Varun2007 commited on
Commit
c9e39ab
Β·
1 Parent(s): d3d09d6

Configure native Gradio SDK for Hugging Face Spaces deployment

Browse files
Files changed (4) hide show
  1. README.md +25 -38
  2. app.py +404 -455
  3. requirements.txt +1 -0
  4. streamlit_app.py +502 -0
README.md CHANGED
@@ -3,8 +3,8 @@ title: Production Document Question Answering (DQA)
3
  emoji: πŸ“„
4
  colorFrom: blue
5
  colorTo: indigo
6
- sdk: streamlit
7
- sdk_version: "1.51.0"
8
  app_file: app.py
9
  pinned: false
10
  license: mit
@@ -74,7 +74,8 @@ flowchart TD
74
  | [`generate.py`](file:///home/ramakrishna/Desktop/Projects/DQA_Project/generate.py) | Strict context-grounded generator quoting evidence and generating `[Page X, Section Y]` citations. Multi-backend: Hugging Face Serverless Inference API, OpenAI-compatible APIs, and offline deterministic grounding engine. |
75
  | [`verify.py`](file:///home/ramakrishna/Desktop/Projects/DQA_Project/verify.py) | Second-pass verifier extracting atomic factual claims and validating them against retrieved passages via cross-encoder entailment, flagging or removing unsupported claims, and reporting Faithfulness Scores. |
76
  | [`evaluate.py`](file:///home/ramakrishna/Desktop/Projects/DQA_Project/evaluate.py) | 50-pair benchmark evaluation suite testing Recall@k, Answer Faithfulness, Answer Correctness, and Hallucination Rate across 4 baseline/proposed configurations. |
77
- | [`app.py`](file:///home/ramakrishna/Desktop/Projects/DQA_Project/app.py) | Streamlit interactive web application with file upload, real-time token streaming, citation inspector, verification audit display, prompt chips, and pipeline ablation toggle. |
 
78
 
79
  ---
80
 
@@ -91,52 +92,38 @@ The system was evaluated against a rigorous test set of **50 Q&A pairs** across
91
 
92
  ---
93
 
94
- ## πŸ€— Deploying to Hugging Face Spaces
95
 
96
- This repository is pre-configured for 1-click deployment on **Hugging Face Spaces** using Streamlit!
97
 
98
- ### Option A: Using Git (Recommended)
99
- 1. Go to [huggingface.co/new-space](https://huggingface.co/new-space).
100
- 2. Choose **Streamlit** as the Space SDK and choose a name (e.g. `document-qa-system`).
101
- 3. Clone your new Hugging Face Space repository to your machine or push directly:
102
  ```bash
103
- git remote add space https://huggingface.co/spaces/<your-username>/<your-space-name>
104
- git add .
105
- git commit -m "Deploy Production Document QA system"
106
- git push space main
107
  ```
108
 
109
- ### Option B: Using the Hugging Face CLI
 
 
 
110
  ```bash
111
- pip install huggingface_hub
112
- huggingface-cli login
113
- huggingface-cli upload <your-username>/<your-space-name> . --repo-type=space
 
114
  ```
115
 
116
- ### Hugging Face Secrets (Optional)
117
- If you wish to use Hugging Face Serverless Inference API with larger open LLMs (like `Qwen/Qwen2.5-7B-Instruct` or `meta-llama/Llama-3.2-3B-Instruct`), add your token in **Space Settings $\rightarrow$ Variables and secrets**:
118
- - Secret Name: `HF_TOKEN`
119
- - Secret Value: Your Hugging Face user access token (`hf_...`)
120
-
121
- *(Even without any secret token, the system runs with 100% functionality using the high-precision offline grounding engine!)*
122
-
123
  ---
124
 
125
- ## πŸ› οΈ Local Setup & Quickstart
126
 
 
127
  ```bash
128
- # 1. Install system prerequisites (Ubuntu / Debian)
129
- sudo apt-get update && sudo apt-get install -y tesseract-ocr poppler-utils
130
-
131
- # 2. Install Python requirements
132
- pip install -r requirements.txt
133
-
134
- # 3. Generate sample documents
135
- python3 create_sample_docs.py
136
-
137
- # 4. Run tests
138
- python3 -m unittest discover -s tests -p "test_*.py"
139
 
140
- # 5. Launch web application
141
- streamlit run app.py
 
142
  ```
 
 
3
  emoji: πŸ“„
4
  colorFrom: blue
5
  colorTo: indigo
6
+ sdk: gradio
7
+ sdk_version: "6.32.0"
8
  app_file: app.py
9
  pinned: false
10
  license: mit
 
74
  | [`generate.py`](file:///home/ramakrishna/Desktop/Projects/DQA_Project/generate.py) | Strict context-grounded generator quoting evidence and generating `[Page X, Section Y]` citations. Multi-backend: Hugging Face Serverless Inference API, OpenAI-compatible APIs, and offline deterministic grounding engine. |
75
  | [`verify.py`](file:///home/ramakrishna/Desktop/Projects/DQA_Project/verify.py) | Second-pass verifier extracting atomic factual claims and validating them against retrieved passages via cross-encoder entailment, flagging or removing unsupported claims, and reporting Faithfulness Scores. |
76
  | [`evaluate.py`](file:///home/ramakrishna/Desktop/Projects/DQA_Project/evaluate.py) | 50-pair benchmark evaluation suite testing Recall@k, Answer Faithfulness, Answer Correctness, and Hallucination Rate across 4 baseline/proposed configurations. |
77
+ | [`app.py`](file:///home/ramakrishna/Desktop/Projects/DQA_Project/app.py) | Gradio interactive web application with file upload, real-time token streaming, citation accordion, verification audit display, and prompt chips. (Default for Hugging Face Spaces). |
78
+ | [`streamlit_app.py`](file:///home/ramakrishna/Desktop/Projects/DQA_Project/streamlit_app.py) | Alternative Streamlit web interface for users who prefer Streamlit. |
79
 
80
  ---
81
 
 
92
 
93
  ---
94
 
95
+ ## πŸ€— Deploying to Hugging Face Spaces (Gradio SDK)
96
 
97
+ This repository is pre-configured for **Gradio on Hugging Face Spaces**:
98
 
99
+ ### Quick Deployment via Script
 
 
 
100
  ```bash
101
+ ./deploy_to_hf.sh <your-username>/<your-space-name>
 
 
 
102
  ```
103
 
104
+ ### Manual Git Deployment
105
+ 1. Go to [huggingface.co/new-space](https://huggingface.co/new-space).
106
+ 2. Choose a name and select **Gradio** as the Space SDK.
107
+ 3. Add the remote and push:
108
  ```bash
109
+ git remote add space https://huggingface.co/spaces/<your-username>/<your-space-name>.git
110
+ git add .
111
+ git commit -m "Deploy Document QA with Gradio on Hugging Face Spaces"
112
+ git push -u space main
113
  ```
114
 
 
 
 
 
 
 
 
115
  ---
116
 
117
+ ## πŸ› οΈ Running Locally
118
 
119
+ ### Gradio (Default)
120
  ```bash
121
+ python3 app.py
122
+ ```
123
+ Open `http://localhost:7860` in your browser.
 
 
 
 
 
 
 
 
124
 
125
+ ### Streamlit (Alternative)
126
+ ```bash
127
+ streamlit run streamlit_app.py
128
  ```
129
+ Open `http://localhost:8501` in your browser.
app.py CHANGED
@@ -1,27 +1,26 @@
1
- """Streamlit Web Application for Production Document Question Answering (DQA).
2
 
3
- Ready for Deployment on Hugging Face Spaces.
4
  Features:
5
- - Auto-loads sample documents for instant one-click demonstration.
6
- - Upload & Ingest PDF (layout-aware, OCR fallback), DOCX, TXT.
7
  - Hybrid Retrieval: BAAI/bge-small-en-v1.5 + BM25Okapi + Reciprocal Rank Fusion (RRF).
8
- - Cross-Encoder Reranking (ms-marco-MiniLM-L-6-v2).
9
- - Multi-Model Generation:
10
- 1. Hugging Face Serverless Inference API (Qwen-2.5, Llama-3.2, Mistral)
11
- 2. High-Precision Offline Grounding Engine (0-cost, 100% cited)
12
- 3. OpenAI / Groq / DeepSeek API
13
  - Strict Grounding with Evidence Quotes & [Page X, Section Y] Citations.
14
- - Second-Pass Verification Audit with Faithfulness scoring.
15
- - Quick-question prompt chips for instant testing.
16
  """
17
 
18
  import os
19
  import sys
20
- import tempfile
21
  import time
22
- from typing import List
23
 
24
- import streamlit as st
25
 
26
  # Local modules
27
  from config import DQAConfig, default_config
@@ -31,307 +30,148 @@ from generate import Generator, OpenAILLMClient, HuggingFaceInferenceClient
31
  from verify import AnswerVerifier
32
 
33
 
34
- st.set_page_config(
35
- page_title="Document QA System - Production RAG",
36
- page_icon="πŸ“„",
37
- layout="wide",
38
- initial_sidebar_state="expanded"
 
 
39
  )
 
 
 
 
 
 
 
 
 
 
 
40
 
41
- # Custom CSS for modern styling
42
- st.markdown("""
43
- <style>
44
- .main-header {
45
- font-size: 2.2rem;
46
- font-weight: 700;
47
- color: #1E3A8A;
48
- margin-bottom: 0.2rem;
49
- }
50
- .sub-header {
51
- font-size: 1.05rem;
52
- color: #4B5563;
53
- margin-bottom: 1.2rem;
54
- }
55
- .citation-box {
56
- background-color: #F8FAFC;
57
- border-left: 4px solid #3B82F6;
58
- padding: 0.8rem 1rem;
59
- border-radius: 4px;
60
- margin-top: 0.6rem;
61
- font-size: 0.9rem;
62
- }
63
- .verified-pill {
64
- display: inline-block;
65
- padding: 3px 10px;
66
- border-radius: 12px;
67
- font-size: 0.82rem;
68
- font-weight: 600;
69
- background-color: #DEF7EC;
70
- color: #03543F;
71
- }
72
- .unverified-pill {
73
- display: inline-block;
74
- padding: 3px 10px;
75
- border-radius: 12px;
76
- font-size: 0.82rem;
77
- font-weight: 600;
78
- background-color: #FDE8E8;
79
- color: #9B1C1C;
80
- }
81
- .chip-btn {
82
- margin: 2px;
83
- }
84
- </style>
85
- """, unsafe_allow_html=True)
86
 
87
 
88
- @st.cache_resource(show_spinner="Loading Neural Embedder & Cross-Encoder Models...")
89
- def load_core_components(embedding_model_name: str):
90
- """Initializes and caches the heavy neural models."""
91
- retriever = HybridRetriever(
92
- embedding_model=embedding_model_name,
93
- cross_encoder_model=default_config.cross_encoder_model_name,
94
- rrf_k=default_config.rrf_k
95
- )
96
- verifier = AnswerVerifier(
97
- cross_encoder=retriever.reranker,
98
- threshold=default_config.verification_threshold,
99
- flag_unsupported=default_config.flag_unsupported_claims
100
- )
101
- pipeline = DocumentIngestionPipeline(
102
- chunk_size=default_config.chunk_size_tokens,
103
- chunk_overlap=default_config.chunk_overlap_tokens,
104
- ocr_fallback=default_config.ocr_fallback_enabled,
105
- ocr_min_chars=default_config.ocr_min_chars_per_page
106
- )
107
- return retriever, verifier, pipeline
108
-
109
-
110
- # Initialize session state variables
111
- if "embedding_model" not in st.session_state:
112
- st.session_state.embedding_model = default_config.embedding_model_name
113
-
114
- if "retriever" not in st.session_state:
115
- retriever, verifier, pipeline = load_core_components(st.session_state.embedding_model)
116
- st.session_state.retriever = retriever
117
- st.session_state.verifier = verifier
118
- st.session_state.pipeline = pipeline
119
- st.session_state.generator = Generator(verifier=verifier)
120
- st.session_state.messages = []
121
- st.session_state.indexed_files = []
122
- st.session_state.total_chunks = 0
123
- st.session_state.all_chunks = []
124
-
125
- # Auto-load pre-built sample documents on first launch for Hugging Face Spaces visitors
126
- sample_paths = [
127
  "data/sample_financial_report.pdf",
128
  "data/sample_cloud_architecture.docx",
129
  "data/sample_security_policy.txt"
130
  ]
131
- parsed_list = []
132
- for p in sample_paths:
133
- if os.path.exists(p):
134
- parsed = pipeline.ingest_file(p)
135
- parsed_list.append(parsed)
136
- st.session_state.indexed_files.append(parsed.file_name)
137
-
138
- if parsed_list:
139
- retriever.index_documents(parsed_list)
140
- st.session_state.total_chunks = len(retriever.chunks)
141
- st.session_state.all_chunks = retriever.chunks
142
-
143
- retriever = st.session_state.retriever
144
- verifier = st.session_state.verifier
145
- pipeline = st.session_state.pipeline
146
- generator = st.session_state.generator
147
-
148
-
149
- # ---------------- SIDEBAR: Controls & Configuration ----------------
150
- with st.sidebar:
151
- st.header("βš™οΈ Settings & Ingestion")
152
-
153
- # Embedding Model selection
154
- st.subheader("1. Embedding Model")
155
- selected_embedder = st.selectbox(
156
- "Dense Embedding Model",
157
- options=["BAAI/bge-small-en-v1.5", "sentence-transformers/all-MiniLM-L6-v2"],
158
- index=0 if st.session_state.embedding_model == "BAAI/bge-small-en-v1.5" else 1,
159
- help="bge-small-en-v1.5 is the top MTEB benchmark retriever. all-MiniLM-L6-v2 is ultra fast."
160
- )
161
- if selected_embedder != st.session_state.embedding_model:
162
- st.session_state.embedding_model = selected_embedder
163
- with st.spinner("Switching embedding model..."):
164
- retriever, verifier, pipeline = load_core_components(selected_embedder)
165
- st.session_state.retriever = retriever
166
- st.session_state.verifier = verifier
167
- st.session_state.pipeline = pipeline
168
- generator.verifier = verifier
169
- if st.session_state.all_chunks:
170
- retriever.index_documents(st.session_state.all_chunks)
171
- st.rerun()
172
-
173
- # Document Upload Section
174
- st.divider()
175
- st.subheader("2. Upload & Ingest Documents")
176
- uploaded_files = st.file_uploader(
177
- "Upload PDF (with OCR), DOCX, or TXT",
178
- type=["pdf", "docx", "txt", "md"],
179
- accept_multiple_files=True
180
- )
181
-
182
- col1, col2 = st.columns(2)
183
- with col1:
184
- load_sample = st.button("πŸ“‚ Reload Samples", use_container_width=True)
185
- with col2:
186
- reindex = st.button("πŸ”„ Clear Index", use_container_width=True)
187
-
188
- if reindex:
189
- retriever.index_documents([])
190
- st.session_state.indexed_files = []
191
- st.session_state.total_chunks = 0
192
- st.session_state.all_chunks = []
193
- st.session_state.messages = []
194
- generator.memory.clear()
195
- st.success("Vector index & conversation memory cleared!")
196
- st.rerun()
197
-
198
- if load_sample:
199
- sample_paths = [
200
- "data/sample_financial_report.pdf",
201
- "data/sample_cloud_architecture.docx",
202
- "data/sample_security_policy.txt"
203
- ]
204
- parsed_list = []
205
- st.session_state.indexed_files = []
206
- for p in sample_paths:
207
- if os.path.exists(p):
208
- parsed = pipeline.ingest_file(p)
209
- parsed_list.append(parsed)
210
- st.session_state.indexed_files.append(parsed.file_name)
211
-
212
- if parsed_list:
213
- retriever.index_documents(parsed_list)
214
- st.session_state.total_chunks = len(retriever.chunks)
215
- st.session_state.all_chunks = retriever.chunks
216
- st.success(f"Indexed {len(parsed_list)} sample documents ({st.session_state.total_chunks} chunks)!")
217
- st.rerun()
218
-
219
- # Process user uploaded files
220
- if uploaded_files:
221
- if st.button("πŸš€ Process Uploaded Files", type="primary", use_container_width=True):
222
- new_parsed = []
223
- with st.spinner("Extracting layout, formatting tables & generating embeddings..."):
224
- for uploaded_file in uploaded_files:
225
- ext = os.path.splitext(uploaded_file.name)[1]
226
- with tempfile.NamedTemporaryFile(delete=False, suffix=ext) as tmp_file:
227
- tmp_file.write(uploaded_file.read())
228
- tmp_path = tmp_file.name
229
-
230
- try:
231
- parsed = pipeline.ingest_file(tmp_path)
232
- parsed.file_name = uploaded_file.name
233
- for c in parsed.chunks:
234
- c.file_name = uploaded_file.name
235
- new_parsed.append(parsed)
236
- if uploaded_file.name not in st.session_state.indexed_files:
237
- st.session_state.indexed_files.append(uploaded_file.name)
238
- finally:
239
- if os.path.exists(tmp_path):
240
- os.remove(tmp_path)
241
-
242
- if new_parsed:
243
- existing_chunks = list(retriever.chunks)
244
- combined = existing_chunks + [c for p in new_parsed for c in p.chunks]
245
- retriever.index_documents(combined)
246
- st.session_state.total_chunks = len(retriever.chunks)
247
- st.session_state.all_chunks = retriever.chunks
248
- st.success(f"Successfully indexed {len(new_parsed)} document(s)!")
249
- st.rerun()
250
-
251
- # Index Status Summary
252
- st.divider()
253
- st.markdown("**Index Status:**")
254
- st.write(f"β€’ **Documents Indexed:** `{len(st.session_state.indexed_files)}`")
255
- st.write(f"β€’ **Total Chunks:** `{st.session_state.total_chunks}`")
256
- if st.session_state.indexed_files:
257
- with st.expander("πŸ“ View Loaded Documents"):
258
- for fname in st.session_state.indexed_files:
259
- st.caption(f"βœ“ {fname}")
260
-
261
- # Chunk Inspector
262
- if st.session_state.all_chunks:
263
- with st.expander("πŸ” Inspect Chunks & Metadata"):
264
- chunk_num = st.selectbox(
265
- "Select Chunk",
266
- range(len(st.session_state.all_chunks)),
267
- format_func=lambda i: f"Chunk {i+1}: {st.session_state.all_chunks[i].file_name} (p. {st.session_state.all_chunks[i].page_number})"
268
- )
269
- selected = st.session_state.all_chunks[chunk_num]
270
- st.markdown(f"**Doc:** `{selected.file_name}` | **Page {selected.page_number}**")
271
- st.markdown(f"**Section:** `{selected.section_title}` | Table: `{selected.metadata.get('is_table', False)}`")
272
- st.text_area("Chunk Content", selected.text, height=120)
273
-
274
- # Retrieval Settings
275
- st.divider()
276
- st.subheader("3. Retrieval & Reranker")
277
- retrieval_mode = st.selectbox(
278
- "Search Pipeline",
279
- options=["hybrid_rerank", "hybrid_no_rerank", "dense_only", "bm25_only"],
280
- format_func=lambda m: {
281
- "hybrid_rerank": "🌟 Hybrid (Dense+BM25) + Cross-Encoder Rerank",
282
- "hybrid_no_rerank": "⚑ Hybrid Search (No Reranker)",
283
- "dense_only": "🎯 Dense Vector Only (FAISS)",
284
- "bm25_only": "πŸ”€ BM25 Lexical Only"
285
- }[m]
286
- )
287
 
288
- rerank_top_k = st.slider("Top-K Context Chunks", min_value=1, max_value=10, value=5)
289
- enable_verification = st.checkbox("Enable Second-Pass Verification", value=True)
290
- enable_decomposition = st.checkbox("Decompose Complex Queries", value=True)
291
-
292
- # LLM Generator Backend
293
- st.divider()
294
- st.subheader("4. LLM Generation Engine")
295
- engine_choice = st.radio(
296
- "Generation Backend",
297
- options=["Offline Grounding Engine", "Hugging Face Inference API", "OpenAI / Groq API"],
298
- index=0,
299
- help="Offline engine runs locally with 0 cost and 100% cited answers. HF and OpenAI support conversational LLMs."
300
- )
301
 
302
- if engine_choice == "Hugging Face Inference API":
303
- hf_token_input = st.text_input(
304
- "Hugging Face Token (Optional)",
305
- type="password",
306
- value=os.getenv("HF_TOKEN", ""),
307
- help="Your HF Token for higher rate limits on the Serverless Inference API."
308
- )
309
- hf_model_choice = st.selectbox(
310
- "Hugging Face Model",
311
- options=[
312
- "Qwen/Qwen2.5-7B-Instruct",
313
- "meta-llama/Llama-3.2-3B-Instruct",
314
- "mistralai/Mistral-7B-Instruct-v0.3"
315
- ]
316
- )
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
317
  generator.llm_client = HuggingFaceInferenceClient(
318
- model=hf_model_choice,
319
- token=hf_token_input or None
320
- )
321
- elif engine_choice == "OpenAI / Groq API":
322
- api_key_input = st.text_input(
323
- "API Key",
324
- type="password",
325
- value=os.getenv("OPENAI_API_KEY", "")
326
  )
327
- base_url_input = st.text_input(
328
- "Base URL (Optional, for Groq/Ollama)",
329
- value=os.getenv("OPENAI_BASE_URL", "")
330
- )
331
- if api_key_input:
332
  generator.llm_client = OpenAILLMClient(
333
- api_key=api_key_input,
334
- base_url=base_url_input or None
335
  )
336
  else:
337
  generator.llm_client = None
@@ -339,164 +179,273 @@ with st.sidebar:
339
  # High-precision offline grounding engine
340
  generator.llm_client = None
341
 
 
342
 
343
- # ---------------- MAIN CHAT INTERFACE ----------------
344
- st.markdown('<div class="main-header">πŸ“„ Production Document Question Answering</div>', unsafe_allow_html=True)
345
- st.markdown(
346
- '<div class="sub-header">Layout-aware Ingestion β€’ Hybrid Retrieval (bge-small-en + BM25) β€’ '
347
- 'Cross-Encoder Reranking β€’ Strict Evidence Grounding β€’ Claim-Level Verification</div>',
348
- unsafe_allow_html=True
349
- )
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
350
 
351
- # Demo Prompt Chips
352
- st.markdown("**πŸ’‘ Try sample questions on the pre-loaded documents:**")
353
- chip_cols = st.columns(4)
354
- sample_prompts = [
355
- "What was the total revenue and net profit margin in Q3?",
356
- "What are the target p99 latency and platform availability SLOs?",
357
- "What hardware MFA and encryption standards are required?",
358
- "Does the system run on Oracle Cloud Infrastructure in Frankfurt?"
359
- ]
360
-
361
- selected_prompt = None
362
- with chip_cols[0]:
363
- if st.button("πŸ“Š Q3 Revenue & Margin", use_container_width=True):
364
- selected_prompt = sample_prompts[0]
365
- with chip_cols[1]:
366
- if st.button("⚑ Latency & SLA Targets", use_container_width=True):
367
- selected_prompt = sample_prompts[1]
368
- with chip_cols[2]:
369
- if st.button("πŸ”’ MFA & TLS Standards", use_container_width=True):
370
- selected_prompt = sample_prompts[2]
371
- with chip_cols[3]:
372
- if st.button("🚫 Oracle Cloud (Negative Test)", use_container_width=True):
373
- selected_prompt = sample_prompts[3]
374
-
375
- # Display Conversation History
376
- for turn in st.session_state.messages:
377
- with st.chat_message(turn["role"]):
378
- st.markdown(turn["content"])
379
-
380
- # Display Citations if available
381
- if turn.get("citations"):
382
- with st.expander(f"πŸ“š View {len(turn['citations'])} Source Citations & Evidence Quotes"):
383
- for idx, c in enumerate(turn["citations"], 1):
384
- score_info = f"(Score: `{c.get('score', 0.0):.3f}`)" if 'score' in c else ""
385
- st.markdown(
386
- f"**Citation {idx}:** `{c['file_name']}` | **Page {c['page']}**, Section *\"{c['section']}\"* {score_info}"
387
- )
388
- st.markdown(f"> *\"{c['text']}\"*")
389
- st.divider()
390
-
391
- # Display Verification details if available
392
- if turn.get("verification"):
393
- ver = turn["verification"]
394
- faithfulness = ver.get("faithfulness_score", 1.0)
395
- status_class = "verified-pill" if faithfulness >= 0.8 else "unverified-pill"
396
- status_label = f"Faithfulness: {int(faithfulness * 100)}%"
397
-
398
- with st.expander(f"πŸ›‘οΈ Second-Pass Verification Audit ({status_label})"):
399
- st.markdown(f"Status: <span class='{status_class}'>{status_label}</span>", unsafe_allow_html=True)
400
- st.markdown(f"β€’ **Supported Claims:** {len(ver.get('supported_claims', []))}")
401
- for sc in ver.get("supported_claims", []):
402
- st.caption(f"βœ“ {sc}")
403
- if ver.get("unsupported_claims"):
404
- st.markdown(f"β€’ **⚠️ Flagged / Unsupported Claims:** {len(ver.get('unsupported_claims', []))}")
405
- for uc in ver.get("unsupported_claims", []):
406
- st.caption(f"⚠️ {uc}")
407
-
408
-
409
- # User Input (via chat input or prompt chip)
410
- chat_input_val = st.chat_input("Ask a question grounded strictly in your uploaded documents...")
411
- active_prompt = selected_prompt or chat_input_val
412
-
413
- if active_prompt:
414
- st.session_state.messages.append({"role": "user", "content": active_prompt})
415
- generator.memory.add_user_message(active_prompt)
416
-
417
- with st.chat_message("user"):
418
- st.markdown(active_prompt)
419
-
420
- with st.chat_message("assistant"):
421
- with st.spinner("Retrieving evidence, reranking passages & generating grounded response..."):
422
- conv_history = generator.memory.get_messages_for_prompt()[:-1]
423
- ret_result = retriever.retrieve(
424
- query=active_prompt,
425
- initial_top_k=20,
426
- rerank_top_k=rerank_top_k,
427
- mode=retrieval_mode,
428
- conversation_history=conv_history,
429
- decompose=enable_decomposition
430
  )
431
 
432
- retrieved_chunks = ret_result["chunks"]
433
- scores = ret_result["scores"]
434
- rewritten_query = ret_result.get("rewritten_query", active_prompt)
435
- sub_queries = ret_result.get("sub_queries", [active_prompt])
 
436
 
437
- if rewritten_query != active_prompt or len(sub_queries) > 1:
438
- st.caption(f"πŸ”„ **Query Reformulation:** `{rewritten_query}` | **Sub-queries:** {sub_queries}")
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
439
 
440
- # Generate and verify answer
441
- answer, verification = generator.generate_answer(
442
- query=rewritten_query,
443
- context_chunks=retrieved_chunks,
444
- verify=enable_verification
445
- )
446
 
447
- citations_data = []
448
- for c, s in zip(retrieved_chunks, scores):
449
- citations_data.append({
450
- "file_name": c.file_name,
451
- "page": c.page_number,
452
- "section": c.section_title,
453
- "text": c.text,
454
- "score": s
455
- })
456
-
457
- # Stream output
458
- output_placeholder = st.empty()
459
- full_text = ""
460
- for word in answer.split(" "):
461
- full_text += word + " "
462
- output_placeholder.markdown(full_text + "β–Œ")
463
- time.sleep(0.012)
464
- output_placeholder.markdown(full_text.strip())
465
-
466
- # Citations expander
467
- if citations_data and "not found in the document" not in full_text.lower():
468
- with st.expander(f"πŸ“š View {len(citations_data)} Source Citations & Evidence Quotes"):
469
- for idx, c in enumerate(citations_data, 1):
470
- st.markdown(
471
- f"**Citation {idx}:** `{c['file_name']}` | **Page {c['page']}**, Section *\"{c['section']}\"* (Score: `{c['score']:.3f}`)"
472
- )
473
- st.markdown(f"> *\"{c['text']}\"*")
474
- st.divider()
475
-
476
- # Verification expander
477
- if verification and "not found in the document" not in full_text.lower():
478
- faithfulness = verification.faithfulness_score
479
- status_class = "verified-pill" if faithfulness >= 0.8 else "unverified-pill"
480
- status_label = f"Faithfulness: {int(faithfulness * 100)}%"
481
-
482
- with st.expander(f"πŸ›‘οΈ Second-Pass Verification Audit ({status_label})"):
483
- st.markdown(f"Status: <span class='{status_class}'>{status_label}</span>", unsafe_allow_html=True)
484
- st.markdown(f"β€’ **Supported Claims ({len(verification.supported_claims)}):**")
485
- for sc in verification.supported_claims:
486
- st.caption(f"βœ“ {sc}")
487
- if verification.unsupported_claims:
488
- st.markdown(f"β€’ **⚠️ Flagged / Unsupported Claims ({len(verification.unsupported_claims)}):**")
489
- for uc in verification.unsupported_claims:
490
- st.caption(f"⚠️ {uc}")
491
-
492
- generator.memory.add_assistant_message(
493
- message=full_text.strip(),
494
- citations=citations_data,
495
- verification=verification
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
496
  )
497
- st.session_state.messages.append({
498
- "role": "assistant",
499
- "content": full_text.strip(),
500
- "citations": citations_data,
501
- "verification": verification.to_dict() if verification else None
502
- })
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Production Document Question Answering (DQA) - Gradio Web Interface.
2
 
3
+ Configured for native Hugging Face Spaces deployment.
4
  Features:
5
+ - Auto-loads sample documents for instant demonstration.
6
+ - Multi-document upload: PDF (layout-aware, OCR fallback), DOCX, TXT.
7
  - Hybrid Retrieval: BAAI/bge-small-en-v1.5 + BM25Okapi + Reciprocal Rank Fusion (RRF).
8
+ - Cross-Encoder Reranking: ms-marco-MiniLM-L-6-v2.
9
+ - Multi-Backend Generator:
10
+ 1. Offline High-Precision Grounding Engine (Default / 0-cost / 100% cited)
11
+ 2. Hugging Face Serverless Inference API (Qwen-2.5, Llama-3.2, Mistral)
12
+ 3. OpenAI / Groq / DeepSeek APIs
13
  - Strict Grounding with Evidence Quotes & [Page X, Section Y] Citations.
14
+ - Second-Pass Verification Audit with Faithfulness Score.
15
+ - Real-time token streaming.
16
  """
17
 
18
  import os
19
  import sys
 
20
  import time
21
+ from typing import List, Dict, Any, Generator as TypingGenerator
22
 
23
+ import gradio as gr
24
 
25
  # Local modules
26
  from config import DQAConfig, default_config
 
30
  from verify import AnswerVerifier
31
 
32
 
33
+ # Global singleton instances
34
+ print("Initializing core components...")
35
+ pipeline = DocumentIngestionPipeline(
36
+ chunk_size=default_config.chunk_size_tokens,
37
+ chunk_overlap=default_config.chunk_overlap_tokens,
38
+ ocr_fallback=default_config.ocr_fallback_enabled,
39
+ ocr_min_chars=default_config.ocr_min_chars_per_page
40
  )
41
+ retriever = HybridRetriever(
42
+ embedding_model=default_config.embedding_model_name,
43
+ cross_encoder_model=default_config.cross_encoder_model_name,
44
+ rrf_k=default_config.rrf_k
45
+ )
46
+ verifier = AnswerVerifier(
47
+ cross_encoder=retriever.reranker,
48
+ threshold=default_config.verification_threshold,
49
+ flag_unsupported=default_config.flag_unsupported_claims
50
+ )
51
+ generator = Generator(verifier=verifier)
52
 
53
+ indexed_filenames = []
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
54
 
55
 
56
+ def load_initial_samples():
57
+ """Auto-loads sample documents on startup for instant demonstration."""
58
+ global indexed_filenames
59
+ sample_files = [
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
60
  "data/sample_financial_report.pdf",
61
  "data/sample_cloud_architecture.docx",
62
  "data/sample_security_policy.txt"
63
  ]
64
+ parsed_docs = []
65
+ indexed_filenames = []
66
+ for sf in sample_files:
67
+ if os.path.exists(sf):
68
+ p = pipeline.ingest_file(sf)
69
+ parsed_docs.append(p)
70
+ indexed_filenames.append(p.file_name)
71
+
72
+ if parsed_docs:
73
+ retriever.index_documents(parsed_docs)
74
+ print(f"Auto-loaded {len(parsed_docs)} sample documents ({len(retriever.chunks)} chunks).")
75
+
76
+
77
+ load_initial_samples()
78
+
79
+
80
+ def get_status_markdown() -> str:
81
+ """Returns markdown summarizing current indexed documents."""
82
+ total_docs = len(indexed_filenames)
83
+ total_chunks = len(retriever.chunks)
84
+ doc_list = "<br>".join([f"β€’ <code>{f}</code>" for f in indexed_filenames]) if indexed_filenames else "<i>No documents indexed.</i>"
85
+ return f"""### πŸ“Š Index Status
86
+ - **Documents Indexed:** `{total_docs}`
87
+ - **Total Chunks:** `{total_chunks}`
88
+ - **Embedding Model:** `{retriever.dense_index.model_name}`
89
+ <br>
90
+ **Active Files:**
91
+ {doc_list}
92
+ """
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
93
 
 
 
 
 
 
 
 
 
 
 
 
 
 
94
 
95
+ def process_uploaded_files(files) -> str:
96
+ """Processes newly uploaded files and adds them to the index."""
97
+ global indexed_filenames
98
+ if not files:
99
+ return get_status_markdown()
100
+
101
+ new_parsed = []
102
+ for file_obj in files:
103
+ file_path = file_obj.name if hasattr(file_obj, "name") else str(file_obj)
104
+ orig_name = os.path.basename(file_path)
105
+ try:
106
+ parsed = pipeline.ingest_file(file_path)
107
+ parsed.file_name = orig_name
108
+ for c in parsed.chunks:
109
+ c.file_name = orig_name
110
+ new_parsed.append(parsed)
111
+ if orig_name not in indexed_filenames:
112
+ indexed_filenames.append(orig_name)
113
+ except Exception as e:
114
+ print(f"Error parsing {orig_name}: {e}")
115
+
116
+ if new_parsed:
117
+ existing = list(retriever.chunks)
118
+ combined = existing + [c for p in new_parsed for c in p.chunks]
119
+ retriever.index_documents(combined)
120
+
121
+ return get_status_markdown()
122
+
123
+
124
+ def reload_samples_action() -> str:
125
+ """Reloads pre-built sample documents."""
126
+ load_initial_samples()
127
+ generator.memory.clear()
128
+ return get_status_markdown()
129
+
130
+
131
+ def clear_index_action() -> str:
132
+ """Clears all documents from the vector store."""
133
+ global indexed_filenames
134
+ retriever.index_documents([])
135
+ indexed_filenames = []
136
+ generator.memory.clear()
137
+ return get_status_markdown()
138
+
139
+
140
+ def chat_response(
141
+ message: str,
142
+ history: List[Dict[str, str]],
143
+ search_strategy: str,
144
+ top_k: int,
145
+ enable_verification: bool,
146
+ enable_decomposition: bool,
147
+ llm_backend: str,
148
+ hf_token: str,
149
+ hf_model: str,
150
+ openai_key: str,
151
+ openai_base_url: str
152
+ ):
153
+ """
154
+ Main conversational QA handler:
155
+ 1. Formulates / rewrites query.
156
+ 2. Performs hybrid search + reranking.
157
+ 3. Streams response tokens.
158
+ 4. Performs second-pass verification and formats citations.
159
+ """
160
+ if not message or not message.strip():
161
+ yield history, "No query provided.", "No query provided."
162
+ return
163
+
164
+ # Update generator LLM client according to settings
165
+ if llm_backend == "Hugging Face Inference API":
166
  generator.llm_client = HuggingFaceInferenceClient(
167
+ model=hf_model or "Qwen/Qwen2.5-7B-Instruct",
168
+ token=hf_token or None
 
 
 
 
 
 
169
  )
170
+ elif llm_backend == "OpenAI / Groq API":
171
+ if openai_key:
 
 
 
172
  generator.llm_client = OpenAILLMClient(
173
+ api_key=openai_key,
174
+ base_url=openai_base_url or None
175
  )
176
  else:
177
  generator.llm_client = None
 
179
  # High-precision offline grounding engine
180
  generator.llm_client = None
181
 
182
+ generator.enable_verification = enable_verification
183
 
184
+ # Map search mode
185
+ mode_map = {
186
+ "🌟 Hybrid (Dense+BM25) + Cross-Encoder Rerank": "hybrid_rerank",
187
+ "⚑ Hybrid Search (No Rerank)": "hybrid_no_rerank",
188
+ "🎯 Dense Vector Only (FAISS)": "dense_only",
189
+ "πŸ”€ BM25 Lexical Only": "bm25_only"
190
+ }
191
+ mode = mode_map.get(search_strategy, "hybrid_rerank")
192
+
193
+ # Format history for conversational memory
194
+ conv_history = []
195
+ for turn in history:
196
+ if isinstance(turn, dict):
197
+ conv_history.append({"role": turn.get("role", "user"), "content": turn.get("content", "")})
198
+ elif isinstance(turn, (list, tuple)) and len(turn) == 2:
199
+ conv_history.append({"role": "user", "content": str(turn[0])})
200
+ conv_history.append({"role": "assistant", "content": str(turn[1])})
201
+
202
+ # Retrieve relevant chunks
203
+ ret_res = retriever.retrieve(
204
+ query=message,
205
+ initial_top_k=20,
206
+ rerank_top_k=int(top_k),
207
+ mode=mode,
208
+ conversation_history=conv_history,
209
+ decompose=enable_decomposition
210
+ )
211
+
212
+ retrieved_chunks = ret_res["chunks"]
213
+ scores = ret_res["scores"]
214
+ rewritten_query = ret_res.get("rewritten_query", message)
215
+ sub_queries = ret_res.get("sub_queries", [message])
216
 
217
+ # Format source citations markdown
218
+ citations_md = "### πŸ“š Retrieved Evidence Passages & Citations\n\n"
219
+ if not retrieved_chunks:
220
+ citations_md += "*No relevant passages retrieved for this query.*\n"
221
+ else:
222
+ for idx, (c, s) in enumerate(zip(retrieved_chunks, scores), 1):
223
+ citations_md += (
224
+ f"**Citation {idx}:** `{c.file_name}` | **Page {c.page_number}**, Section *\"{c.section_title}\"* "
225
+ f"(Score: `{s:.3f}`, Table: `{c.metadata.get('is_table', False)}`)\n\n"
226
+ f"> \"{c.text}\"\n\n---\n"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
227
  )
228
 
229
+ # Initialize assistant turn in history
230
+ updated_history = list(history) + [
231
+ {"role": "user", "content": message},
232
+ {"role": "assistant", "content": ""}
233
+ ]
234
 
235
+ # Generate answer
236
+ raw_answer, verification = generator.generate_answer(
237
+ query=rewritten_query,
238
+ context_chunks=retrieved_chunks,
239
+ verify=enable_verification
240
+ )
241
+
242
+ # Stream output word by word for interactive response
243
+ words = raw_answer.split(" ")
244
+ streamed_text = ""
245
+ for i, w in enumerate(words):
246
+ streamed_text += w + (" " if i < len(words) - 1 else "")
247
+ updated_history[-1]["content"] = streamed_text
248
+ if i % 2 == 0 or i == len(words) - 1:
249
+ time.sleep(0.015)
250
+ # Yield progressive updates
251
+ yield updated_history, citations_md, "⏳ Verifying claims..."
252
+
253
+ # Format verification audit markdown
254
+ if verification and "not found in the document" not in raw_answer.lower():
255
+ status_badge = "βœ… High Faithfulness" if verification.faithfulness_score >= 0.8 else "⚠️ Potential Hallucination"
256
+ ver_md = f"### πŸ›‘οΈ Claim-Level Verification Audit\n\n"
257
+ ver_md += f"**Overall Status:** {status_badge} (`{int(verification.faithfulness_score * 100)}%` Faithfulness)\n\n"
258
+ ver_md += f"- **Total Claims Analyzed:** `{verification.total_claims}`\n"
259
+ ver_md += f"- **Supported Claims:** `{len(verification.supported_claims)}`\n"
260
+ for sc in verification.supported_claims:
261
+ ver_md += f" - βœ“ {sc}\n"
262
+ if verification.unsupported_claims:
263
+ ver_md += f"\n- **⚠️ Flagged / Unsupported Claims:** `{len(verification.unsupported_claims)}`\n"
264
+ for uc in verification.unsupported_claims:
265
+ ver_md += f" - ⚠️ {uc}\n"
266
+ else:
267
+ ver_md = "### πŸ›‘οΈ Claim-Level Verification Audit\n\n*Response is ungrounded negative control or refusal ('Not found in the document.'). Zero hallucination detected.*"
268
+
269
+ yield updated_history, citations_md, ver_md
270
 
 
 
 
 
 
 
271
 
272
+ # ---------------- GRADIO UI LAYOUT ----------------
273
+ custom_theme = gr.themes.Soft(
274
+ primary_hue="blue",
275
+ secondary_hue="indigo",
276
+ neutral_hue="slate"
277
+ )
278
+
279
+ with gr.Blocks(title="Document QA System") as demo:
280
+ gr.Markdown("""
281
+ # πŸ“„ Production Document Question Answering (DQA)
282
+ **Layout-aware Ingestion β€’ Hybrid Dense & BM25 Search β€’ Cross-Encoder Reranking β€’ Strict Evidence Citations β€’ Second-Pass Verification**
283
+ """)
284
+
285
+ with gr.Row():
286
+ # LEFT COLUMN: Ingestion & Configuration
287
+ with gr.Column(scale=4):
288
+ status_box = gr.Markdown(value=get_status_markdown())
289
+
290
+ with gr.Accordion("πŸ“ Upload & Manage Documents", open=True):
291
+ file_input = gr.File(
292
+ label="Upload PDF, DOCX, or TXT",
293
+ file_count="multiple",
294
+ file_types=[".pdf", ".docx", ".txt", ".md"]
295
+ )
296
+ with gr.Row():
297
+ upload_btn = gr.Button("πŸš€ Process Files", variant="primary")
298
+ reload_btn = gr.Button("πŸ“‚ Reload Samples")
299
+ clear_btn = gr.Button("πŸ”„ Clear All")
300
+
301
+ with gr.Accordion("βš™οΈ Search & Reranking Settings", open=True):
302
+ search_mode = gr.Dropdown(
303
+ label="Retrieval Strategy",
304
+ choices=[
305
+ "🌟 Hybrid (Dense+BM25) + Cross-Encoder Rerank",
306
+ "⚑ Hybrid Search (No Rerank)",
307
+ "🎯 Dense Vector Only (FAISS)",
308
+ "πŸ”€ BM25 Lexical Only"
309
+ ],
310
+ value="🌟 Hybrid (Dense+BM25) + Cross-Encoder Rerank"
311
+ )
312
+ top_k_slider = gr.Slider(
313
+ minimum=1, maximum=10, value=5, step=1,
314
+ label="Top-K Chunks for Generation"
315
+ )
316
+ verify_chk = gr.Checkbox(
317
+ value=True,
318
+ label="Enable Claim-Level Verification"
319
+ )
320
+ decomp_chk = gr.Checkbox(
321
+ value=True,
322
+ label="Decompose Complex / Comparative Questions"
323
+ )
324
+
325
+ with gr.Accordion("πŸ€– Generation Engine", open=False):
326
+ backend_select = gr.Radio(
327
+ label="LLM Backend",
328
+ choices=[
329
+ "Offline High-Precision Grounding Engine",
330
+ "Hugging Face Inference API",
331
+ "OpenAI / Groq API"
332
+ ],
333
+ value="Offline High-Precision Grounding Engine"
334
+ )
335
+ hf_token_box = gr.Textbox(
336
+ label="Hugging Face Token (Optional)",
337
+ type="password",
338
+ value=os.getenv("HF_TOKEN", "")
339
+ )
340
+ hf_model_box = gr.Dropdown(
341
+ label="Hugging Face Model",
342
+ choices=[
343
+ "Qwen/Qwen2.5-7B-Instruct",
344
+ "meta-llama/Llama-3.2-3B-Instruct",
345
+ "mistralai/Mistral-7B-Instruct-v0.3"
346
+ ],
347
+ value="Qwen/Qwen2.5-7B-Instruct"
348
+ )
349
+ openai_key_box = gr.Textbox(
350
+ label="OpenAI / Groq API Key (Optional)",
351
+ type="password",
352
+ value=os.getenv("OPENAI_API_KEY", "")
353
+ )
354
+ openai_url_box = gr.Textbox(
355
+ label="Base URL (Optional, for Groq/Ollama)",
356
+ value=os.getenv("OPENAI_BASE_URL", "")
357
+ )
358
+
359
+ # RIGHT COLUMN: Chat & Audit Inspection
360
+ with gr.Column(scale=8):
361
+ chatbot = gr.Chatbot(
362
+ label="Conversation & Grounded Answers",
363
+ height=460
364
  )
365
+
366
+ with gr.Row():
367
+ msg_input = gr.Textbox(
368
+ label="Question",
369
+ placeholder="Ask a question strictly grounded in the loaded documents...",
370
+ scale=9,
371
+ lines=1
372
+ )
373
+ submit_btn = gr.Button("Ask πŸ’¬", variant="primary", scale=2)
374
+
375
+ # Quick clickable example prompts
376
+ gr.Markdown("**πŸ’‘ Quick Examples:**")
377
+ with gr.Row():
378
+ ex1 = gr.Button("πŸ“Š Q3 Revenue & Margin", size="sm")
379
+ ex2 = gr.Button("⚑ Latency & SLA Targets", size="sm")
380
+ ex3 = gr.Button("πŸ”’ MFA & TLS Standards", size="sm")
381
+ ex4 = gr.Button("🚫 Oracle Cloud (Negative)", size="sm")
382
+
383
+ clear_chat_btn = gr.Button("🧹 Clear Conversation", size="sm")
384
+
385
+ with gr.Accordion("πŸ“š Source Citations & Evidence Quotes", open=True):
386
+ citations_display = gr.Markdown(
387
+ value="*Ask a question to inspect retrieved citations, page numbers, and similarity scores.*"
388
+ )
389
+
390
+ with gr.Accordion("πŸ›‘οΈ Second-Pass Verification Audit", open=True):
391
+ verification_display = gr.Markdown(
392
+ value="*Verification audit will show supported vs flagged claims.*"
393
+ )
394
+
395
+ # Event Handlers
396
+ upload_btn.click(
397
+ process_uploaded_files,
398
+ inputs=[file_input],
399
+ outputs=[status_box]
400
+ )
401
+ reload_btn.click(
402
+ reload_samples_action,
403
+ inputs=[],
404
+ outputs=[status_box]
405
+ )
406
+ clear_btn.click(
407
+ clear_index_action,
408
+ inputs=[],
409
+ outputs=[status_box]
410
+ )
411
+
412
+ # Chat execution
413
+ chat_inputs = [
414
+ msg_input,
415
+ chatbot,
416
+ search_mode,
417
+ top_k_slider,
418
+ verify_chk,
419
+ decomp_chk,
420
+ backend_select,
421
+ hf_token_box,
422
+ hf_model_box,
423
+ openai_key_box,
424
+ openai_url_box
425
+ ]
426
+ chat_outputs = [chatbot, citations_display, verification_display]
427
+
428
+ submit_btn.click(
429
+ chat_response,
430
+ inputs=chat_inputs,
431
+ outputs=chat_outputs
432
+ ).then(lambda: "", None, msg_input)
433
+
434
+ msg_input.submit(
435
+ chat_response,
436
+ inputs=chat_inputs,
437
+ outputs=chat_outputs
438
+ ).then(lambda: "", None, msg_input)
439
+
440
+ # Example buttons
441
+ ex1.click(lambda: "What was the total revenue and net profit margin in Q3?", None, msg_input)
442
+ ex2.click(lambda: "What are the target p99 latency and platform availability SLOs?", None, msg_input)
443
+ ex3.click(lambda: "What hardware MFA and encryption standards are required?", None, msg_input)
444
+ ex4.click(lambda: "Does the system run on Oracle Cloud Infrastructure in Frankfurt?", None, msg_input)
445
+
446
+ clear_chat_btn.click(lambda: [], None, chatbot)
447
+
448
+
449
+ if __name__ == "__main__":
450
+ demo.queue()
451
+ demo.launch(server_name="0.0.0.0", server_port=7860)
requirements.txt CHANGED
@@ -1,4 +1,5 @@
1
  # Production Document Question Answering (DQA) Dependencies
 
2
  sentence-transformers>=2.2.0
3
  faiss-cpu>=1.7.4
4
  rank-bm25>=0.2.2
 
1
  # Production Document Question Answering (DQA) Dependencies
2
+ gradio>=5.0.0
3
  sentence-transformers>=2.2.0
4
  faiss-cpu>=1.7.4
5
  rank-bm25>=0.2.2
streamlit_app.py ADDED
@@ -0,0 +1,502 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Streamlit Web Application for Production Document Question Answering (DQA).
2
+
3
+ Ready for Deployment on Hugging Face Spaces.
4
+ Features:
5
+ - Auto-loads sample documents for instant one-click demonstration.
6
+ - Upload & Ingest PDF (layout-aware, OCR fallback), DOCX, TXT.
7
+ - Hybrid Retrieval: BAAI/bge-small-en-v1.5 + BM25Okapi + Reciprocal Rank Fusion (RRF).
8
+ - Cross-Encoder Reranking (ms-marco-MiniLM-L-6-v2).
9
+ - Multi-Model Generation:
10
+ 1. Hugging Face Serverless Inference API (Qwen-2.5, Llama-3.2, Mistral)
11
+ 2. High-Precision Offline Grounding Engine (0-cost, 100% cited)
12
+ 3. OpenAI / Groq / DeepSeek API
13
+ - Strict Grounding with Evidence Quotes & [Page X, Section Y] Citations.
14
+ - Second-Pass Verification Audit with Faithfulness scoring.
15
+ - Quick-question prompt chips for instant testing.
16
+ """
17
+
18
+ import os
19
+ import sys
20
+ import tempfile
21
+ import time
22
+ from typing import List
23
+
24
+ import streamlit as st
25
+
26
+ # Local modules
27
+ from config import DQAConfig, default_config
28
+ from ingest import DocumentIngestionPipeline, DocumentChunk
29
+ from retrieve import HybridRetriever
30
+ from generate import Generator, OpenAILLMClient, HuggingFaceInferenceClient
31
+ from verify import AnswerVerifier
32
+
33
+
34
+ st.set_page_config(
35
+ page_title="Document QA System - Production RAG",
36
+ page_icon="πŸ“„",
37
+ layout="wide",
38
+ initial_sidebar_state="expanded"
39
+ )
40
+
41
+ # Custom CSS for modern styling
42
+ st.markdown("""
43
+ <style>
44
+ .main-header {
45
+ font-size: 2.2rem;
46
+ font-weight: 700;
47
+ color: #1E3A8A;
48
+ margin-bottom: 0.2rem;
49
+ }
50
+ .sub-header {
51
+ font-size: 1.05rem;
52
+ color: #4B5563;
53
+ margin-bottom: 1.2rem;
54
+ }
55
+ .citation-box {
56
+ background-color: #F8FAFC;
57
+ border-left: 4px solid #3B82F6;
58
+ padding: 0.8rem 1rem;
59
+ border-radius: 4px;
60
+ margin-top: 0.6rem;
61
+ font-size: 0.9rem;
62
+ }
63
+ .verified-pill {
64
+ display: inline-block;
65
+ padding: 3px 10px;
66
+ border-radius: 12px;
67
+ font-size: 0.82rem;
68
+ font-weight: 600;
69
+ background-color: #DEF7EC;
70
+ color: #03543F;
71
+ }
72
+ .unverified-pill {
73
+ display: inline-block;
74
+ padding: 3px 10px;
75
+ border-radius: 12px;
76
+ font-size: 0.82rem;
77
+ font-weight: 600;
78
+ background-color: #FDE8E8;
79
+ color: #9B1C1C;
80
+ }
81
+ .chip-btn {
82
+ margin: 2px;
83
+ }
84
+ </style>
85
+ """, unsafe_allow_html=True)
86
+
87
+
88
+ @st.cache_resource(show_spinner="Loading Neural Embedder & Cross-Encoder Models...")
89
+ def load_core_components(embedding_model_name: str):
90
+ """Initializes and caches the heavy neural models."""
91
+ retriever = HybridRetriever(
92
+ embedding_model=embedding_model_name,
93
+ cross_encoder_model=default_config.cross_encoder_model_name,
94
+ rrf_k=default_config.rrf_k
95
+ )
96
+ verifier = AnswerVerifier(
97
+ cross_encoder=retriever.reranker,
98
+ threshold=default_config.verification_threshold,
99
+ flag_unsupported=default_config.flag_unsupported_claims
100
+ )
101
+ pipeline = DocumentIngestionPipeline(
102
+ chunk_size=default_config.chunk_size_tokens,
103
+ chunk_overlap=default_config.chunk_overlap_tokens,
104
+ ocr_fallback=default_config.ocr_fallback_enabled,
105
+ ocr_min_chars=default_config.ocr_min_chars_per_page
106
+ )
107
+ return retriever, verifier, pipeline
108
+
109
+
110
+ # Initialize session state variables
111
+ if "embedding_model" not in st.session_state:
112
+ st.session_state.embedding_model = default_config.embedding_model_name
113
+
114
+ if "retriever" not in st.session_state:
115
+ retriever, verifier, pipeline = load_core_components(st.session_state.embedding_model)
116
+ st.session_state.retriever = retriever
117
+ st.session_state.verifier = verifier
118
+ st.session_state.pipeline = pipeline
119
+ st.session_state.generator = Generator(verifier=verifier)
120
+ st.session_state.messages = []
121
+ st.session_state.indexed_files = []
122
+ st.session_state.total_chunks = 0
123
+ st.session_state.all_chunks = []
124
+
125
+ # Auto-load pre-built sample documents on first launch for Hugging Face Spaces visitors
126
+ sample_paths = [
127
+ "data/sample_financial_report.pdf",
128
+ "data/sample_cloud_architecture.docx",
129
+ "data/sample_security_policy.txt"
130
+ ]
131
+ parsed_list = []
132
+ for p in sample_paths:
133
+ if os.path.exists(p):
134
+ parsed = pipeline.ingest_file(p)
135
+ parsed_list.append(parsed)
136
+ st.session_state.indexed_files.append(parsed.file_name)
137
+
138
+ if parsed_list:
139
+ retriever.index_documents(parsed_list)
140
+ st.session_state.total_chunks = len(retriever.chunks)
141
+ st.session_state.all_chunks = retriever.chunks
142
+
143
+ retriever = st.session_state.retriever
144
+ verifier = st.session_state.verifier
145
+ pipeline = st.session_state.pipeline
146
+ generator = st.session_state.generator
147
+
148
+
149
+ # ---------------- SIDEBAR: Controls & Configuration ----------------
150
+ with st.sidebar:
151
+ st.header("βš™οΈ Settings & Ingestion")
152
+
153
+ # Embedding Model selection
154
+ st.subheader("1. Embedding Model")
155
+ selected_embedder = st.selectbox(
156
+ "Dense Embedding Model",
157
+ options=["BAAI/bge-small-en-v1.5", "sentence-transformers/all-MiniLM-L6-v2"],
158
+ index=0 if st.session_state.embedding_model == "BAAI/bge-small-en-v1.5" else 1,
159
+ help="bge-small-en-v1.5 is the top MTEB benchmark retriever. all-MiniLM-L6-v2 is ultra fast."
160
+ )
161
+ if selected_embedder != st.session_state.embedding_model:
162
+ st.session_state.embedding_model = selected_embedder
163
+ with st.spinner("Switching embedding model..."):
164
+ retriever, verifier, pipeline = load_core_components(selected_embedder)
165
+ st.session_state.retriever = retriever
166
+ st.session_state.verifier = verifier
167
+ st.session_state.pipeline = pipeline
168
+ generator.verifier = verifier
169
+ if st.session_state.all_chunks:
170
+ retriever.index_documents(st.session_state.all_chunks)
171
+ st.rerun()
172
+
173
+ # Document Upload Section
174
+ st.divider()
175
+ st.subheader("2. Upload & Ingest Documents")
176
+ uploaded_files = st.file_uploader(
177
+ "Upload PDF (with OCR), DOCX, or TXT",
178
+ type=["pdf", "docx", "txt", "md"],
179
+ accept_multiple_files=True
180
+ )
181
+
182
+ col1, col2 = st.columns(2)
183
+ with col1:
184
+ load_sample = st.button("πŸ“‚ Reload Samples", use_container_width=True)
185
+ with col2:
186
+ reindex = st.button("πŸ”„ Clear Index", use_container_width=True)
187
+
188
+ if reindex:
189
+ retriever.index_documents([])
190
+ st.session_state.indexed_files = []
191
+ st.session_state.total_chunks = 0
192
+ st.session_state.all_chunks = []
193
+ st.session_state.messages = []
194
+ generator.memory.clear()
195
+ st.success("Vector index & conversation memory cleared!")
196
+ st.rerun()
197
+
198
+ if load_sample:
199
+ sample_paths = [
200
+ "data/sample_financial_report.pdf",
201
+ "data/sample_cloud_architecture.docx",
202
+ "data/sample_security_policy.txt"
203
+ ]
204
+ parsed_list = []
205
+ st.session_state.indexed_files = []
206
+ for p in sample_paths:
207
+ if os.path.exists(p):
208
+ parsed = pipeline.ingest_file(p)
209
+ parsed_list.append(parsed)
210
+ st.session_state.indexed_files.append(parsed.file_name)
211
+
212
+ if parsed_list:
213
+ retriever.index_documents(parsed_list)
214
+ st.session_state.total_chunks = len(retriever.chunks)
215
+ st.session_state.all_chunks = retriever.chunks
216
+ st.success(f"Indexed {len(parsed_list)} sample documents ({st.session_state.total_chunks} chunks)!")
217
+ st.rerun()
218
+
219
+ # Process user uploaded files
220
+ if uploaded_files:
221
+ if st.button("πŸš€ Process Uploaded Files", type="primary", use_container_width=True):
222
+ new_parsed = []
223
+ with st.spinner("Extracting layout, formatting tables & generating embeddings..."):
224
+ for uploaded_file in uploaded_files:
225
+ ext = os.path.splitext(uploaded_file.name)[1]
226
+ with tempfile.NamedTemporaryFile(delete=False, suffix=ext) as tmp_file:
227
+ tmp_file.write(uploaded_file.read())
228
+ tmp_path = tmp_file.name
229
+
230
+ try:
231
+ parsed = pipeline.ingest_file(tmp_path)
232
+ parsed.file_name = uploaded_file.name
233
+ for c in parsed.chunks:
234
+ c.file_name = uploaded_file.name
235
+ new_parsed.append(parsed)
236
+ if uploaded_file.name not in st.session_state.indexed_files:
237
+ st.session_state.indexed_files.append(uploaded_file.name)
238
+ finally:
239
+ if os.path.exists(tmp_path):
240
+ os.remove(tmp_path)
241
+
242
+ if new_parsed:
243
+ existing_chunks = list(retriever.chunks)
244
+ combined = existing_chunks + [c for p in new_parsed for c in p.chunks]
245
+ retriever.index_documents(combined)
246
+ st.session_state.total_chunks = len(retriever.chunks)
247
+ st.session_state.all_chunks = retriever.chunks
248
+ st.success(f"Successfully indexed {len(new_parsed)} document(s)!")
249
+ st.rerun()
250
+
251
+ # Index Status Summary
252
+ st.divider()
253
+ st.markdown("**Index Status:**")
254
+ st.write(f"β€’ **Documents Indexed:** `{len(st.session_state.indexed_files)}`")
255
+ st.write(f"β€’ **Total Chunks:** `{st.session_state.total_chunks}`")
256
+ if st.session_state.indexed_files:
257
+ with st.expander("πŸ“ View Loaded Documents"):
258
+ for fname in st.session_state.indexed_files:
259
+ st.caption(f"βœ“ {fname}")
260
+
261
+ # Chunk Inspector
262
+ if st.session_state.all_chunks:
263
+ with st.expander("πŸ” Inspect Chunks & Metadata"):
264
+ chunk_num = st.selectbox(
265
+ "Select Chunk",
266
+ range(len(st.session_state.all_chunks)),
267
+ format_func=lambda i: f"Chunk {i+1}: {st.session_state.all_chunks[i].file_name} (p. {st.session_state.all_chunks[i].page_number})"
268
+ )
269
+ selected = st.session_state.all_chunks[chunk_num]
270
+ st.markdown(f"**Doc:** `{selected.file_name}` | **Page {selected.page_number}**")
271
+ st.markdown(f"**Section:** `{selected.section_title}` | Table: `{selected.metadata.get('is_table', False)}`")
272
+ st.text_area("Chunk Content", selected.text, height=120)
273
+
274
+ # Retrieval Settings
275
+ st.divider()
276
+ st.subheader("3. Retrieval & Reranker")
277
+ retrieval_mode = st.selectbox(
278
+ "Search Pipeline",
279
+ options=["hybrid_rerank", "hybrid_no_rerank", "dense_only", "bm25_only"],
280
+ format_func=lambda m: {
281
+ "hybrid_rerank": "🌟 Hybrid (Dense+BM25) + Cross-Encoder Rerank",
282
+ "hybrid_no_rerank": "⚑ Hybrid Search (No Reranker)",
283
+ "dense_only": "🎯 Dense Vector Only (FAISS)",
284
+ "bm25_only": "πŸ”€ BM25 Lexical Only"
285
+ }[m]
286
+ )
287
+
288
+ rerank_top_k = st.slider("Top-K Context Chunks", min_value=1, max_value=10, value=5)
289
+ enable_verification = st.checkbox("Enable Second-Pass Verification", value=True)
290
+ enable_decomposition = st.checkbox("Decompose Complex Queries", value=True)
291
+
292
+ # LLM Generator Backend
293
+ st.divider()
294
+ st.subheader("4. LLM Generation Engine")
295
+ engine_choice = st.radio(
296
+ "Generation Backend",
297
+ options=["Offline Grounding Engine", "Hugging Face Inference API", "OpenAI / Groq API"],
298
+ index=0,
299
+ help="Offline engine runs locally with 0 cost and 100% cited answers. HF and OpenAI support conversational LLMs."
300
+ )
301
+
302
+ if engine_choice == "Hugging Face Inference API":
303
+ hf_token_input = st.text_input(
304
+ "Hugging Face Token (Optional)",
305
+ type="password",
306
+ value=os.getenv("HF_TOKEN", ""),
307
+ help="Your HF Token for higher rate limits on the Serverless Inference API."
308
+ )
309
+ hf_model_choice = st.selectbox(
310
+ "Hugging Face Model",
311
+ options=[
312
+ "Qwen/Qwen2.5-7B-Instruct",
313
+ "meta-llama/Llama-3.2-3B-Instruct",
314
+ "mistralai/Mistral-7B-Instruct-v0.3"
315
+ ]
316
+ )
317
+ generator.llm_client = HuggingFaceInferenceClient(
318
+ model=hf_model_choice,
319
+ token=hf_token_input or None
320
+ )
321
+ elif engine_choice == "OpenAI / Groq API":
322
+ api_key_input = st.text_input(
323
+ "API Key",
324
+ type="password",
325
+ value=os.getenv("OPENAI_API_KEY", "")
326
+ )
327
+ base_url_input = st.text_input(
328
+ "Base URL (Optional, for Groq/Ollama)",
329
+ value=os.getenv("OPENAI_BASE_URL", "")
330
+ )
331
+ if api_key_input:
332
+ generator.llm_client = OpenAILLMClient(
333
+ api_key=api_key_input,
334
+ base_url=base_url_input or None
335
+ )
336
+ else:
337
+ generator.llm_client = None
338
+ else:
339
+ # High-precision offline grounding engine
340
+ generator.llm_client = None
341
+
342
+
343
+ # ---------------- MAIN CHAT INTERFACE ----------------
344
+ st.markdown('<div class="main-header">πŸ“„ Production Document Question Answering</div>', unsafe_allow_html=True)
345
+ st.markdown(
346
+ '<div class="sub-header">Layout-aware Ingestion β€’ Hybrid Retrieval (bge-small-en + BM25) β€’ '
347
+ 'Cross-Encoder Reranking β€’ Strict Evidence Grounding β€’ Claim-Level Verification</div>',
348
+ unsafe_allow_html=True
349
+ )
350
+
351
+ # Demo Prompt Chips
352
+ st.markdown("**πŸ’‘ Try sample questions on the pre-loaded documents:**")
353
+ chip_cols = st.columns(4)
354
+ sample_prompts = [
355
+ "What was the total revenue and net profit margin in Q3?",
356
+ "What are the target p99 latency and platform availability SLOs?",
357
+ "What hardware MFA and encryption standards are required?",
358
+ "Does the system run on Oracle Cloud Infrastructure in Frankfurt?"
359
+ ]
360
+
361
+ selected_prompt = None
362
+ with chip_cols[0]:
363
+ if st.button("πŸ“Š Q3 Revenue & Margin", use_container_width=True):
364
+ selected_prompt = sample_prompts[0]
365
+ with chip_cols[1]:
366
+ if st.button("⚑ Latency & SLA Targets", use_container_width=True):
367
+ selected_prompt = sample_prompts[1]
368
+ with chip_cols[2]:
369
+ if st.button("πŸ”’ MFA & TLS Standards", use_container_width=True):
370
+ selected_prompt = sample_prompts[2]
371
+ with chip_cols[3]:
372
+ if st.button("🚫 Oracle Cloud (Negative Test)", use_container_width=True):
373
+ selected_prompt = sample_prompts[3]
374
+
375
+ # Display Conversation History
376
+ for turn in st.session_state.messages:
377
+ with st.chat_message(turn["role"]):
378
+ st.markdown(turn["content"])
379
+
380
+ # Display Citations if available
381
+ if turn.get("citations"):
382
+ with st.expander(f"πŸ“š View {len(turn['citations'])} Source Citations & Evidence Quotes"):
383
+ for idx, c in enumerate(turn["citations"], 1):
384
+ score_info = f"(Score: `{c.get('score', 0.0):.3f}`)" if 'score' in c else ""
385
+ st.markdown(
386
+ f"**Citation {idx}:** `{c['file_name']}` | **Page {c['page']}**, Section *\"{c['section']}\"* {score_info}"
387
+ )
388
+ st.markdown(f"> *\"{c['text']}\"*")
389
+ st.divider()
390
+
391
+ # Display Verification details if available
392
+ if turn.get("verification"):
393
+ ver = turn["verification"]
394
+ faithfulness = ver.get("faithfulness_score", 1.0)
395
+ status_class = "verified-pill" if faithfulness >= 0.8 else "unverified-pill"
396
+ status_label = f"Faithfulness: {int(faithfulness * 100)}%"
397
+
398
+ with st.expander(f"πŸ›‘οΈ Second-Pass Verification Audit ({status_label})"):
399
+ st.markdown(f"Status: <span class='{status_class}'>{status_label}</span>", unsafe_allow_html=True)
400
+ st.markdown(f"β€’ **Supported Claims:** {len(ver.get('supported_claims', []))}")
401
+ for sc in ver.get("supported_claims", []):
402
+ st.caption(f"βœ“ {sc}")
403
+ if ver.get("unsupported_claims"):
404
+ st.markdown(f"β€’ **⚠️ Flagged / Unsupported Claims:** {len(ver.get('unsupported_claims', []))}")
405
+ for uc in ver.get("unsupported_claims", []):
406
+ st.caption(f"⚠️ {uc}")
407
+
408
+
409
+ # User Input (via chat input or prompt chip)
410
+ chat_input_val = st.chat_input("Ask a question grounded strictly in your uploaded documents...")
411
+ active_prompt = selected_prompt or chat_input_val
412
+
413
+ if active_prompt:
414
+ st.session_state.messages.append({"role": "user", "content": active_prompt})
415
+ generator.memory.add_user_message(active_prompt)
416
+
417
+ with st.chat_message("user"):
418
+ st.markdown(active_prompt)
419
+
420
+ with st.chat_message("assistant"):
421
+ with st.spinner("Retrieving evidence, reranking passages & generating grounded response..."):
422
+ conv_history = generator.memory.get_messages_for_prompt()[:-1]
423
+ ret_result = retriever.retrieve(
424
+ query=active_prompt,
425
+ initial_top_k=20,
426
+ rerank_top_k=rerank_top_k,
427
+ mode=retrieval_mode,
428
+ conversation_history=conv_history,
429
+ decompose=enable_decomposition
430
+ )
431
+
432
+ retrieved_chunks = ret_result["chunks"]
433
+ scores = ret_result["scores"]
434
+ rewritten_query = ret_result.get("rewritten_query", active_prompt)
435
+ sub_queries = ret_result.get("sub_queries", [active_prompt])
436
+
437
+ if rewritten_query != active_prompt or len(sub_queries) > 1:
438
+ st.caption(f"πŸ”„ **Query Reformulation:** `{rewritten_query}` | **Sub-queries:** {sub_queries}")
439
+
440
+ # Generate and verify answer
441
+ answer, verification = generator.generate_answer(
442
+ query=rewritten_query,
443
+ context_chunks=retrieved_chunks,
444
+ verify=enable_verification
445
+ )
446
+
447
+ citations_data = []
448
+ for c, s in zip(retrieved_chunks, scores):
449
+ citations_data.append({
450
+ "file_name": c.file_name,
451
+ "page": c.page_number,
452
+ "section": c.section_title,
453
+ "text": c.text,
454
+ "score": s
455
+ })
456
+
457
+ # Stream output
458
+ output_placeholder = st.empty()
459
+ full_text = ""
460
+ for word in answer.split(" "):
461
+ full_text += word + " "
462
+ output_placeholder.markdown(full_text + "β–Œ")
463
+ time.sleep(0.012)
464
+ output_placeholder.markdown(full_text.strip())
465
+
466
+ # Citations expander
467
+ if citations_data and "not found in the document" not in full_text.lower():
468
+ with st.expander(f"πŸ“š View {len(citations_data)} Source Citations & Evidence Quotes"):
469
+ for idx, c in enumerate(citations_data, 1):
470
+ st.markdown(
471
+ f"**Citation {idx}:** `{c['file_name']}` | **Page {c['page']}**, Section *\"{c['section']}\"* (Score: `{c['score']:.3f}`)"
472
+ )
473
+ st.markdown(f"> *\"{c['text']}\"*")
474
+ st.divider()
475
+
476
+ # Verification expander
477
+ if verification and "not found in the document" not in full_text.lower():
478
+ faithfulness = verification.faithfulness_score
479
+ status_class = "verified-pill" if faithfulness >= 0.8 else "unverified-pill"
480
+ status_label = f"Faithfulness: {int(faithfulness * 100)}%"
481
+
482
+ with st.expander(f"πŸ›‘οΈ Second-Pass Verification Audit ({status_label})"):
483
+ st.markdown(f"Status: <span class='{status_class}'>{status_label}</span>", unsafe_allow_html=True)
484
+ st.markdown(f"β€’ **Supported Claims ({len(verification.supported_claims)}):**")
485
+ for sc in verification.supported_claims:
486
+ st.caption(f"βœ“ {sc}")
487
+ if verification.unsupported_claims:
488
+ st.markdown(f"β€’ **⚠️ Flagged / Unsupported Claims ({len(verification.unsupported_claims)}):**")
489
+ for uc in verification.unsupported_claims:
490
+ st.caption(f"⚠️ {uc}")
491
+
492
+ generator.memory.add_assistant_message(
493
+ message=full_text.strip(),
494
+ citations=citations_data,
495
+ verification=verification
496
+ )
497
+ st.session_state.messages.append({
498
+ "role": "assistant",
499
+ "content": full_text.strip(),
500
+ "citations": citations_data,
501
+ "verification": verification.to_dict() if verification else None
502
+ })