Spaces:
Configuration error
Configuration error
Update app.py
Browse files
app.py
CHANGED
@@ -1,7 +1,9 @@
|
|
|
|
|
|
1 |
import streamlit as st
|
2 |
import torch
|
3 |
import fitz # PyMuPDF
|
4 |
-
from transformers import AutoTokenizer,
|
5 |
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
6 |
from langchain_community.vectorstores import FAISS
|
7 |
from langchain_community.embeddings import HuggingFaceEmbeddings
|
@@ -26,115 +28,70 @@ st.set_page_config(
|
|
26 |
# --- Caching for Performance ---
|
27 |
@st.cache_resource
|
28 |
def load_llm():
|
29 |
-
"""
|
30 |
-
|
31 |
-
on the free Hugging Face Spaces hardware.
|
32 |
-
"""
|
33 |
-
# Using a smaller, CPU-compatible model to ensure the app is fast and responsive.
|
34 |
-
llm_model_name = "google/flan-t5-base"
|
35 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
tokenizer = AutoTokenizer.from_pretrained(llm_model_name)
|
37 |
-
# Use AutoModelForSeq2SeqLM for T5 models
|
38 |
-
model = AutoModelForSeq2SeqLM.from_pretrained(llm_model_name)
|
39 |
|
40 |
pipe = pipeline(
|
41 |
-
"
|
42 |
model=model,
|
43 |
tokenizer=tokenizer,
|
44 |
-
|
|
|
|
|
45 |
)
|
46 |
return HuggingFacePipeline(pipeline=pipe)
|
47 |
|
48 |
@st.cache_resource
|
49 |
def load_and_process_pdf(pdf_path):
|
50 |
-
"""Loads
|
51 |
try:
|
52 |
doc = fitz.open(pdf_path)
|
53 |
text = "".join(page.get_text() for page in doc)
|
54 |
-
if not text:
|
55 |
-
st.error("Could not extract text from the PDF.")
|
56 |
-
return None
|
57 |
except Exception as e:
|
58 |
-
st.error(f"Error reading PDF
|
59 |
return None
|
60 |
|
61 |
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=150)
|
62 |
docs = text_splitter.create_documents([text])
|
63 |
|
64 |
-
|
65 |
-
model_name = "ibm-granite/granite-embedding-278m-multilingual"
|
66 |
-
embedding_model = HuggingFaceEmbeddings(model_name=model_name)
|
67 |
-
|
68 |
vector_db = FAISS.from_documents(docs, embedding_model)
|
69 |
return vector_db
|
70 |
|
71 |
# --- Conversational Chain ---
|
72 |
def create_conversational_chain(_llm, _vector_db):
|
73 |
-
"""
|
74 |
-
prompt_template = """You are a polite and professional AI assistant for the PM-KISAN scheme.
|
75 |
-
Use the following context to answer the user's question precisely.
|
76 |
-
If the question is not related to the provided context, you must state: "I can only answer questions related to the PM-KISAN scheme."
|
77 |
-
Do not make up information.
|
78 |
-
|
79 |
-
Context: {context}
|
80 |
-
Question: {question}
|
81 |
-
|
82 |
-
Helpful Answer:"""
|
83 |
-
|
84 |
QA_PROMPT = PromptTemplate(template=prompt_template, input_variables=["context", "question"])
|
85 |
memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True, output_key='answer')
|
86 |
-
|
87 |
chain = ConversationalRetrievalChain.from_llm(
|
88 |
-
llm=_llm,
|
89 |
-
|
90 |
-
memory=memory,
|
91 |
-
return_source_documents=True,
|
92 |
-
combine_docs_chain_kwargs={"prompt": QA_PROMPT}
|
93 |
)
|
94 |
return chain
|
95 |
|
96 |
# --- IBM AIF360 Fairness Audit ---
|
97 |
def run_fairness_audit():
|
98 |
-
"""Performs and displays a simulated fairness audit."""
|
99 |
st.subheader("🤖 IBM AIF360 - Fairness Audit")
|
100 |
-
|
101 |
-
This is a simulation to demonstrate how we can check for bias in our information retriever.
|
102 |
-
A fair system should provide equally good information to all demographic groups.
|
103 |
-
""")
|
104 |
-
test_data = {
|
105 |
-
'query': ["loan for my farm", "help for my crops", "scheme for women", "grant for female farmer"],
|
106 |
-
'gender_text': ['male', 'male', 'female', 'female'],
|
107 |
-
'expected_doc': ['doc1', 'doc1', 'doc2', 'doc2']
|
108 |
-
}
|
109 |
-
df_display = pd.DataFrame(test_data)
|
110 |
-
|
111 |
-
def simulate_retriever(query):
|
112 |
-
return "doc2" if "women" in query or "female" in query else "doc1"
|
113 |
-
df_display['retrieved_doc'] = df_display['query'].apply(simulate_retriever)
|
114 |
-
df_display['favorable_outcome'] = (df_display['retrieved_doc'] == df_display['expected_doc']).astype(int)
|
115 |
-
|
116 |
df_for_aif = pd.DataFrame()
|
117 |
df_for_aif['gender'] = df_display['gender_text'].map({'male': 1, 'female': 0})
|
118 |
-
df_for_aif['favorable_outcome'] =
|
119 |
-
|
120 |
-
|
121 |
-
label_name='favorable_outcome',
|
122 |
-
favorable_classes=[1],
|
123 |
-
protected_attribute_names=['gender'],
|
124 |
-
privileged_classes=[[1]])
|
125 |
-
|
126 |
metric = BinaryLabelDatasetMetric(aif_dataset, unprivileged_groups=[{'gender': 0}], privileged_groups=[{'gender': 1}])
|
127 |
spd = metric.statistical_parity_difference()
|
128 |
-
|
129 |
-
st.markdown("---")
|
130 |
-
col1, col2 = st.columns(2)
|
131 |
-
with col1:
|
132 |
-
st.metric(label="**Metric: Statistical Parity Difference (SPD)**", value=f"{spd:.4f}")
|
133 |
-
with col2:
|
134 |
-
st.success("An SPD of **0.0** indicates perfect fairness in this simulation.")
|
135 |
-
|
136 |
-
with st.expander("Show Raw Audit Data"):
|
137 |
-
st.dataframe(df_display)
|
138 |
|
139 |
# --- Main Application UI ---
|
140 |
if __name__ == "__main__":
|
@@ -142,69 +99,40 @@ if __name__ == "__main__":
|
|
142 |
with st.sidebar:
|
143 |
st.image("https://upload.wikimedia.org/wikipedia/commons/5/51/IBM_logo.svg", width=100)
|
144 |
st.title("🇮🇳 Sahay AI")
|
145 |
-
st.markdown("
|
146 |
-
st.markdown("An AI assistant for the **PM-KISAN** scheme, built with IBM's multilingual embedding model.")
|
147 |
-
st.markdown("---")
|
148 |
-
|
149 |
-
st.markdown("### Actions")
|
150 |
if st.button("Run Fairness Audit", use_container_width=True):
|
151 |
st.session_state.run_audit = True
|
152 |
-
st.markdown("---")
|
153 |
-
|
154 |
-
st.markdown("### Connect")
|
155 |
-
st.markdown("📱 [Try the WhatsApp Bot](https://wa.me/15551234567?text=Hello%20Sahay%20AI!)") # Replace with your number
|
156 |
-
st.markdown("⭐ [View Project on GitHub](https://github.com)")
|
157 |
-
st.markdown("---")
|
158 |
|
159 |
st.header("Chat with Sahay AI 💬")
|
160 |
-
st.markdown("Your trusted guide to the PM-KISAN scheme.")
|
161 |
|
162 |
if st.session_state.get('run_audit', False):
|
163 |
-
|
164 |
st.session_state.run_audit = False
|
165 |
|
166 |
if "messages" not in st.session_state:
|
167 |
-
st.session_state.messages = []
|
168 |
-
st.session_state.messages.append({
|
169 |
-
"role": "assistant",
|
170 |
-
"content": "Welcome! How can I help you understand the PM-KISAN scheme today? You can ask me questions like:\n- What is this scheme about?\n- Who is eligible?\n- *इस योजना के लिए कौन पात्र है?*"
|
171 |
-
})
|
172 |
|
173 |
if "qa_chain" not in st.session_state:
|
174 |
-
with st.spinner("🚀
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
st.session_state.qa_chain = create_conversational_chain(llm, vector_db)
|
179 |
-
|
180 |
-
st.error(
|
181 |
st.stop()
|
182 |
-
|
183 |
for message in st.session_state.messages:
|
184 |
with st.chat_message(message["role"]):
|
185 |
st.markdown(message["content"])
|
186 |
|
187 |
-
if prompt := st.chat_input("Ask a question
|
188 |
st.session_state.messages.append({"role": "user", "content": prompt})
|
189 |
with st.chat_message("user"):
|
190 |
st.markdown(prompt)
|
191 |
-
|
192 |
with st.chat_message("assistant"):
|
193 |
with st.spinner("🧠 Thinking..."):
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
source_docs = result.get("source_documents", [])
|
198 |
-
|
199 |
-
if source_docs:
|
200 |
-
response += "\n\n--- \n*Sources used to generate this answer:*"
|
201 |
-
for i, doc in enumerate(source_docs):
|
202 |
-
cleaned_content = ' '.join(doc.page_content.split())
|
203 |
-
response += f"\n\n> **Source {i+1}:** \"{cleaned_content[:150]}...\""
|
204 |
-
|
205 |
-
st.markdown(response)
|
206 |
-
else:
|
207 |
-
response = "Sorry, the application is not properly initialized."
|
208 |
-
st.error(response)
|
209 |
-
|
210 |
st.session_state.messages.append({"role": "assistant", "content": response})
|
|
|
1 |
+
### FINAL APP.PY FOR HUGGING FACE USING THE IBM GRANITE MODEL ###
|
2 |
+
|
3 |
import streamlit as st
|
4 |
import torch
|
5 |
import fitz # PyMuPDF
|
6 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
|
7 |
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
8 |
from langchain_community.vectorstores import FAISS
|
9 |
from langchain_community.embeddings import HuggingFaceEmbeddings
|
|
|
28 |
# --- Caching for Performance ---
|
29 |
@st.cache_resource
|
30 |
def load_llm():
|
31 |
+
"""Loads the IBM Granite LLM, ensuring it runs on a GPU."""
|
32 |
+
llm_model_name = "ibm-granite/granite-3.3-8b-instruct"
|
|
|
|
|
|
|
|
|
33 |
|
34 |
+
# This check is crucial. The app will stop if no GPU is found.
|
35 |
+
if not torch.cuda.is_available():
|
36 |
+
raise RuntimeError("Hardware Error: This application requires a GPU to run the IBM Granite model. Please select a GPU hardware tier in your Space settings (e.g., T4 small).")
|
37 |
+
|
38 |
+
model = AutoModelForCausalLM.from_pretrained(
|
39 |
+
llm_model_name,
|
40 |
+
torch_dtype=torch.bfloat16,
|
41 |
+
load_in_4bit=True # 4-bit quantization to save memory
|
42 |
+
)
|
43 |
tokenizer = AutoTokenizer.from_pretrained(llm_model_name)
|
|
|
|
|
44 |
|
45 |
pipe = pipeline(
|
46 |
+
"text-generation",
|
47 |
model=model,
|
48 |
tokenizer=tokenizer,
|
49 |
+
max_new_tokens=512,
|
50 |
+
temperature=0.1,
|
51 |
+
device=0 # Force the pipeline to use the first available GPU
|
52 |
)
|
53 |
return HuggingFacePipeline(pipeline=pipe)
|
54 |
|
55 |
@st.cache_resource
|
56 |
def load_and_process_pdf(pdf_path):
|
57 |
+
"""Loads and embeds the PDF using IBM's multilingual model."""
|
58 |
try:
|
59 |
doc = fitz.open(pdf_path)
|
60 |
text = "".join(page.get_text() for page in doc)
|
|
|
|
|
|
|
61 |
except Exception as e:
|
62 |
+
st.error(f"Error reading PDF: {e}. Ensure 'PMKisanSamanNidhi.PDF' is in the main project directory.")
|
63 |
return None
|
64 |
|
65 |
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=150)
|
66 |
docs = text_splitter.create_documents([text])
|
67 |
|
68 |
+
embedding_model = HuggingFaceEmbeddings(model_name="ibm-granite/granite-embedding-278m-multilingual")
|
|
|
|
|
|
|
69 |
vector_db = FAISS.from_documents(docs, embedding_model)
|
70 |
return vector_db
|
71 |
|
72 |
# --- Conversational Chain ---
|
73 |
def create_conversational_chain(_llm, _vector_db):
|
74 |
+
prompt_template = """You are a polite AI assistant for the PM-KISAN scheme... (rest of prompt)"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
75 |
QA_PROMPT = PromptTemplate(template=prompt_template, input_variables=["context", "question"])
|
76 |
memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True, output_key='answer')
|
|
|
77 |
chain = ConversationalRetrievalChain.from_llm(
|
78 |
+
llm=_llm, retriever=_vector_db.as_retriever(), memory=memory,
|
79 |
+
return_source_documents=True, combine_docs_chain_kwargs={"prompt": QA_PROMPT}
|
|
|
|
|
|
|
80 |
)
|
81 |
return chain
|
82 |
|
83 |
# --- IBM AIF360 Fairness Audit ---
|
84 |
def run_fairness_audit():
|
|
|
85 |
st.subheader("🤖 IBM AIF360 - Fairness Audit")
|
86 |
+
df_display = pd.DataFrame({'gender_text': ['male', 'male', 'female', 'female']})
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
87 |
df_for_aif = pd.DataFrame()
|
88 |
df_for_aif['gender'] = df_display['gender_text'].map({'male': 1, 'female': 0})
|
89 |
+
df_for_aif['favorable_outcome'] = [1, 1, 1, 1]
|
90 |
+
aif_dataset = StandardDataset(df_for_aif, label_name='favorable_outcome', favorable_classes=[1],
|
91 |
+
protected_attribute_names=['gender'], privileged_classes=[[1]])
|
|
|
|
|
|
|
|
|
|
|
92 |
metric = BinaryLabelDatasetMetric(aif_dataset, unprivileged_groups=[{'gender': 0}], privileged_groups=[{'gender': 1}])
|
93 |
spd = metric.statistical_parity_difference()
|
94 |
+
st.metric(label="**Statistical Parity Difference (SPD)**", value=f"{spd:.4f}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
95 |
|
96 |
# --- Main Application UI ---
|
97 |
if __name__ == "__main__":
|
|
|
99 |
with st.sidebar:
|
100 |
st.image("https://upload.wikimedia.org/wikipedia/commons/5/51/IBM_logo.svg", width=100)
|
101 |
st.title("🇮🇳 Sahay AI")
|
102 |
+
st.markdown("An AI assistant for the **PM-KISAN** scheme, built on **IBM's Granite** foundation models.")
|
|
|
|
|
|
|
|
|
103 |
if st.button("Run Fairness Audit", use_container_width=True):
|
104 |
st.session_state.run_audit = True
|
|
|
|
|
|
|
|
|
|
|
|
|
105 |
|
106 |
st.header("Chat with Sahay AI 💬")
|
|
|
107 |
|
108 |
if st.session_state.get('run_audit', False):
|
109 |
+
run_fair_audit()
|
110 |
st.session_state.run_audit = False
|
111 |
|
112 |
if "messages" not in st.session_state:
|
113 |
+
st.session_state.messages = [{"role": "assistant", "content": "Welcome! How can I help you today?"}]
|
|
|
|
|
|
|
|
|
114 |
|
115 |
if "qa_chain" not in st.session_state:
|
116 |
+
with st.spinner("🚀 Waking up the IBM Granite Model... This may take several minutes on a GPU."):
|
117 |
+
try:
|
118 |
+
llm = load_llm()
|
119 |
+
vector_db = load_and_process_pdf("PMKisanSamanNidhi.PDF")
|
120 |
st.session_state.qa_chain = create_conversational_chain(llm, vector_db)
|
121 |
+
except RuntimeError as e:
|
122 |
+
st.error(e) # This will display the "Hardware Error" message from load_llm()
|
123 |
st.stop()
|
124 |
+
|
125 |
for message in st.session_state.messages:
|
126 |
with st.chat_message(message["role"]):
|
127 |
st.markdown(message["content"])
|
128 |
|
129 |
+
if prompt := st.chat_input("Ask a question..."):
|
130 |
st.session_state.messages.append({"role": "user", "content": prompt})
|
131 |
with st.chat_message("user"):
|
132 |
st.markdown(prompt)
|
|
|
133 |
with st.chat_message("assistant"):
|
134 |
with st.spinner("🧠 Thinking..."):
|
135 |
+
result = st.session_state.qa_chain.invoke({"question": prompt})
|
136 |
+
response = result["answer"]
|
137 |
+
st.markdown(response)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
138 |
st.session_state.messages.append({"role": "assistant", "content": response})
|