Spaces:
Sleeping
Fix: Graceful fallback when company filter finds nothing + top 30 results
Browse filesProblem 1: Company filter too strict
- Query "Sinopharm covid vaccine" found 0 results
- Filter failed silently with no fallback
- User got "No matching trials found" even though related trials exist
Fix:
- Save original keyword scores before filtering
- If company filter returns 0 results, fall back to original scores
- Add warning note to context for LLM:
"No trials found with 'Sinopharm' as sponsor. Related trials below.
Please provide general knowledge if available."
- LLM can now answer: "While I don't have Sinopharm trials in database,
here are related COVID vaccine trials..."
Problem 2: Only returned 3 trials (too few)
- Changed top_k from 3 → 30 in all search calls
- More comprehensive results for LLM to analyze
- Better for comparison queries (30 trials per treatment)
Result:
- No more silent failures on company queries
- LLM provides useful answer even when exact match not found
- 10x more trial context for better responses
🤖 Generated with Claude Code https://claude.com/claude-code
Co-Authored-By: Claude <noreply@anthropic.com>
- foundation_engine.py +27 -8
|
@@ -627,10 +627,14 @@ def retrieve_context_with_embeddings(query, top_k=10, entities=None):
|
|
| 627 |
logger.info(f"[HYBRID] Inverted index scoring: {len(keyword_scores)} trials matched ({time.time()-t_kw:.2f}s)")
|
| 628 |
|
| 629 |
# 1.5. STRICT COMPANY FILTERING (if companies specified)
|
|
|
|
| 630 |
if entities and entities.get('companies'):
|
| 631 |
companies = [c.lower() for c in entities['companies']]
|
| 632 |
logger.info(f"[STRICT FILTER] Enforcing company filter: {companies}")
|
| 633 |
|
|
|
|
|
|
|
|
|
|
| 634 |
# Filter keyword_scores to ONLY trials with these companies
|
| 635 |
filtered_keyword_scores = {}
|
| 636 |
sponsor_field_patterns = ['sponsor:', 'collaborator:', 'manufacturer:']
|
|
@@ -660,14 +664,17 @@ def retrieve_context_with_embeddings(query, top_k=10, entities=None):
|
|
| 660 |
# If no company match, EXCLUDE this trial
|
| 661 |
|
| 662 |
before_count = len(keyword_scores)
|
| 663 |
-
|
| 664 |
-
after_count = len(keyword_scores)
|
| 665 |
|
| 666 |
logger.info(f"[STRICT FILTER] Filtered {before_count} → {after_count} trials (only those from {companies})")
|
| 667 |
|
| 668 |
-
|
| 669 |
-
|
| 670 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 671 |
|
| 672 |
# 2. SEMANTIC SCORING
|
| 673 |
load_embedder()
|
|
@@ -767,6 +774,18 @@ def retrieve_context_with_embeddings(query, top_k=10, entities=None):
|
|
| 767 |
logger.info(f"[355M RANKING] Added ranking metadata to context for final LLM")
|
| 768 |
|
| 769 |
context = "\n\n---\n\n".join(context_chunks) # Use --- as separator between trials
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 770 |
logger.info(f"[HYBRID] TOTAL TIME: {time.time()-t0:.2f}s")
|
| 771 |
logger.info(f"[HYBRID] Filtered context length: {len(context)} chars (was ~{sum(len(c) for c in raw_chunks)} chars)")
|
| 772 |
|
|
@@ -1370,7 +1389,7 @@ def process_query_simple_test(conversation):
|
|
| 1370 |
|
| 1371 |
# Try to search
|
| 1372 |
start = time.time()
|
| 1373 |
-
context = retrieve_context_with_embeddings(conversation, top_k=
|
| 1374 |
search_time = time.time() - start
|
| 1375 |
|
| 1376 |
if not context:
|
|
@@ -1539,7 +1558,7 @@ def process_query(conversation):
|
|
| 1539 |
context_parts = []
|
| 1540 |
for i, treatment in enumerate(treatments[:2], 1): # Compare first 2
|
| 1541 |
logger.info(f"[COMPARE] Searching trials for {treatment}...")
|
| 1542 |
-
treatment_trials = retrieve_context_with_embeddings(treatment, top_k=
|
| 1543 |
|
| 1544 |
if treatment_trials:
|
| 1545 |
context_parts.append(f"=== TRIALS FOR {treatment.upper()} ===\n{treatment_trials}\n")
|
|
@@ -1571,7 +1590,7 @@ def process_query(conversation):
|
|
| 1571 |
logger.info("Step 1: RAG search...")
|
| 1572 |
output_parts.append("✓ Step 1: RAG search started...\n")
|
| 1573 |
# Pass entities for STRICT company filtering
|
| 1574 |
-
context = retrieve_context_with_embeddings(search_query, top_k=
|
| 1575 |
|
| 1576 |
if not context:
|
| 1577 |
return "No matching trials found in RAG search."
|
|
|
|
| 627 |
logger.info(f"[HYBRID] Inverted index scoring: {len(keyword_scores)} trials matched ({time.time()-t_kw:.2f}s)")
|
| 628 |
|
| 629 |
# 1.5. STRICT COMPANY FILTERING (if companies specified)
|
| 630 |
+
company_filter_failed = False
|
| 631 |
if entities and entities.get('companies'):
|
| 632 |
companies = [c.lower() for c in entities['companies']]
|
| 633 |
logger.info(f"[STRICT FILTER] Enforcing company filter: {companies}")
|
| 634 |
|
| 635 |
+
# Save original scores in case we need to fall back
|
| 636 |
+
original_keyword_scores = keyword_scores.copy()
|
| 637 |
+
|
| 638 |
# Filter keyword_scores to ONLY trials with these companies
|
| 639 |
filtered_keyword_scores = {}
|
| 640 |
sponsor_field_patterns = ['sponsor:', 'collaborator:', 'manufacturer:']
|
|
|
|
| 664 |
# If no company match, EXCLUDE this trial
|
| 665 |
|
| 666 |
before_count = len(keyword_scores)
|
| 667 |
+
after_count = len(filtered_keyword_scores)
|
|
|
|
| 668 |
|
| 669 |
logger.info(f"[STRICT FILTER] Filtered {before_count} → {after_count} trials (only those from {companies})")
|
| 670 |
|
| 671 |
+
# If no company matches, fall back to original search but flag it
|
| 672 |
+
if len(filtered_keyword_scores) == 0:
|
| 673 |
+
logger.warning(f"[STRICT FILTER] No trials found from companies {companies}, falling back to general search")
|
| 674 |
+
company_filter_failed = True
|
| 675 |
+
keyword_scores = original_keyword_scores # Restore original
|
| 676 |
+
else:
|
| 677 |
+
keyword_scores = filtered_keyword_scores
|
| 678 |
|
| 679 |
# 2. SEMANTIC SCORING
|
| 680 |
load_embedder()
|
|
|
|
| 774 |
logger.info(f"[355M RANKING] Added ranking metadata to context for final LLM")
|
| 775 |
|
| 776 |
context = "\n\n---\n\n".join(context_chunks) # Use --- as separator between trials
|
| 777 |
+
|
| 778 |
+
# Add note if company filter failed
|
| 779 |
+
if company_filter_failed and entities and entities.get('companies'):
|
| 780 |
+
company_names = ', '.join(entities['companies'])
|
| 781 |
+
warning_note = f"""
|
| 782 |
+
⚠️ IMPORTANT: No trials found with '{company_names}' as the sponsor/collaborator.
|
| 783 |
+
The trials below are related to the topic but may not be from {company_names} specifically.
|
| 784 |
+
If you have general knowledge about {company_names}, please provide that information along with the context below.
|
| 785 |
+
|
| 786 |
+
"""
|
| 787 |
+
context = warning_note + context
|
| 788 |
+
|
| 789 |
logger.info(f"[HYBRID] TOTAL TIME: {time.time()-t0:.2f}s")
|
| 790 |
logger.info(f"[HYBRID] Filtered context length: {len(context)} chars (was ~{sum(len(c) for c in raw_chunks)} chars)")
|
| 791 |
|
|
|
|
| 1389 |
|
| 1390 |
# Try to search
|
| 1391 |
start = time.time()
|
| 1392 |
+
context = retrieve_context_with_embeddings(conversation, top_k=30)
|
| 1393 |
search_time = time.time() - start
|
| 1394 |
|
| 1395 |
if not context:
|
|
|
|
| 1558 |
context_parts = []
|
| 1559 |
for i, treatment in enumerate(treatments[:2], 1): # Compare first 2
|
| 1560 |
logger.info(f"[COMPARE] Searching trials for {treatment}...")
|
| 1561 |
+
treatment_trials = retrieve_context_with_embeddings(treatment, top_k=30, entities=parsed_query)
|
| 1562 |
|
| 1563 |
if treatment_trials:
|
| 1564 |
context_parts.append(f"=== TRIALS FOR {treatment.upper()} ===\n{treatment_trials}\n")
|
|
|
|
| 1590 |
logger.info("Step 1: RAG search...")
|
| 1591 |
output_parts.append("✓ Step 1: RAG search started...\n")
|
| 1592 |
# Pass entities for STRICT company filtering
|
| 1593 |
+
context = retrieve_context_with_embeddings(search_query, top_k=30, entities=parsed_query)
|
| 1594 |
|
| 1595 |
if not context:
|
| 1596 |
return "No matching trials found in RAG search."
|