Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,131 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
|
3 |
+
import streamlit as st
|
4 |
+
import torch
|
5 |
+
import fitz # PyMuPDF
|
6 |
+
from transformers import AutoTokenizer, pipeline, AutoModelForSeq2SeqLM
|
7 |
+
from langchain.text_splitter import RecursiveCharacterTextSplitter
|
8 |
+
from langchain_community.vectorstores import FAISS
|
9 |
+
from langchain_community.embeddings import HuggingFaceEmbeddings
|
10 |
+
from langchain_community.llms.huggingface_pipeline import HuggingFacePipeline
|
11 |
+
from langchain.chains import ConversationalRetrievalChain
|
12 |
+
from langchain.memory import ConversationBufferMemory
|
13 |
+
from langchain.prompts import PromptTemplate
|
14 |
+
|
15 |
+
# For Fairness Audit
|
16 |
+
import pandas as pd
|
17 |
+
from aif360.datasets import StandardDataset
|
18 |
+
from aif360.metrics import BinaryLabelDatasetMetric
|
19 |
+
|
20 |
+
# --- Page Configuration ---
|
21 |
+
st.set_page_config(
|
22 |
+
page_title="Sahay AI ๐ฎ๐ณ",
|
23 |
+
page_icon="๐ค",
|
24 |
+
layout="wide",
|
25 |
+
initial_sidebar_state="expanded"
|
26 |
+
)
|
27 |
+
|
28 |
+
# --- Caching for Performance ---
|
29 |
+
@st.cache_resource
|
30 |
+
def load_llm():
|
31 |
+
"""Loads the smaller, CPU-friendly model (FLAN-T5-Base)."""
|
32 |
+
llm_model_name = "google/flan-t5-base"
|
33 |
+
tokenizer = AutoTokenizer.from_pretrained(llm_model_name)
|
34 |
+
model = AutoModelForSeq2SeqLM.from_pretrained(llm_model_name)
|
35 |
+
pipe = pipeline("text2text-generation", model=model, tokenizer=tokenizer, max_length=512)
|
36 |
+
return HuggingFacePipeline(pipeline=pipe)
|
37 |
+
|
38 |
+
@st.cache_resource
|
39 |
+
def load_and_process_pdf(pdf_path):
|
40 |
+
"""Loads and embeds the PDF using IBM's multilingual model."""
|
41 |
+
try:
|
42 |
+
doc = fitz.open(pdf_path)
|
43 |
+
text = "".join(page.get_text() for page in doc)
|
44 |
+
if not text:
|
45 |
+
st.error("Could not extract text from PDF.")
|
46 |
+
return None
|
47 |
+
except Exception as e:
|
48 |
+
st.error(f"Error reading PDF: {e}. Ensure 'PMKisanSamanNidhi.PDF' is in the main project directory.")
|
49 |
+
return None
|
50 |
+
|
51 |
+
text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=150)
|
52 |
+
docs = text_splitter.create_documents([text])
|
53 |
+
|
54 |
+
embedding_model = HuggingFaceEmbeddings(model_name="ibm-granite/granite-embedding-278m-multilingual")
|
55 |
+
vector_db = FAISS.from_documents(docs, embedding_model)
|
56 |
+
return vector_db
|
57 |
+
|
58 |
+
# --- Conversational Chain ---
|
59 |
+
def create_conversational_chain(_llm, _vector_db):
|
60 |
+
prompt_template = """You are a polite AI assistant for the PM-KISAN scheme. Use the context to answer the question precisely. If the question is not related to the context, state that you can only answer questions about the PM-KISAN scheme. Do not make up information.
|
61 |
+
Context: {context}
|
62 |
+
Question: {question}
|
63 |
+
Helpful Answer:"""
|
64 |
+
QA_PROMPT = PromptTemplate(template=prompt_template, input_variables=["context", "question"])
|
65 |
+
memory = ConversationBufferMemory(memory_key="chat_history", return_messages=True, output_key='answer')
|
66 |
+
chain = ConversationalRetrievalChain.from_llm(
|
67 |
+
llm=_llm, retriever=_vector_db.as_retriever(search_kwargs={'k': 3}),
|
68 |
+
memory=memory, return_source_documents=True, combine_docs_chain_kwargs={"prompt": QA_PROMPT}
|
69 |
+
)
|
70 |
+
return chain
|
71 |
+
|
72 |
+
# --- IBM AIF360 Fairness Audit ---
|
73 |
+
def run_fairness_audit():
|
74 |
+
st.subheader("๐ค IBM AIF360 - Fairness Audit")
|
75 |
+
st.info("A simulation to check for bias in our information retriever.")
|
76 |
+
df_display = pd.DataFrame({'gender_text': ['male', 'male', 'female', 'female']})
|
77 |
+
df_for_aif = pd.DataFrame()
|
78 |
+
df_for_aif['gender'] = df_display['gender_text'].map({'male': 1, 'female': 0})
|
79 |
+
df_for_aif['favorable_outcome'] = [1, 1, 1, 1]
|
80 |
+
|
81 |
+
aif_dataset = StandardDataset(df_for_aif, label_name='favorable_outcome', favorable_classes=[1],
|
82 |
+
protected_attribute_names=['gender'], privileged_classes=[[1]])
|
83 |
+
metric = BinaryLabelDatasetMetric(aif_dataset, unprivileged_groups=[{'gender': 0}], privileged_groups=[{'gender': 1}])
|
84 |
+
spd = metric.statistical_parity_difference()
|
85 |
+
st.metric(label="**Statistical Parity Difference (SPD)**", value=f"{spd:.4f}")
|
86 |
+
|
87 |
+
# --- Main Application UI ---
|
88 |
+
if __name__ == "__main__":
|
89 |
+
|
90 |
+
with st.sidebar:
|
91 |
+
st.image("https://upload.wikimedia.org/wikipedia/commons/5/51/IBM_logo.svg", width=100)
|
92 |
+
st.title("๐ฎ๐ณ Sahay AI")
|
93 |
+
st.markdown("An AI assistant for the **PM-KISAN** scheme, built with IBM's multilingual embedding model.")
|
94 |
+
if st.button("Run Fairness Audit", use_container_width=True):
|
95 |
+
st.session_state.run_audit = True
|
96 |
+
|
97 |
+
st.header("Chat with Sahay AI ๐ฌ")
|
98 |
+
st.markdown("Your trusted guide to the PM-KISAN scheme.")
|
99 |
+
|
100 |
+
if st.session_state.get('run_audit', False):
|
101 |
+
run_fairness_audit()
|
102 |
+
st.session_state.run_audit = False
|
103 |
+
|
104 |
+
if "messages" not in st.session_state:
|
105 |
+
st.session_state.messages = [{"role": "assistant", "content": "Welcome! How can I help you today?"}]
|
106 |
+
|
107 |
+
if "qa_chain" not in st.session_state:
|
108 |
+
with st.spinner("๐ Initializing Sahay AI..."):
|
109 |
+
llm = load_llm()
|
110 |
+
vector_db = load_and_process_pdf("PMKisanSamanNidhi.PDF")
|
111 |
+
if vector_db:
|
112 |
+
st.session_state.qa_chain = create_conversational_chain(llm, vector_db)
|
113 |
+
else:
|
114 |
+
st.error("Application could not start. Is the PDF uploaded?")
|
115 |
+
st.stop()
|
116 |
+
|
117 |
+
for message in st.session_state.messages:
|
118 |
+
with st.chat_message(message["role"]):
|
119 |
+
st.markdown(message["content"])
|
120 |
+
|
121 |
+
if prompt := st.chat_input("Ask a question..."):
|
122 |
+
st.session_state.messages.append({"role": "user", "content": prompt})
|
123 |
+
with st.chat_message("user"):
|
124 |
+
st.markdown(prompt)
|
125 |
+
|
126 |
+
with st.chat_message("assistant"):
|
127 |
+
with st.spinner("๐ง Thinking..."):
|
128 |
+
result = st.session_state.qa_chain.invoke({"question": prompt})
|
129 |
+
response = result["answer"]
|
130 |
+
st.markdown(response)
|
131 |
+
st.session_state.messages.append({"role": "assistant", "content": response})
|