
|
287 |
-
|
288 |
-
inputs=[
|
289 |
-
outputs=[evaluation_results]
|
290 |
-
)
|
291 |
-
|
292 |
-
qachain_btn.click(
|
293 |
-
initialize_llmchain, # Fixed function name here
|
294 |
-
inputs=[llm_btn, slider_temperature, slider_maxtokens, slider_topk, vector_db],
|
295 |
outputs=[qa_chain, llm_progress]
|
296 |
-
).then(
|
297 |
-
lambda: [None, "", 0, "", 0, "", 0],
|
298 |
-
inputs=None,
|
299 |
-
outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page],
|
300 |
-
queue=False
|
301 |
)
|
302 |
|
303 |
-
msg.submit(
|
|
|
304 |
inputs=[qa_chain, msg, chatbot],
|
305 |
-
outputs=[qa_chain, msg, chatbot,
|
306 |
-
queue=False
|
307 |
)
|
308 |
|
309 |
-
submit_btn.click(
|
|
|
310 |
inputs=[qa_chain, msg, chatbot],
|
311 |
-
outputs=[qa_chain, msg, chatbot,
|
312 |
-
queue=False
|
313 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
314 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
315 |
clear_btn.click(
|
316 |
lambda: [None, "", 0, "", 0, "", 0],
|
317 |
-
|
318 |
-
outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page],
|
319 |
-
queue=False
|
320 |
)
|
321 |
-
|
|
|
322 |
demo.queue().launch(debug=True)
|
323 |
|
324 |
if __name__ == "__main__":
|
|
|
16 |
from langchain.memory import ConversationBufferMemory
|
17 |
from sentence_transformers import SentenceTransformer, util
|
18 |
import torch
|
19 |
+
from ragas import evaluate
|
20 |
+
from ragas.metrics import (
|
21 |
+
ContextRecall,
|
22 |
+
AnswerRelevancy,
|
23 |
+
Faithfulness,
|
24 |
+
ContextPrecision
|
25 |
+
)
|
26 |
+
import pandas as pd
|
27 |
|
28 |
# Constants and setup
|
29 |
list_llm = ["meta-llama/Meta-Llama-3-8B-Instruct", "mistralai/Mistral-7B-Instruct-v0.2"]
|
30 |
list_llm_simple = [os.path.basename(llm) for llm in list_llm]
|
31 |
api_token = os.getenv("HF_TOKEN")
|
32 |
|
33 |
+
CHUNK_SIZES = {
|
34 |
+
"small": {"recursive": 512, "fixed": 512, "token": 256},
|
35 |
+
"medium": {"recursive": 1024, "fixed": 1024, "token": 512}
|
36 |
+
}
|
37 |
+
|
38 |
# Initialize sentence transformer for evaluation
|
39 |
sentence_model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
|
40 |
|
41 |
+
class RAGEvaluator:
|
42 |
+
def __init__(self):
|
43 |
+
self.datasets = {
|
44 |
+
"squad": "squad_v2",
|
45 |
+
"msmarco": "ms_marco"
|
46 |
+
}
|
47 |
+
self.current_dataset = None
|
48 |
+
self.test_samples = []
|
49 |
+
|
50 |
+
def load_dataset(self, dataset_name: str, num_samples: int = 10):
|
51 |
+
"""Load a smaller subset of questions with proper error handling"""
|
52 |
+
try:
|
53 |
+
if dataset_name == "squad":
|
54 |
+
dataset = load_dataset("squad_v2", split="validation")
|
55 |
+
# Select diverse questions
|
56 |
+
samples = dataset.select(range(0, 1000, 100))[:num_samples]
|
57 |
+
|
58 |
+
self.test_samples = []
|
59 |
+
for sample in samples:
|
60 |
+
# Check if answers exist and are not empty
|
61 |
+
if sample.get("answers") and isinstance(sample["answers"], dict) and sample["answers"].get("text"):
|
62 |
+
self.test_samples.append({
|
63 |
+
"question": sample["question"],
|
64 |
+
"ground_truth": sample["answers"]["text"][0],
|
65 |
+
"context": sample["context"]
|
66 |
+
})
|
67 |
+
|
68 |
+
elif dataset_name == "msmarco":
|
69 |
+
dataset = load_dataset("ms_marco", "v2.1", split="dev")
|
70 |
+
samples = dataset.select(range(0, 1000, 100))[:num_samples]
|
71 |
+
|
72 |
+
self.test_samples = []
|
73 |
+
for sample in samples:
|
74 |
+
# Check for valid answers
|
75 |
+
if sample.get("answers") and sample["answers"]:
|
76 |
+
self.test_samples.append({
|
77 |
+
"question": sample["query"],
|
78 |
+
"ground_truth": sample["answers"][0],
|
79 |
+
"context": sample["passages"][0]["passage_text"]
|
80 |
+
if isinstance(sample["passages"], list)
|
81 |
+
else sample["passages"]["passage_text"][0]
|
82 |
+
})
|
83 |
+
|
84 |
+
self.current_dataset = dataset_name
|
85 |
+
|
86 |
+
# Return dataset info
|
87 |
+
return {
|
88 |
+
"dataset": dataset_name,
|
89 |
+
"num_samples": len(self.test_samples),
|
90 |
+
"sample_questions": [s["question"] for s in self.test_samples[:3]],
|
91 |
+
"status": "success"
|
92 |
+
}
|
93 |
+
|
94 |
+
except Exception as e:
|
95 |
+
print(f"Error loading dataset: {str(e)}")
|
96 |
+
return {
|
97 |
+
"dataset": dataset_name,
|
98 |
+
"error": str(e),
|
99 |
+
"status": "failed"
|
100 |
+
}
|
101 |
+
|
102 |
+
def evaluate_configuration(self, vector_db, qa_chain, splitting_strategy: str, chunk_size: str) -> Dict:
|
103 |
+
"""Evaluate with progress tracking and error handling"""
|
104 |
+
if not self.test_samples:
|
105 |
+
return {"error": "No dataset loaded"}
|
106 |
+
|
107 |
+
results = []
|
108 |
+
total_questions = len(self.test_samples)
|
109 |
+
|
110 |
+
# Add progress tracking
|
111 |
+
for i, sample in enumerate(self.test_samples):
|
112 |
+
print(f"Evaluating question {i+1}/{total_questions}")
|
113 |
+
|
114 |
+
try:
|
115 |
+
response = qa_chain.invoke({
|
116 |
+
"question": sample["question"],
|
117 |
+
"chat_history": []
|
118 |
+
})
|
119 |
+
|
120 |
+
results.append({
|
121 |
+
"question": sample["question"],
|
122 |
+
"answer": response["answer"],
|
123 |
+
"contexts": [doc.page_content for doc in response["source_documents"]],
|
124 |
+
"ground_truths": [sample["ground_truth"]]
|
125 |
+
})
|
126 |
+
except Exception as e:
|
127 |
+
print(f"Error processing question {i+1}: {str(e)}")
|
128 |
+
continue
|
129 |
+
|
130 |
+
if not results:
|
131 |
+
return {
|
132 |
+
"configuration": f"{splitting_strategy}_{chunk_size}",
|
133 |
+
"error": "No successful evaluations",
|
134 |
+
"questions_evaluated": 0
|
135 |
+
}
|
136 |
+
|
137 |
+
try:
|
138 |
+
# Calculate RAGAS metrics
|
139 |
+
eval_dataset = Dataset.from_list(results)
|
140 |
+
metrics = [ContextRecall(), AnswerRelevancy(), Faithfulness(), ContextPrecision()]
|
141 |
+
scores = evaluate(eval_dataset, metrics=metrics)
|
142 |
+
|
143 |
+
return {
|
144 |
+
"configuration": f"{splitting_strategy}_{chunk_size}",
|
145 |
+
"questions_evaluated": len(results),
|
146 |
+
"context_recall": float(scores['context_recall']),
|
147 |
+
"answer_relevancy": float(scores['answer_relevancy']),
|
148 |
+
"faithfulness": float(scores['faithfulness']),
|
149 |
+
"context_precision": float(scores['context_precision']),
|
150 |
+
"average_score": float(np.mean([
|
151 |
+
scores['context_recall'],
|
152 |
+
scores['answer_relevancy'],
|
153 |
+
scores['faithfulness'],
|
154 |
+
scores['context_precision']
|
155 |
+
]))
|
156 |
+
}
|
157 |
+
except Exception as e:
|
158 |
+
return {
|
159 |
+
"configuration": f"{splitting_strategy}_{chunk_size}",
|
160 |
+
"error": str(e),
|
161 |
+
"questions_evaluated": len(results)
|
162 |
+
}
|
163 |
+
|
164 |
+
# Text splitting and database functions
|
165 |
def get_text_splitter(strategy: str, chunk_size: int = 1024, chunk_overlap: int = 64):
|
166 |
splitters = {
|
167 |
"recursive": RecursiveCharacterTextSplitter(
|
|
|
179 |
}
|
180 |
return splitters.get(strategy)
|
181 |
|
182 |
+
def load_doc(list_file_path: List[str], splitting_strategy: str, chunk_size: str):
|
183 |
+
chunk_size_value = CHUNK_SIZES[chunk_size][splitting_strategy]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
184 |
loaders = [PyPDFLoader(x) for x in list_file_path]
|
185 |
pages = []
|
186 |
for loader in loaders:
|
187 |
pages.extend(loader.load())
|
188 |
|
189 |
+
text_splitter = get_text_splitter(splitting_strategy, chunk_size_value)
|
190 |
doc_splits = text_splitter.split_documents(pages)
|
191 |
return doc_splits
|
192 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
193 |
def create_db(splits, db_choice: str = "faiss"):
|
194 |
embeddings = HuggingFaceEmbeddings()
|
195 |
db_creators = {
|
196 |
+
"faiss": lambda: FAISS.from_documents(splits, embeddings),
|
197 |
+
"chroma": lambda: Chroma.from_documents(splits, embeddings),
|
198 |
+
"qdrant": lambda: Qdrant.from_documents(
|
199 |
+
splits,
|
200 |
+
embeddings,
|
201 |
+
location=":memory:",
|
202 |
+
collection_name="pdf_docs"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
203 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
204 |
}
|
205 |
+
return db_creators[db_choice]()
|
206 |
+
|
207 |
+
def initialize_database(list_file_obj, splitting_strategy, chunk_size, db_choice, progress=gr.Progress()):
|
208 |
+
list_file_path = [x.name for x in list_file_obj if x is not None]
|
209 |
+
doc_splits = load_doc(list_file_path, splitting_strategy, chunk_size)
|
210 |
+
vector_db = create_db(doc_splits, db_choice)
|
211 |
+
return vector_db, f"Database created using {splitting_strategy} splitting and {db_choice} vector database!"
|
212 |
|
|
|
213 |
def initialize_llmchain(llm_choice, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
|
|
|
214 |
llm_model = list_llm[llm_choice]
|
215 |
|
216 |
llm = HuggingFaceEndpoint(
|
|
|
218 |
huggingfacehub_api_token=api_token,
|
219 |
temperature=temperature,
|
220 |
max_new_tokens=max_tokens,
|
221 |
+
top_k=top_k
|
|
|
222 |
)
|
223 |
|
224 |
memory = ConversationBufferMemory(
|
|
|
231 |
qa_chain = ConversationalRetrievalChain.from_llm(
|
232 |
llm,
|
233 |
retriever=retriever,
|
|
|
234 |
memory=memory,
|
235 |
+
return_source_documents=True
|
|
|
236 |
)
|
237 |
return qa_chain, "LLM initialized successfully!"
|
238 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
239 |
def conversation(qa_chain, message, history):
|
240 |
+
"""Fixed conversation function returning all required outputs"""
|
241 |
response = qa_chain.invoke({
|
242 |
"question": message,
|
243 |
+
"chat_history": [(hist[0], hist[1]) for hist in history]
|
244 |
})
|
245 |
|
246 |
response_answer = response["answer"]
|
247 |
+
if "Helpful Answer:" in response_answer:
|
248 |
response_answer = response_answer.split("Helpful Answer:")[-1]
|
249 |
|
250 |
+
# Get source documents, ensure we have exactly 3
|
251 |
+
sources = response["source_documents"][:3]
|
252 |
+
source_contents = []
|
253 |
+
source_pages = []
|
254 |
|
255 |
+
# Process available sources
|
256 |
+
for source in sources:
|
257 |
+
source_contents.append(source.page_content.strip())
|
258 |
+
source_pages.append(source.metadata.get("page", 0) + 1)
|
259 |
|
260 |
+
# Pad with empty values if we have fewer than 3 sources
|
261 |
+
while len(source_contents) < 3:
|
262 |
+
source_contents.append("")
|
263 |
+
source_pages.append(0)
|
264 |
|
265 |
+
# Return all required outputs in correct order
|
266 |
+
return (
|
267 |
+
qa_chain, # State
|
268 |
+
gr.update(value=""), # Clear message box
|
269 |
+
history + [(message, response_answer)], # Updated chat history
|
270 |
+
source_contents[0], # First source
|
271 |
+
source_pages[0], # First page
|
272 |
+
source_contents[1], # Second source
|
273 |
+
source_pages[1], # Second page
|
274 |
+
source_contents[2], # Third source
|
275 |
+
source_pages[2] # Third page
|
276 |
+
)
|
277 |
|
278 |
def demo():
|
279 |
+
evaluator = RAGEvaluator()
|
280 |
+
|
281 |
with gr.Blocks(theme=gr.themes.Default(primary_hue="red", secondary_hue="pink", neutral_hue="sky")) as demo:
|
282 |
vector_db = gr.State()
|
283 |
qa_chain = gr.State()
|
284 |
|
285 |
+
gr.HTML("<center><h1>Enhanced RAG PDF Chatbot with Evaluation</h1></center>")
|
|
|
286 |
|
287 |
+
with gr.Tabs():
|
288 |
+
# Custom PDF Tab
|
289 |
+
with gr.Tab("Custom PDF Chat"):
|
290 |
with gr.Row():
|
291 |
+
with gr.Column(scale=86):
|
292 |
+
gr.Markdown("<b>Step 1 - Configure and Initialize RAG Pipeline</b>")
|
293 |
+
with gr.Row():
|
294 |
+
document = gr.Files(
|
295 |
+
height=300,
|
296 |
+
file_count="multiple",
|
297 |
+
file_types=["pdf"],
|
298 |
+
interactive=True,
|
299 |
+
label="Upload PDF documents"
|
300 |
+
)
|
301 |
+
|
302 |
+
with gr.Row():
|
303 |
+
splitting_strategy = gr.Radio(
|
304 |
+
["recursive", "fixed", "token"],
|
305 |
+
label="Text Splitting Strategy",
|
306 |
+
value="recursive"
|
307 |
+
)
|
308 |
+
db_choice = gr.Radio(
|
309 |
+
["faiss", "chroma", "qdrant"],
|
310 |
+
label="Vector Database",
|
311 |
+
value="faiss"
|
312 |
+
)
|
313 |
+
chunk_size = gr.Radio(
|
314 |
+
["small", "medium"],
|
315 |
+
label="Chunk Size",
|
316 |
+
value="medium"
|
317 |
+
)
|
318 |
+
|
319 |
+
with gr.Row():
|
320 |
+
db_btn = gr.Button("Create vector database")
|
321 |
+
db_progress = gr.Textbox(
|
322 |
+
value="Not initialized",
|
323 |
+
show_label=False
|
324 |
+
)
|
325 |
+
|
326 |
+
gr.Markdown("<b>Step 2 - Configure LLM</b>")
|
327 |
+
with gr.Row():
|
328 |
+
llm_choice = gr.Radio(
|
329 |
+
list_llm_simple,
|
330 |
+
label="Available LLMs",
|
331 |
+
value=list_llm_simple[0],
|
332 |
+
type="index"
|
333 |
+
)
|
334 |
+
|
335 |
+
with gr.Row():
|
336 |
+
with gr.Accordion("LLM Parameters", open=False):
|
337 |
+
temperature = gr.Slider(
|
338 |
+
minimum=0.01,
|
339 |
+
maximum=1.0,
|
340 |
+
value=0.5,
|
341 |
+
step=0.1,
|
342 |
+
label="Temperature"
|
343 |
+
)
|
344 |
+
max_tokens = gr.Slider(
|
345 |
+
minimum=128,
|
346 |
+
maximum=4096,
|
347 |
+
value=2048,
|
348 |
+
step=128,
|
349 |
+
label="Max Tokens"
|
350 |
+
)
|
351 |
+
top_k = gr.Slider(
|
352 |
+
minimum=1,
|
353 |
+
maximum=10,
|
354 |
+
value=3,
|
355 |
+
step=1,
|
356 |
+
label="Top K"
|
357 |
+
)
|
358 |
+
|
359 |
+
with gr.Row():
|
360 |
+
init_llm_btn = gr.Button("Initialize LLM")
|
361 |
+
llm_progress = gr.Textbox(
|
362 |
+
value="Not initialized",
|
363 |
+
show_label=False
|
364 |
+
)
|
365 |
+
|
366 |
+
with gr.Column(scale=200):
|
367 |
+
gr.Markdown("<b>Step 3 - Chat with Documents</b>")
|
368 |
+
chatbot = gr.Chatbot(height=505)
|
369 |
+
|
370 |
+
with gr.Accordion("Source References", open=False):
|
371 |
+
with gr.Row():
|
372 |
+
source1 = gr.Textbox(label="Source 1", lines=2)
|
373 |
+
page1 = gr.Number(label="Page")
|
374 |
+
with gr.Row():
|
375 |
+
source2 = gr.Textbox(label="Source 2", lines=2)
|
376 |
+
page2 = gr.Number(label="Page")
|
377 |
+
with gr.Row():
|
378 |
+
source3 = gr.Textbox(label="Source 3", lines=2)
|
379 |
+
page3 = gr.Number(label="Page")
|
380 |
+
|
381 |
+
with gr.Row():
|
382 |
+
msg = gr.Textbox(
|
383 |
+
placeholder="Ask a question",
|
384 |
+
show_label=False
|
385 |
+
)
|
386 |
+
with gr.Row():
|
387 |
+
submit_btn = gr.Button("Submit")
|
388 |
+
clear_btn = gr.ClearButton(
|
389 |
+
[msg, chatbot],
|
390 |
+
value="Clear Chat"
|
391 |
+
)
|
392 |
+
|
393 |
+
# Evaluation Tab
|
394 |
+
with gr.Tab("RAG Evaluation"):
|
395 |
+
with gr.Row():
|
396 |
+
dataset_choice = gr.Dropdown(
|
397 |
+
choices=list(evaluator.datasets.keys()),
|
398 |
+
label="Select Evaluation Dataset",
|
399 |
+
value="squad"
|
400 |
+
)
|
401 |
+
load_dataset_btn = gr.Button("Load Dataset")
|
402 |
|
403 |
with gr.Row():
|
404 |
+
dataset_info = gr.JSON(label="Dataset Information")
|
405 |
+
|
406 |
+
with gr.Row():
|
407 |
+
eval_splitting_strategy = gr.Radio(
|
408 |
["recursive", "fixed", "token"],
|
409 |
label="Text Splitting Strategy",
|
410 |
value="recursive"
|
411 |
)
|
412 |
+
eval_chunk_size = gr.Radio(
|
413 |
+
["small", "medium"],
|
414 |
+
label="Chunk Size",
|
415 |
+
value="medium"
|
416 |
)
|
417 |
|
418 |
with gr.Row():
|
419 |
+
evaluate_btn = gr.Button("Run Evaluation")
|
420 |
+
evaluation_results = gr.DataFrame(label="Evaluation Results")
|
421 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
422 |
# Event handlers
|
423 |
db_btn.click(
|
424 |
initialize_database,
|
425 |
+
inputs=[document, splitting_strategy, chunk_size, db_choice],
|
426 |
outputs=[vector_db, db_progress]
|
427 |
)
|
428 |
|
429 |
+
init_llm_btn.click(
|
430 |
+
initialize_llmchain,
|
431 |
+
inputs=[llm_choice, temperature, max_tokens, top_k, vector_db],
|
|
|
|
|
|
|
|
|
|
|
|
|
432 |
outputs=[qa_chain, llm_progress]
|
|
|
|
|
|
|
|
|
|
|
433 |
)
|
434 |
|
435 |
+
msg.submit(
|
436 |
+
conversation,
|
437 |
inputs=[qa_chain, msg, chatbot],
|
438 |
+
outputs=[qa_chain, msg, chatbot, source1, page1, source2, page2, source3, page3]
|
|
|
439 |
)
|
440 |
|
441 |
+
submit_btn.click(
|
442 |
+
conversation,
|
443 |
inputs=[qa_chain, msg, chatbot],
|
444 |
+
outputs=[qa_chain, msg, chatbot, source1, page1, source2, page2, source3, page3]
|
|
|
445 |
)
|
446 |
+
|
447 |
+
def load_dataset_handler(dataset_name):
|
448 |
+
try:
|
449 |
+
result = evaluator.load_dataset(dataset_name)
|
450 |
+
if result.get("status") == "success":
|
451 |
+
return {
|
452 |
+
"dataset": result["dataset"],
|
453 |
+
"samples_loaded": result["num_samples"],
|
454 |
+
"example_questions": result["sample_questions"],
|
455 |
+
"status": "ready for evaluation"
|
456 |
+
}
|
457 |
+
else:
|
458 |
+
return {
|
459 |
+
"error": result.get("error", "Unknown error occurred"),
|
460 |
+
"status": "failed to load dataset"
|
461 |
+
}
|
462 |
+
except Exception as e:
|
463 |
+
return {
|
464 |
+
"error": str(e),
|
465 |
+
"status": "failed to load dataset"
|
466 |
+
}
|
467 |
|
468 |
+
def run_evaluation(dataset_choice, splitting_strategy, chunk_size, vector_db, qa_chain):
|
469 |
+
if not evaluator.current_dataset:
|
470 |
+
return pd.DataFrame()
|
471 |
+
|
472 |
+
results = evaluator.evaluate_configuration(
|
473 |
+
vector_db=vector_db,
|
474 |
+
qa_chain=qa_chain,
|
475 |
+
splitting_strategy=splitting_strategy,
|
476 |
+
chunk_size=chunk_size
|
477 |
+
)
|
478 |
+
|
479 |
+
return pd.DataFrame([results])
|
480 |
+
|
481 |
+
load_dataset_btn.click(
|
482 |
+
load_dataset_handler,
|
483 |
+
inputs=[dataset_choice],
|
484 |
+
outputs=[dataset_info]
|
485 |
+
)
|
486 |
+
|
487 |
+
evaluate_btn.click(
|
488 |
+
run_evaluation,
|
489 |
+
inputs=[
|
490 |
+
dataset_choice,
|
491 |
+
eval_splitting_strategy,
|
492 |
+
eval_chunk_size,
|
493 |
+
vector_db,
|
494 |
+
qa_chain
|
495 |
+
],
|
496 |
+
outputs=[evaluation_results]
|
497 |
+
)
|
498 |
+
|
499 |
+
# Clear button handlers
|
500 |
clear_btn.click(
|
501 |
lambda: [None, "", 0, "", 0, "", 0],
|
502 |
+
outputs=[chatbot, source1, page1, source2, page2, source3, page3]
|
|
|
|
|
503 |
)
|
504 |
+
|
505 |
+
# Launch the demo
|
506 |
demo.queue().launch(debug=True)
|
507 |
|
508 |
if __name__ == "__main__":
|