File size: 19,249 Bytes
47ea66a
 
 
df1c03e
47ea66a
df1c03e
 
 
 
 
 
47ea66a
c9e39ab
d3d09d6
df1c03e
d3d09d6
c9e39ab
 
d3d09d6
c9e39ab
 
 
 
 
d3d09d6
c9e39ab
 
d3d09d6
 
 
 
 
c9e39ab
d3d09d6
c9e39ab
d3d09d6
 
 
 
74f3acd
d3d09d6
 
 
 
c9e39ab
 
 
 
 
 
 
d3d09d6
c9e39ab
 
 
 
 
 
 
 
 
 
 
d3d09d6
c9e39ab
d3d09d6
 
c9e39ab
 
 
 
d3d09d6
 
 
 
c9e39ab
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d3d09d6
 
c9e39ab
 
 
 
 
 
 
b0716ab
 
c9e39ab
b0716ab
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c9e39ab
b0716ab
 
 
 
 
 
 
c9e39ab
b0716ab
c9e39ab
 
 
 
 
 
b0716ab
 
 
 
c9e39ab
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
cfc8ffd
df1c03e
cfc8ffd
 
 
 
df1c03e
 
 
 
 
 
 
 
 
cfc8ffd
 
df1c03e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c9e39ab
 
 
 
 
 
 
 
 
 
 
 
 
 
 
df1c03e
 
 
 
c9e39ab
 
 
 
 
 
 
d3d09d6
c9e39ab
 
d3d09d6
c9e39ab
 
d3d09d6
c9e39ab
 
d3d09d6
 
 
 
 
 
 
c9e39ab
d3d09d6
c9e39ab
 
 
 
 
 
 
 
 
 
 
 
 
74f3acd
 
 
 
c9e39ab
74f3acd
 
c9e39ab
cfc8ffd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c9e39ab
 
 
d3d09d6
c9e39ab
 
 
 
 
 
 
 
 
 
d3d09d6
 
c9e39ab
 
 
 
 
d3d09d6
c9e39ab
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d3d09d6
 
c9e39ab
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d3d09d6
c9e39ab
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
# IMPORTANT: spaces must be imported at the very top for Hugging Face Spaces / ZeroGPU
try:
    import spaces
    gpu_decorator = spaces.GPU
except Exception:
    def gpu_decorator(fn=None, **kwargs):
        if fn is not None:
            return fn
        def wrapper(f):
            return f
        return wrapper

"""Production Document Question Answering (DQA) - Gradio Web Interface.

Configured for native Hugging Face Spaces deployment (CPU & ZeroGPU).
Features:
- Auto-loads sample documents for instant demonstration.
- Multi-document upload: PDF (layout-aware, OCR fallback), DOCX, TXT.
- Hybrid Retrieval: BAAI/bge-small-en-v1.5 + BM25Okapi + Reciprocal Rank Fusion (RRF).
- Cross-Encoder Reranking: ms-marco-MiniLM-L-6-v2.
- Multi-Backend Generator:
    1. Offline High-Precision Grounding Engine (Default / 0-cost / 100% cited)
    2. Hugging Face Serverless Inference API (Qwen-2.5, Llama-3.2, Mistral)
    3. OpenAI / Groq / DeepSeek APIs
- Strict Grounding with Evidence Quotes & [Page X, Section Y] Citations.
- Second-Pass Verification Audit with Faithfulness Score.
- Real-time token streaming.
"""

import os
import sys
import time
from typing import List, Dict, Any, Generator as TypingGenerator

import gradio as gr

# Local modules
from config import DQAConfig, default_config
from ingest import DocumentIngestionPipeline, DocumentChunk
from retrieve import HybridRetriever, extract_text_from_content
from generate import Generator, OpenAILLMClient, HuggingFaceInferenceClient
from verify import AnswerVerifier


# Global singleton instances
print("Initializing core components...")
pipeline = DocumentIngestionPipeline(
    chunk_size=default_config.chunk_size_tokens,
    chunk_overlap=default_config.chunk_overlap_tokens,
    ocr_fallback=default_config.ocr_fallback_enabled,
    ocr_min_chars=default_config.ocr_min_chars_per_page
)
retriever = HybridRetriever(
    embedding_model=default_config.embedding_model_name,
    cross_encoder_model=default_config.cross_encoder_model_name,
    rrf_k=default_config.rrf_k
)
verifier = AnswerVerifier(
    cross_encoder=retriever.reranker,
    threshold=default_config.verification_threshold,
    flag_unsupported=default_config.flag_unsupported_claims
)
generator = Generator(verifier=verifier)

indexed_filenames = []


def load_initial_samples():
    """Auto-loads sample documents on startup for instant demonstration."""
    global indexed_filenames
    sample_files = [
        "data/sample_financial_report.pdf",
        "data/sample_cloud_architecture.docx",
        "data/sample_security_policy.txt"
    ]
    parsed_docs = []
    indexed_filenames = []
    for sf in sample_files:
        if os.path.exists(sf):
            p = pipeline.ingest_file(sf)
            parsed_docs.append(p)
            indexed_filenames.append(p.file_name)

    if parsed_docs:
        retriever.index_documents(parsed_docs)
        print(f"Auto-loaded {len(parsed_docs)} sample documents ({len(retriever.chunks)} chunks).")


load_initial_samples()


def get_status_markdown() -> str:
    """Returns markdown summarizing current indexed documents."""
    total_docs = len(indexed_filenames)
    total_chunks = len(retriever.chunks)
    doc_list = "<br>".join([f"β€’ <code>{f}</code>" for f in indexed_filenames]) if indexed_filenames else "<i>No documents indexed.</i>"
    return f"""### πŸ“Š Index Status
- **Documents Indexed:** `{total_docs}`
- **Total Chunks:** `{total_chunks}`
- **Embedding Model:** `{retriever.dense_index.model_name}`
<br>
**Active Files:**
{doc_list}
"""


def process_uploaded_files(files) -> str:
    """Processes newly uploaded files and adds them to the index."""
    global indexed_filenames
    if not files:
        return get_status_markdown()

    new_parsed = []
    error_messages = []

    for file_obj in files:
        # Robust path and filename extraction across Gradio FileData / wrappers
        if hasattr(file_obj, "path"):
            file_path = file_obj.path
            orig_name = getattr(file_obj, "orig_name", None) or os.path.basename(file_path)
        elif hasattr(file_obj, "name"):
            file_path = file_obj.name
            orig_name = getattr(file_obj, "orig_name", None) or os.path.basename(file_path)
        elif isinstance(file_obj, dict):
            file_path = file_obj.get("path") or file_obj.get("name")
            orig_name = file_obj.get("orig_name") or os.path.basename(file_path)
        else:
            file_path = str(file_obj)
            orig_name = os.path.basename(file_path)

        if not file_path or not os.path.exists(file_path):
            error_messages.append(f"File not found: {orig_name}")
            continue

        try:
            parsed = pipeline.ingest_file(file_path, original_filename=orig_name)
            if parsed.chunks:
                new_parsed.append(parsed)
                if orig_name not in indexed_filenames:
                    indexed_filenames.append(orig_name)
            else:
                error_messages.append(f"No extractable text found in: {orig_name}")
        except Exception as e:
            error_messages.append(f"Error parsing {orig_name}: {str(e)}")

    if new_parsed:
        existing = list(retriever.chunks)
        combined = existing + [c for p in new_parsed for c in p.chunks]
        retriever.index_documents(combined)

    status_md = get_status_markdown()
    if error_messages:
        status_md += "\n\n**⚠️ Upload Warnings:**\n" + "\n".join([f"- {msg}" for msg in error_messages])
    return status_md


def reload_samples_action() -> str:
    """Reloads pre-built sample documents."""
    load_initial_samples()
    generator.memory.clear()
    return get_status_markdown()


def clear_index_action() -> str:
    """Clears all documents from the vector store."""
    global indexed_filenames
    retriever.index_documents([])
    indexed_filenames = []
    generator.memory.clear()
    return get_status_markdown()


# Minimal anchor function to satisfy Hugging Face ZeroGPU startup scanner
@gpu_decorator
def _spaces_startup_anchor():
    return True


def run_neural_pipeline(
    message: str,
    conv_history: list,
    mode: str,
    top_k: int,
    enable_decomposition: bool,
    enable_verification: bool
):
    """
    Executes in-process dense/hybrid search, cross-encoder reranking, and verification
    with direct in-memory access to FAISS index and BM25 tables.
    """
    ret_res = retriever.retrieve(
        query=message,
        initial_top_k=20,
        rerank_top_k=int(top_k),
        mode=mode,
        conversation_history=conv_history,
        decompose=enable_decomposition
    )
    retrieved_chunks = ret_res["chunks"]
    rewritten_query = ret_res.get("rewritten_query", message)
    
    raw_answer, verification = generator.generate_answer(
        query=rewritten_query,
        context_chunks=retrieved_chunks,
        verify=enable_verification
    )
    return ret_res, raw_answer, verification


def chat_response(
    message: str,
    history: List[Dict[str, str]],
    search_strategy: str,
    top_k: int,
    enable_verification: bool,
    enable_decomposition: bool,
    llm_backend: str,
    hf_token: str,
    hf_model: str,
    openai_key: str,
    openai_base_url: str
):
    """
    Main conversational QA handler:
    1. Updates LLM backend configuration.
    2. Runs neural retrieval, reranking, and answer generation.
    3. Streams response tokens to the UI.
    4. Displays evidence citations and claim verification audit.
    """
    if not message or not message.strip():
        yield history, "No query provided.", "No query provided."
        return

    # Update generator LLM client according to settings
    if llm_backend == "Hugging Face Inference API":
        generator.llm_client = HuggingFaceInferenceClient(
            model=hf_model or "Qwen/Qwen2.5-7B-Instruct",
            token=hf_token or None
        )
    elif llm_backend == "OpenAI / Groq API":
        if openai_key:
            generator.llm_client = OpenAILLMClient(
                api_key=openai_key,
                base_url=openai_base_url or None
            )
        else:
            generator.llm_client = None
    else:
        # High-precision offline grounding engine
        generator.llm_client = None

    generator.enable_verification = enable_verification

    # Map search mode
    mode_map = {
        "🌟 Hybrid (Dense+BM25) + Cross-Encoder Rerank": "hybrid_rerank",
        "⚑ Hybrid Search (No Rerank)": "hybrid_no_rerank",
        "🎯 Dense Vector Only (FAISS)": "dense_only",
        "πŸ”€ BM25 Lexical Only": "bm25_only"
    }
    mode = mode_map.get(search_strategy, "hybrid_rerank")

    # Format history for conversational memory
    conv_history = []
    for turn in history:
        if isinstance(turn, dict):
            conv_history.append({
                "role": turn.get("role", "user"),
                "content": extract_text_from_content(turn.get("content", ""))
            })
        elif isinstance(turn, (list, tuple)) and len(turn) == 2:
            conv_history.append({"role": "user", "content": extract_text_from_content(turn[0])})
            conv_history.append({"role": "assistant", "content": extract_text_from_content(turn[1])})

    # Execute in-process neural retrieval + rerank + generation
    try:
        ret_res, raw_answer, verification = run_neural_pipeline(
            message=message,
            conv_history=conv_history,
            mode=mode,
            top_k=int(top_k),
            enable_decomposition=enable_decomposition,
            enable_verification=enable_verification
        )
    except Exception as e:
        import traceback
        err_msg = f"⚠️ Pipeline error: {str(e)}"
        print(f"Error in run_neural_pipeline: {traceback.format_exc()}")
        updated_history = list(history) + [
            {"role": "user", "content": message},
            {"role": "assistant", "content": err_msg}
        ]
        yield updated_history, "No citations available due to an internal error.", "Verification aborted."
        return

    retrieved_chunks = ret_res["chunks"]
    scores = ret_res["scores"]

    # Format source citations markdown
    citations_md = "### πŸ“š Retrieved Evidence Passages & Citations\n\n"
    if not retrieved_chunks:
        citations_md += "*No relevant passages retrieved for this query.*\n"
    else:
        for idx, (c, s) in enumerate(zip(retrieved_chunks, scores), 1):
            citations_md += (
                f"**Citation {idx}:** `{c.file_name}` | **Page {c.page_number}**, Section *\"{c.section_title}\"* "
                f"(Score: `{s:.3f}`, Table: `{c.metadata.get('is_table', False)}`)\n\n"
                f"> \"{c.text}\"\n\n---\n"
            )

    # Initialize assistant turn in history
    updated_history = list(history) + [
        {"role": "user", "content": message},
        {"role": "assistant", "content": ""}
    ]

    # Stream output word by word for interactive response
    words = raw_answer.split(" ")
    streamed_text = ""
    for i, w in enumerate(words):
        streamed_text += w + (" " if i < len(words) - 1 else "")
        updated_history[-1]["content"] = streamed_text
        if i % 2 == 0 or i == len(words) - 1:
            time.sleep(0.015)
            yield updated_history, citations_md, "⏳ Verifying claims..."

    # Format verification audit markdown
    if verification and "not found in the document" not in raw_answer.lower():
        status_badge = "βœ… High Faithfulness" if verification.faithfulness_score >= 0.8 else "⚠️ Potential Hallucination"
        ver_md = f"### πŸ›‘οΈ Claim-Level Verification Audit\n\n"
        ver_md += f"**Overall Status:** {status_badge} (`{int(verification.faithfulness_score * 100)}%` Faithfulness)\n\n"
        ver_md += f"- **Total Claims Analyzed:** `{verification.total_claims}`\n"
        ver_md += f"- **Supported Claims:** `{len(verification.supported_claims)}`\n"
        for sc in verification.supported_claims:
            ver_md += f"  - βœ“ {sc}\n"
        if verification.unsupported_claims:
            ver_md += f"\n- **⚠️ Flagged / Unsupported Claims:** `{len(verification.unsupported_claims)}`\n"
            for uc in verification.unsupported_claims:
                ver_md += f"  - ⚠️ {uc}\n"
    else:
        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.*"

    yield updated_history, citations_md, ver_md


# ---------------- GRADIO UI LAYOUT ----------------
with gr.Blocks(title="Document QA System") as demo:
    gr.Markdown("""
    # πŸ“„ Production Document Question Answering (DQA)
    **Layout-aware Ingestion β€’ Hybrid Dense & BM25 Search β€’ Cross-Encoder Reranking β€’ Strict Evidence Citations β€’ Second-Pass Verification**
    """)

    with gr.Row():
        # LEFT COLUMN: Ingestion & Configuration
        with gr.Column(scale=4):
            status_box = gr.Markdown(value=get_status_markdown())

            with gr.Accordion("πŸ“ Upload & Manage Documents", open=True):
                file_input = gr.File(
                    label="Upload PDF, DOCX, or TXT",
                    file_count="multiple",
                    file_types=[".pdf", ".docx", ".txt", ".md"]
                )
                with gr.Row():
                    upload_btn = gr.Button("πŸš€ Process Files", variant="primary")
                    reload_btn = gr.Button("πŸ“‚ Reload Samples")
                    clear_btn = gr.Button("πŸ”„ Clear All")

            with gr.Accordion("βš™οΈ Search & Reranking Settings", open=True):
                search_mode = gr.Dropdown(
                    label="Retrieval Strategy",
                    choices=[
                        "🌟 Hybrid (Dense+BM25) + Cross-Encoder Rerank",
                        "⚑ Hybrid Search (No Rerank)",
                        "🎯 Dense Vector Only (FAISS)",
                        "πŸ”€ BM25 Lexical Only"
                    ],
                    value="🌟 Hybrid (Dense+BM25) + Cross-Encoder Rerank"
                )
                top_k_slider = gr.Slider(
                    minimum=1, maximum=10, value=5, step=1,
                    label="Top-K Chunks for Generation"
                )
                verify_chk = gr.Checkbox(
                    value=True,
                    label="Enable Claim-Level Verification"
                )
                decomp_chk = gr.Checkbox(
                    value=True,
                    label="Decompose Complex / Comparative Questions"
                )

            with gr.Accordion("πŸ€– Generation Engine", open=False):
                backend_select = gr.Radio(
                    label="LLM Backend",
                    choices=[
                        "Offline High-Precision Grounding Engine",
                        "Hugging Face Inference API",
                        "OpenAI / Groq API"
                    ],
                    value="Offline High-Precision Grounding Engine"
                )
                hf_token_box = gr.Textbox(
                    label="Hugging Face Token (Optional)",
                    type="password",
                    value=os.getenv("HF_TOKEN", "")
                )
                hf_model_box = gr.Dropdown(
                    label="Hugging Face Model",
                    choices=[
                        "Qwen/Qwen2.5-7B-Instruct",
                        "meta-llama/Llama-3.2-3B-Instruct",
                        "mistralai/Mistral-7B-Instruct-v0.3"
                    ],
                    value="Qwen/Qwen2.5-7B-Instruct"
                )
                openai_key_box = gr.Textbox(
                    label="OpenAI / Groq API Key (Optional)",
                    type="password",
                    value=os.getenv("OPENAI_API_KEY", "")
                )
                openai_url_box = gr.Textbox(
                    label="Base URL (Optional, for Groq/Ollama)",
                    value=os.getenv("OPENAI_BASE_URL", "")
                )

        # RIGHT COLUMN: Chat & Audit Inspection
        with gr.Column(scale=8):
            chatbot = gr.Chatbot(
                label="Conversation & Grounded Answers",
                height=460
            )

            with gr.Row():
                msg_input = gr.Textbox(
                    label="Question",
                    placeholder="Ask a question strictly grounded in the loaded documents...",
                    scale=9,
                    lines=1
                )
                submit_btn = gr.Button("Ask πŸ’¬", variant="primary", scale=2)

            # Quick clickable example prompts
            gr.Markdown("**πŸ’‘ Quick Examples:**")
            with gr.Row():
                ex1 = gr.Button("πŸ“Š Q3 Revenue & Margin", size="sm")
                ex2 = gr.Button("⚑ Latency & SLA Targets", size="sm")
                ex3 = gr.Button("πŸ”’ MFA & TLS Standards", size="sm")
                ex4 = gr.Button("🚫 Oracle Cloud (Negative)", size="sm")

            clear_chat_btn = gr.Button("🧹 Clear Conversation", size="sm")

            with gr.Accordion("πŸ“š Source Citations & Evidence Quotes", open=True):
                citations_display = gr.Markdown(
                    value="*Ask a question to inspect retrieved citations, page numbers, and similarity scores.*"
                )

            with gr.Accordion("πŸ›‘οΈ Second-Pass Verification Audit", open=True):
                verification_display = gr.Markdown(
                    value="*Verification audit will show supported vs flagged claims.*"
                )

    # Event Handlers
    upload_btn.click(
        process_uploaded_files,
        inputs=[file_input],
        outputs=[status_box]
    )
    reload_btn.click(
        reload_samples_action,
        inputs=[],
        outputs=[status_box]
    )
    clear_btn.click(
        clear_index_action,
        inputs=[],
        outputs=[status_box]
    )

    # Chat execution
    chat_inputs = [
        msg_input,
        chatbot,
        search_mode,
        top_k_slider,
        verify_chk,
        decomp_chk,
        backend_select,
        hf_token_box,
        hf_model_box,
        openai_key_box,
        openai_url_box
    ]
    chat_outputs = [chatbot, citations_display, verification_display]

    submit_btn.click(
        chat_response,
        inputs=chat_inputs,
        outputs=chat_outputs
    ).then(lambda: "", None, msg_input)

    msg_input.submit(
        chat_response,
        inputs=chat_inputs,
        outputs=chat_outputs
    ).then(lambda: "", None, msg_input)

    # Example buttons
    ex1.click(lambda: "What was the total revenue and net profit margin in Q3?", None, msg_input)
    ex2.click(lambda: "What are the target p99 latency and platform availability SLOs?", None, msg_input)
    ex3.click(lambda: "What hardware MFA and encryption standards are required?", None, msg_input)
    ex4.click(lambda: "Does the system run on Oracle Cloud Infrastructure in Frankfurt?", None, msg_input)

    clear_chat_btn.click(lambda: [], None, chatbot)


if __name__ == "__main__":
    demo.queue()
    demo.launch(server_name="0.0.0.0", server_port=7860)