Spaces:
Sleeping
Sleeping
File size: 16,198 Bytes
1bbe15b | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 | """
VDHF - Streamlit Interface
Verification-Driven Hallucination Firewall
"""
import os
import sys
import time
import streamlit as st
# Add project root to path
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
DATA_DIR = os.path.join(os.path.dirname(__file__), "data", "sample_docs")
# βββ Page Config βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
st.set_page_config(
page_title="VDHF - Hallucination Firewall",
page_icon="π‘οΈ",
layout="wide",
initial_sidebar_state="expanded",
)
# βββ Custom CSS ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
st.markdown("""
<style>
.main-header {
font-size: 2.2rem;
font-weight: 700;
color: #1E3A5F;
text-align: center;
padding: 0.5rem 0 0.2rem 0;
}
.sub-header {
font-size: 1rem;
color: #6B7280;
text-align: center;
margin-bottom: 1.5rem;
}
.metric-card {
background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
padding: 1.2rem;
border-radius: 12px;
color: white;
text-align: center;
}
.metric-card h3 { margin: 0; font-size: 2rem; }
.metric-card p { margin: 0; font-size: 0.85rem; opacity: 0.9; }
.claim-supported {
background-color: #D1FAE5;
border-left: 4px solid #10B981;
padding: 0.7rem 1rem;
border-radius: 6px;
margin: 0.4rem 0;
}
.claim-unsupported {
background-color: #FEE2E2;
border-left: 4px solid #EF4444;
padding: 0.7rem 1rem;
border-radius: 6px;
margin: 0.4rem 0;
}
.evidence-box {
background-color: #EFF6FF;
border-left: 4px solid #3B82F6;
padding: 0.7rem 1rem;
border-radius: 6px;
margin: 0.4rem 0;
font-size: 0.9rem;
}
.status-pass {
background: linear-gradient(135deg, #10B981 0%, #059669 100%);
color: white;
padding: 0.8rem 1.5rem;
border-radius: 10px;
text-align: center;
font-size: 1.1rem;
font-weight: 600;
}
.status-fail {
background: linear-gradient(135deg, #F59E0B 0%, #D97706 100%);
color: white;
padding: 0.8rem 1.5rem;
border-radius: 10px;
text-align: center;
font-size: 1.1rem;
font-weight: 600;
}
</style>
""", unsafe_allow_html=True)
# βββ Pipeline Init (cached) βββββββββββββββββββββββββββββββββββββββββββββββββ
@st.cache_resource(show_spinner=False)
def load_pipeline():
"""Initialize the VDHF pipeline and load documents."""
from core.pipeline import VDHFPipeline
pipeline = VDHFPipeline()
# Load sample documents
if os.path.exists(DATA_DIR):
for filename in os.listdir(DATA_DIR):
if filename.endswith('.txt'):
filepath = os.path.join(DATA_DIR, filename)
pipeline.ingest_file(filepath)
return pipeline
# βββ Sidebar βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
with st.sidebar:
st.markdown("## βοΈ Configuration")
similarity_threshold = st.slider(
"Similarity Threshold (ΞΈ_sim)",
min_value=0.0, max_value=1.0, value=0.75, step=0.05,
help="Minimum semantic similarity for a claim to be considered supported."
)
firewall_threshold = st.slider(
"Firewall Threshold (Ο)",
min_value=0.0, max_value=1.0, value=0.80, step=0.05,
help="Minimum support ratio to pass the firewall."
)
top_k = st.slider(
"Top-K Evidence Chunks",
min_value=1, max_value=15, value=7,
help="Number of evidence chunks to retrieve."
)
st.markdown("---")
st.markdown("## π Knowledge Base")
# Custom document upload
uploaded_file = st.file_uploader(
"Upload a document (.txt)", type=["txt"],
help="Add your own document to the knowledge base."
)
st.markdown("---")
st.markdown("## βΉοΈ How It Works")
st.markdown("""
1. **Retrieve** relevant evidence from documents
2. **Generate** an LLM response using context
3. **Extract** atomic factual claims
4. **Verify** each claim against evidence
5. **Firewall** blocks hallucinated responses
6. **Regenerate** using only verified evidence
""")
# βββ Header ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
st.markdown('<div class="main-header">π‘οΈ Hallucination Firewall</div>', unsafe_allow_html=True)
st.markdown('<div class="sub-header">Verification-Driven Hallucination Firewall for RAG Systems</div>', unsafe_allow_html=True)
# βββ Load Pipeline βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
with st.spinner("π Loading models and documents... (first load may take a moment)"):
pipeline = load_pipeline()
# Handle file upload
if uploaded_file is not None:
content = uploaded_file.read().decode("utf-8")
if f"uploaded_{uploaded_file.name}" not in st.session_state:
pipeline.ingest_text(content, source=uploaded_file.name)
st.session_state[f"uploaded_{uploaded_file.name}"] = True
st.sidebar.success(f"β
Uploaded: {uploaded_file.name}")
# Show doc count
st.sidebar.metric("Document Chunks Loaded", pipeline.document_count)
# βββ Tabs ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
tab_query, tab_analyze, tab_about = st.tabs(["π Query", "π§ͺ Analyze Claims", "π About"])
# βββ TAB 1: Query ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
with tab_query:
query = st.text_input(
"Ask a question about the knowledge base:",
placeholder="e.g., When was Python released and who created it?",
)
col_btn, col_examples = st.columns([1, 3])
with col_btn:
run_query = st.button("π Run Query", type="primary", use_container_width=True)
with col_examples:
example = st.selectbox("Or try an example:", [
"",
"When was Python released and who created it?",
"What caused World War I?",
"Tell me about artificial intelligence history.",
"How does the human body work?",
"What is climate change and what causes it?",
"Tell me about the Renaissance period.",
"How did the internet develop?",
], label_visibility="collapsed")
if example and not query:
query = example
run_query = True
if run_query and query:
# Update pipeline thresholds
pipeline.similarity_threshold = similarity_threshold
pipeline.firewall_threshold = firewall_threshold
pipeline.top_k = top_k
pipeline.verifier.similarity_threshold = similarity_threshold
pipeline.firewall.similarity_threshold = similarity_threshold
pipeline.firewall.decision_engine.threshold = firewall_threshold
pipeline.firewall.decision_engine.scoring_module.threshold = firewall_threshold
with st.spinner("Processing query through the VDHF pipeline..."):
start_time = time.time()
result = pipeline.query(query, verbose=False)
elapsed = time.time() - start_time
# ββ Status Banner ββ
if result.is_verified:
st.markdown(
f'<div class="status-pass">β
VERIFIED β Support Ratio: {result.support_ratio:.0%} '
f'({result.supported_claims}/{result.total_claims} claims supported)</div>',
unsafe_allow_html=True
)
else:
st.markdown(
f'<div class="status-fail">β οΈ PARTIALLY VERIFIED β Support Ratio: {result.support_ratio:.0%} '
f'({result.supported_claims}/{result.total_claims} claims supported)</div>',
unsafe_allow_html=True
)
st.markdown("")
# ββ Metrics Row ββ
m1, m2, m3, m4 = st.columns(4)
m1.metric("Support Ratio", f"{result.support_ratio:.0%}")
m2.metric("Total Claims", result.total_claims)
m3.metric("Supported", result.supported_claims)
m4.metric("Regenerations", result.regeneration_attempts)
st.markdown("---")
# ββ Response ββ
st.subheader("π Response")
st.info(result.final_response)
st.caption(f"β±οΈ Processed in {elapsed:.2f}s")
# ββ Claims Breakdown ββ
if result.verification_results:
st.subheader("π¬ Claims Verification")
for vr in result.verification_results:
if vr.is_supported:
st.markdown(
f'<div class="claim-supported">'
f'<strong>β
SUPPORTED</strong> (similarity: {vr.similarity_score:.3f}, '
f'entailment: {vr.entailment_label})<br/>'
f'{vr.claim.text}</div>',
unsafe_allow_html=True
)
else:
st.markdown(
f'<div class="claim-unsupported">'
f'<strong>β UNSUPPORTED</strong> (similarity: {vr.similarity_score:.3f}, '
f'entailment: {vr.entailment_label})<br/>'
f'{vr.claim.text}</div>',
unsafe_allow_html=True
)
# ββ Retrieved Evidence ββ
if result.retrieved_evidence:
with st.expander(f"π Retrieved Evidence ({len(result.retrieved_evidence)} chunks)", expanded=False):
for i, ev in enumerate(result.retrieved_evidence, 1):
source = os.path.basename(ev.metadata.get("source", "Unknown"))
st.markdown(
f'<div class="evidence-box">'
f'<strong>[{i}]</strong> Score: {ev.similarity_score:.3f} | Source: {source}<br/>'
f'{ev.content[:300]}{"..." if len(ev.content) > 300 else ""}</div>',
unsafe_allow_html=True
)
# βββ TAB 2: Analyze Claims βββββββββββββββββββββββββββββββββββββββββββββββββββ
with tab_analyze:
st.subheader("Test Custom Claims Against the Knowledge Base")
st.markdown("Enter individual claims to verify them against the loaded documents.")
claims_input = st.text_area(
"Enter claims (one per line):",
placeholder="Python was created by Guido van Rossum.\nPython was released in 2005.\nPython is a compiled language.",
height=150,
)
if st.button("π Verify Claims", type="primary"):
if claims_input.strip():
from core.claim_extractor import Claim
from retrieval.retriever import RetrievedEvidence
lines = [l.strip() for l in claims_input.strip().split("\n") if l.strip()]
with st.spinner("Verifying claims..."):
# Retrieve evidence for all claims combined
combined_query = " ".join(lines)
evidence_list = pipeline.retriever.retrieve(combined_query, top_k=top_k)
claims = [Claim(text=line, claim_id=i) for i, line in enumerate(lines)]
results = pipeline.verifier.verify_all_claims(claims, evidence_list)
supported = sum(1 for r in results if r.is_supported)
total = len(results)
ratio = supported / total if total > 0 else 0
st.markdown(f"**Results: {supported}/{total} claims supported ({ratio:.0%})**")
st.progress(ratio)
for vr in results:
if vr.is_supported:
st.markdown(
f'<div class="claim-supported">'
f'<strong>β
SUPPORTED</strong> (score: {vr.similarity_score:.3f})<br/>'
f'{vr.claim.text}</div>',
unsafe_allow_html=True
)
else:
st.markdown(
f'<div class="claim-unsupported">'
f'<strong>β UNSUPPORTED</strong> (score: {vr.similarity_score:.3f})<br/>'
f'{vr.claim.text}</div>',
unsafe_allow_html=True
)
if vr.best_evidence:
with st.expander(f"Evidence for: {vr.claim.text[:50]}..."):
st.markdown(
f'<div class="evidence-box">{vr.best_evidence[:500]}</div>',
unsafe_allow_html=True
)
else:
st.warning("Please enter at least one claim.")
# βββ TAB 3: About ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
with tab_about:
st.subheader("About VDHF")
st.markdown("""
The **Verification-Driven Hallucination Firewall (VDHF)** is a post-generation
verification system that detects and mitigates hallucinations in LLM-generated responses.
### Architecture
```
User Query
β
βΌ
βββββββββββββββββββ
β RAG Retrieval β β Sentence-BERT + ChromaDB
ββββββββββ¬βββββββββ
βΌ
βββββββββββββββββββ
β LLM Generation β β Groq API / Mock
ββββββββββ¬βββββββββ
βΌ
βββββββββββββββββββ
β Claim Extraction β β Rule-based decomposition
ββββββββββ¬βββββββββ
βΌ
βββββββββββββββββββ
β Verification β β Semantic Similarity + NLI
ββββββββββ¬βββββββββ
βΌ
βββββββββββββββββββ
β Firewall β β Support Ratio β₯ Ο ?
ββββββββββ¬βββββββββ
βββββ΄ββββ
βΌ βΌ
PASS REGENERATE
```
### Key Parameters
| Parameter | Default | Description |
|-----------|---------|-------------|
| Similarity Threshold (ΞΈ_sim) | 0.75 | Min cosine similarity for support |
| Firewall Threshold (Ο) | 0.80 | Min support ratio to pass |
| Top-K | 7 | Evidence chunks retrieved |
| Max Regenerations | 2 | Retry attempts on failure |
### Models Used
- **Embeddings**: `all-MiniLM-L6-v2` (Sentence-BERT)
- **NLI**: `microsoft/deberta-base-mnli`
- **LLM**: `llama-3.3-70b-versatile` (via Groq API)
### Knowledge Base
The system comes preloaded with 12 sample documents covering:
Python, Ancient Egypt, AI, Climate Change, Economics, Human Body,
Internet Technology, Music History, Quantum Physics, Renaissance,
Solar System, and World War II.
""")
|