Spaces:
Sleeping
Sleeping
File size: 16,963 Bytes
9f98759 4dcbdd1 f33d22c 4dcbdd1 9f98759 4dcbdd1 9f98759 1d5bf01 1d7f6cb 1d5bf01 1d7f6cb 1d5bf01 f33d22c 1d5bf01 f33d22c 1d5bf01 1d7f6cb 1d5bf01 9f98759 f33d22c 1d5bf01 9f98759 1d5bf01 9f98759 1d5bf01 9f98759 f33d22c 1d7f6cb f33d22c 1d7f6cb 9f98759 f33d22c 9f98759 1d7f6cb f33d22c 9f98759 f33d22c 1d7f6cb f33d22c 1d7f6cb 9f98759 f33d22c 1d7f6cb f33d22c 9f98759 f33d22c 1d7f6cb f33d22c 1d7f6cb 9f98759 1d7f6cb 1d5bf01 1d7f6cb f33d22c 1d7f6cb 4dcbdd1 9f98759 1d7f6cb 9f98759 1d7f6cb 9f98759 1d7f6cb 9f98759 1d7f6cb 9f98759 1d7f6cb 9f98759 1d7f6cb 9f98759 1d7f6cb 9f98759 1d7f6cb 9f98759 1d7f6cb 9f98759 1d7f6cb f33d22c 1d7f6cb f33d22c 1d7f6cb f33d22c 9f98759 4dcbdd1 9f98759 1d7f6cb f33d22c 9f98759 f33d22c 1d7f6cb f33d22c 1d7f6cb f33d22c 1d7f6cb f33d22c 1d7f6cb f33d22c 1d7f6cb f33d22c 1d7f6cb 9f98759 4dcbdd1 9f98759 |
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 |
import os
os.environ['ANONYMIZED_TELEMETRY'] = 'False'
import zipfile
import chromadb
from sentence_transformers import SentenceTransformer
import gradio as gr
from fastapi import FastAPI
from pydantic import BaseModel
from typing import List, Optional
import re
import time
# Extract and load database
DB_PATH = "./medqa_db"
if not os.path.exists(DB_PATH) and os.path.exists("./medqa_db.zip"):
print("π¦ Extracting database...")
with zipfile.ZipFile("./medqa_db.zip", 'r') as z:
z.extractall(".")
print("β
Database extracted")
print("π Loading ChromaDB...")
client = chromadb.PersistentClient(path=DB_PATH)
collection = client.get_collection("medqa")
print(f"β
Loaded {collection.count()} questions")
print("π§ Loading MedCPT model...")
model = SentenceTransformer('ncbi/MedCPT-Query-Encoder')
print("β
Model ready")
# ============================================================================
# Deduplication function
# ============================================================================
def deduplicate_results(results, target_count):
"""
Remove duplicate questions based on:
1. High text similarity (>0.92) - catches near-exact duplicates
2. Same answer + moderate similarity (>0.85) - catches conceptual duplicates
"""
if not results['documents'][0]:
return results
documents = results['documents'][0]
metadatas = results['metadatas'][0]
distances = results['distances'][0]
selected_indices = []
for i in range(len(documents)):
is_duplicate = False
current_answer = metadatas[i].get('answer', '')
for j in selected_indices:
selected_answer = metadatas[j].get('answer', '')
dist_diff = abs(distances[i] - distances[j])
if dist_diff < 0.08:
is_duplicate = True
break
if current_answer == selected_answer and dist_diff < 0.15:
is_duplicate = True
break
if not is_duplicate:
selected_indices.append(i)
if len(selected_indices) >= target_count:
break
return {
'documents': [[documents[i] for i in selected_indices]],
'metadatas': [[metadatas[i] for i in selected_indices]],
'distances': [[distances[i] for i in selected_indices]],
'ids': [[results['ids'][0][i] for i in selected_indices]] if 'ids' in results else None
}
# ============================================================================
# Search function with deduplication
# ============================================================================
def search(query, num_results=3, source_filter=None):
emb = model.encode(query).tolist()
where_clause = None
if source_filter and source_filter != "all":
where_clause = {"source": source_filter}
fetch_count = min(num_results * 4, 50)
results = collection.query(
query_embeddings=[emb],
n_results=fetch_count,
where=where_clause
)
return deduplicate_results(results, num_results)
# ============================================================================
# Parser to extract question structure
# ============================================================================
def parse_question_document(doc_text, metadata):
"""Extract question and choices from document text - NO TRUNCATION."""
lines = doc_text.split('\n')
question_lines = []
options_started = False
options = {}
for line in lines:
line = line.strip()
if not line:
continue
# Check if this is an option line (A., B., C., etc.)
option_match = re.match(r'^([A-E])[\.\)]\s*(.+)$', line)
if option_match:
options_started = True
letter = option_match.group(1)
text = option_match.group(2).strip()
options[letter] = text
elif not options_started:
question_lines.append(line)
# Reconstruct FULL question text - no truncation
question_text = ' '.join(question_lines).strip()
answer_idx = metadata.get('answer_idx', 'N/A')
answer_text = metadata.get('answer', 'N/A')
# If answer_text is just the letter, map it to the actual option text
if answer_text in options:
answer_text = options[answer_text]
return {
'question': question_text,
'choices': options,
'correct_answer_letter': answer_idx,
'correct_answer_text': answer_text
}
# ============================================================================
# Enhanced Gradio UI
# ============================================================================
def ui_search(query, num_results=3, source_filter="all"):
if not query.strip():
return "π‘ Enter a medical query to search"
try:
r = search(query, num_results, source_filter if source_filter != "all" else None)
if not r['documents'][0]:
return "β No results found"
out = f"π Found {len(r['documents'][0])} unique results\n\n"
for i in range(len(r['documents'][0])):
source = r['metadatas'][0][i].get('source', 'unknown')
distance = r['distances'][0][i]
similarity = 1 - distance
# Source emoji
if source == 'medgemini':
source_icon = "π¬"
source_name = "Med-Gemini"
elif source.startswith('medqa_'):
source_icon = "π"
split = source.replace('medqa_', '').upper()
source_name = f"MedQA {split}"
else:
source_icon = "π"
source_name = source.upper()
out += f"\n{'='*70}\n"
out += f"{source_icon} Result {i+1} | {source_name} | Similarity: {similarity:.3f}\n"
out += f"{'='*70}\n\n"
out += r['documents'][0][i]
answer = r['metadatas'][0][i].get('answer', 'N/A')
out += f"\n\nβ
CORRECT ANSWER: {answer}\n"
explanation = r['metadatas'][0][i].get('explanation', '')
if explanation and explanation.strip():
out += f"\nπ‘ EXPLANATION:\n{explanation}\n"
out += "\n"
return out
except Exception as e:
return f"β Error: {e}"
# Create Gradio interface
with gr.Blocks(theme=gr.themes.Soft(), title="MedQA Search") as demo:
gr.Markdown("""
# π₯Ό MedQA Semantic Search
Search across **Med-Gemini** (expert explanations) and **MedQA** (USMLE questions) databases.
Uses medical-specific embeddings (MedCPT) for accurate retrieval.
β¨ **Features**: Automatic deduplication, structured output for AI integration
""")
with gr.Row():
with gr.Column(scale=3):
query_input = gr.Textbox(
label="Medical Query",
placeholder="e.g., hyponatremia, myocardial infarction, diabetes management...",
lines=2
)
with gr.Column(scale=1):
num_results = gr.Slider(
minimum=1,
maximum=10,
value=3,
step=1,
label="Number of Results"
)
with gr.Row():
source_filter = gr.Radio(
choices=["all", "medgemini", "medqa_train", "medqa_dev", "medqa_test"],
value="all",
label="Filter by Source"
)
search_btn = gr.Button("π Search", variant="primary", size="lg")
output = gr.Textbox(
label="Search Results",
lines=25,
max_lines=50
)
search_btn.click(
fn=ui_search,
inputs=[query_input, num_results, source_filter],
outputs=output
)
query_input.submit(
fn=ui_search,
inputs=[query_input, num_results, source_filter],
outputs=output
)
gr.Markdown("""
### π Database Info
**Med-Gemini**: Expert-relabeled questions with detailed explanations
**MedQA**: USMLE-style questions (Train/Dev/Test splits)
**Total Questions**: ~10,000+ USMLE-style questions
""")
gr.Examples(
examples=[
["hyponatremia", 3, "all"],
["myocardial infarction treatment", 2, "medgemini"],
["diabetes complications", 3, "all"],
["antibiotics for pneumonia", 2, "medqa_train"]
],
inputs=[query_input, num_results, source_filter]
)
# ============================================================================
# FastAPI with structured JSON output (for OpenAI integration)
# ============================================================================
app = FastAPI()
class SearchRequest(BaseModel):
query: str
num_results: int = 3
source_filter: str = None
class BatchSearchRequest(BaseModel):
queries: List[str]
num_results_per_query: int = 10
source_filter: Optional[str] = None
@app.post("/search_medqa")
def api_search(req: SearchRequest):
"""
Search MedQA and return structured exemplars.
Returns COMPLETE question text with no truncation.
"""
r = search(req.query, req.num_results, req.source_filter)
if not r['documents'][0]:
return {"results": []}
results = []
for i in range(len(r['documents'][0])):
doc_text = r['documents'][0][i]
metadata = r['metadatas'][0][i]
# Parse the document into structured format
parsed = parse_question_document(doc_text, metadata)
# Build complete result object
result = {
"result_number": i + 1,
"question": parsed['question'], # FULL question text
"choices": parsed['choices'],
"correct_answer": parsed['correct_answer_letter'],
"correct_answer_text": parsed['correct_answer_text'],
"explanation": metadata.get('explanation', ''),
"has_explanation": bool(metadata.get('explanation', '').strip()),
"source": metadata.get('source', 'unknown'),
"exam_type": metadata.get('exam_type', 'unknown'),
"split": metadata.get('split', 'unknown'),
"similarity": round(1 - r['distances'][0][i], 3),
"metamap_phrases": metadata.get('metamap_phrases', '')
}
results.append(result)
return {"results": results}
@app.post("/batch_search_medqa")
def batch_api_search(req: BatchSearchRequest):
"""
NEW: Batch search for multiple learning objectives.
Processes all queries, tracks duplicates, and returns organized results.
Returns:
- results_by_objective: List of results organized by each objective
- unique_questions: Deduplicated list of all questions
- statistics: Coverage and quality metrics
"""
start_time = time.time()
# Track all questions and their objective mappings
all_questions = {} # key: question_text, value: question data + objectives list
results_by_objective = []
for obj_idx, query in enumerate(req.queries):
objective_id = obj_idx + 1
# Search for this objective
r = search(query, req.num_results_per_query, req.source_filter)
objective_results = []
similarities = []
if r['documents'][0]:
for i in range(len(r['documents'][0])):
doc_text = r['documents'][0][i]
metadata = r['metadatas'][0][i]
similarity = round(1 - r['distances'][0][i], 3)
similarities.append(similarity)
# Parse the document
parsed = parse_question_document(doc_text, metadata)
# Create unique key for deduplication
question_key = parsed['question'][:200] # Use first 200 chars as key
# Build result object
result = {
"question": parsed['question'],
"choices": parsed['choices'],
"correct_answer": parsed['correct_answer_letter'],
"correct_answer_text": parsed['correct_answer_text'],
"explanation": metadata.get('explanation', ''),
"has_explanation": bool(metadata.get('explanation', '').strip()),
"source": metadata.get('source', 'unknown'),
"similarity": similarity
}
# Track for global deduplication
if question_key in all_questions:
# This question already exists - add this objective to its list
all_questions[question_key]['matches_objectives'].append(objective_id)
# Update similarity if higher
if similarity > all_questions[question_key]['max_similarity']:
all_questions[question_key]['max_similarity'] = similarity
else:
# First time seeing this question
all_questions[question_key] = {
**result,
'matches_objectives': [objective_id],
'max_similarity': similarity,
'first_seen_at': objective_id
}
objective_results.append(result)
# Store results for this objective
results_by_objective.append({
"objective_id": objective_id,
"objective_text": query,
"num_results": len(objective_results),
"avg_similarity": round(sum(similarities) / len(similarities), 3) if similarities else 0,
"results": objective_results
})
# Prepare unique questions list
unique_questions = list(all_questions.values())
# Calculate statistics
execution_time = round(time.time() - start_time, 2)
total_retrieved = sum(obj['num_results'] for obj in results_by_objective)
# Coverage analysis
coverage = {
"excellent": [obj for obj in results_by_objective if obj['num_results'] >= 5],
"moderate": [obj for obj in results_by_objective if 2 <= obj['num_results'] < 5],
"limited": [obj for obj in results_by_objective if obj['num_results'] == 1],
"none": [obj for obj in results_by_objective if obj['num_results'] == 0]
}
# Multi-objective questions
multi_objective_questions = [q for q in unique_questions if len(q['matches_objectives']) > 1]
# Source distribution
sources = {}
for q in unique_questions:
source = q['source']
sources[source] = sources.get(source, 0) + 1
# Similarity distribution
all_similarities = [q['max_similarity'] for q in unique_questions]
high_sim = len([s for s in all_similarities if s > 0.8])
med_sim = len([s for s in all_similarities if 0.7 <= s <= 0.8])
low_sim = len([s for s in all_similarities if s < 0.7])
statistics = {
"total_objectives": len(req.queries),
"total_retrieved": total_retrieved,
"unique_questions": len(unique_questions),
"deduplication_rate": round((total_retrieved - len(unique_questions)) / total_retrieved * 100, 1) if total_retrieved > 0 else 0,
"execution_time_seconds": execution_time,
"coverage": {
"excellent_coverage_count": len(coverage["excellent"]),
"moderate_coverage_count": len(coverage["moderate"]),
"limited_coverage_count": len(coverage["limited"]),
"no_coverage_count": len(coverage["none"]),
"no_coverage_objectives": [obj['objective_id'] for obj in coverage["none"]]
},
"cross_objective": {
"multi_objective_questions": len(multi_objective_questions),
"multi_objective_percentage": round(len(multi_objective_questions) / len(unique_questions) * 100, 1) if unique_questions else 0
},
"sources": sources,
"similarity_distribution": {
"high_similarity_count": high_sim,
"medium_similarity_count": med_sim,
"low_similarity_count": low_sim,
"average_similarity": round(sum(all_similarities) / len(all_similarities), 3) if all_similarities else 0
}
}
return {
"results_by_objective": results_by_objective,
"unique_questions": unique_questions,
"statistics": statistics
}
app = gr.mount_gradio_app(app, demo, path="/")
if __name__ == "__main__":
import uvicorn
uvicorn.run(app, host="0.0.0.0", port=7860) |