Skip to main content
Configure StreamingDiskANN query behavior at runtime to dynamically tune the accuracy vs performance trade-off for individual queries or sessions.
  • Adjust accuracy and speed trade-offs per query
  • Fine-tune search quality without rebuilding indexes
  • Optimize for different use cases (fast approximate vs high-accuracy search)
  • Control rescoring behavior for better accuracy

​
Samples

​
Increase accuracy for important queries

-- Temporarily increase search list size and rescoring
BEGIN;
SET LOCAL diskann.query_search_list_size = 200;
SET LOCAL diskann.query_rescore = 100;

SELECT * FROM document_embedding
ORDER BY embedding <=> '[...]'
LIMIT 10;

COMMIT;

-- Reduce search list size for faster queries
SET diskann.query_search_list_size = 50;
SET diskann.query_rescore = 20;

SELECT * FROM document_embedding
ORDER BY embedding <=> '[...]'
LIMIT 10;

​
Disable rescoring for maximum speed

-- Use quantized results without rescoring
SET diskann.query_rescore = 0;

SELECT * FROM document_embedding
ORDER BY embedding <=> '[...]'
LIMIT 10;

​
Session-wide configuration

-- Set parameters for the entire session
SET diskann.query_search_list_size = 150;
SET diskann.query_rescore = 75;

-- All queries in this session use these settings
SELECT * FROM document_embedding ORDER BY embedding <=> '[...]' LIMIT 10;
SELECT * FROM document_embedding ORDER BY embedding <=> '[...]' LIMIT 20;

​
Parameters

ParameterTypeDefaultDescription
diskann.query_search_list_sizeint100Number of additional candidates considered during graph search. Higher values improve accuracy but slow queries
diskann.query_rescoreint50Number of elements rescored with full precision. Set to 0 to disable rescoring

​
diskann.query_search_list_size

Controls the size of the candidate list during graph traversal:
  • Lower values (20-50): Faster queries, reduced accuracy, fewer candidates explored
  • Default (100): Balanced performance and accuracy
  • Higher values (100-200): Better accuracy, slower queries, more thorough search
This parameter has the most direct impact on the accuracy/speed trade-off.

​
diskann.query_rescore

The number of top candidates to rescore using full-precision vectors:
  • 0: Disable rescoring, use quantized results directly (fastest)
  • Default (50): Rescore top 50 candidates for improved accuracy
  • Higher values (50-200): More thorough rescoring, better accuracy, slower queries
Rescoring helps correct errors introduced by quantization in memory-optimized indexes. For plain storage indexes, rescoring has minimal effect.

​
Setting parameters

​
Session-level (persistent)

SET diskann.query_search_list_size = 150;
Applies to all queries in the current session until changed or the session ends.

​
Transaction-local (temporary)

BEGIN;
SET LOCAL diskann.query_rescore = 100;
-- Parameter applies only within this transaction
SELECT * FROM document_embedding ORDER BY embedding <=> '[...]' LIMIT 10;
COMMIT;
-- Parameter is reset after commit
The LOCAL keyword ensures parameters reset after the transaction ends.

​
Tuning recommendations

​
High-accuracy applications

For applications where accuracy is critical (e.g., medical diagnosis, legal document search): 
SET diskann.query_search_list_size = 200;
SET diskann.query_rescore = 100;

​
Real-time applications

For latency-sensitive applications (e.g., chatbots, autocomplete): 
SET diskann.query_search_list_size = 50;
SET diskann.query_rescore = 20;

​
Batch processing

For offline batch processing where throughput matters: 
SET diskann.query_search_list_size = 75;
SET diskann.query_rescore = 30;

​
A/B testing different configurations

-- Test configuration A
BEGIN;
SET LOCAL diskann.query_search_list_size = 100;
SET LOCAL diskann.query_rescore = 50;
EXPLAIN ANALYZE SELECT * FROM document_embedding ORDER BY embedding <=> '[...]' LIMIT 10;
COMMIT;

-- Test configuration B
BEGIN;
SET LOCAL diskann.query_search_list_size = 150;
SET LOCAL diskann.query_rescore = 75;
EXPLAIN ANALYZE SELECT * FROM document_embedding ORDER BY embedding <=> '[...]' LIMIT 10;
COMMIT;

​
Performance considerations

  • Start with default values and adjust diskann.query_rescore first
  • Use transaction-local settings (SET LOCAL) for experimentation
  • Monitor query latency with EXPLAIN ANALYZE when tuning
  • Higher values consume more CPU but not significantly more memory