Manju080's picture
Initial commit
c399543
raw
history blame
12.5 kB
"""
SQL Retriever for RAG System
Intelligent retrieval of relevant SQL examples based on question similarity and table schema analysis.
"""
import re
from typing import List, Dict, Any, Optional, Tuple
from collections import defaultdict
import numpy as np
from loguru import logger
from .vector_store import VectorStore
class SQLRetriever:
"""Intelligent SQL example retriever with schema-aware filtering."""
def __init__(self, vector_store: VectorStore):
"""
Initialize the SQL retriever.
Args:
vector_store: Initialized vector store instance
"""
self.vector_store = vector_store
self.schema_cache = {} # Cache for table schema analysis
def retrieve_examples(self,
question: str,
table_headers: List[str],
top_k: int = 5,
use_schema_filtering: bool = True) -> List[Dict[str, Any]]:
"""
Retrieve relevant SQL examples using multiple retrieval strategies.
Args:
question: Natural language question
table_headers: List of table column names
top_k: Number of examples to retrieve
use_schema_filtering: Whether to use schema-aware filtering
Returns:
List of retrieved examples with relevance scores
"""
# Strategy 1: Vector similarity search
vector_results = self.vector_store.search_similar(
query=question,
table_headers=table_headers,
top_k=top_k * 2, # Get more for filtering
similarity_threshold=0.6
)
if not vector_results:
logger.warning("No vector search results found")
return []
# Strategy 2: Schema-aware filtering and ranking
if use_schema_filtering:
filtered_results = self._apply_schema_filtering(
vector_results, question, table_headers
)
else:
filtered_results = vector_results
# Strategy 3: Question type classification and boosting
enhanced_results = self._enhance_with_question_analysis(
filtered_results, question, table_headers
)
# Strategy 4: Final ranking and selection
final_results = self._final_ranking(
enhanced_results, question, table_headers, top_k
)
logger.info(f"Retrieved {len(final_results)} relevant examples")
return final_results
def _apply_schema_filtering(self,
results: List[Dict[str, Any]],
question: str,
table_headers: List[str]) -> List[Dict[str, Any]]:
"""Apply schema-aware filtering to improve relevance."""
filtered_results = []
# Analyze current table schema
current_schema = self._analyze_schema(table_headers)
for result in results:
# Analyze example table schema
example_headers = result["table_headers"]
if isinstance(example_headers, str):
example_headers = [h.strip() for h in example_headers.split(",")]
example_schema = self._analyze_schema(example_headers)
# Calculate schema similarity
schema_similarity = self._calculate_schema_similarity(
current_schema, example_schema
)
# Boost score based on schema similarity
result["schema_similarity"] = schema_similarity
result["enhanced_score"] = (
result["similarity_score"] * 0.7 +
schema_similarity * 0.3
)
# Filter out examples with very low schema similarity
if schema_similarity > 0.3:
filtered_results.append(result)
return filtered_results
def _analyze_schema(self, table_headers: List[str]) -> Dict[str, Any]:
"""Analyze table schema for intelligent matching."""
if not table_headers:
return {}
schema_info = {
"column_count": len(table_headers),
"column_types": {},
"has_numeric": False,
"has_text": False,
"has_date": False,
"has_boolean": False,
"primary_key_candidates": [],
"foreign_key_candidates": []
}
for header in table_headers:
header_lower = header.lower()
# Detect column types based on naming patterns
if any(word in header_lower for word in ['id', 'key', 'pk', 'fk']):
if 'id' in header_lower:
schema_info["primary_key_candidates"].append(header)
if 'fk' in header_lower or 'foreign' in header_lower:
schema_info["foreign_key_candidates"].append(header)
# Detect data types
if any(word in header_lower for word in ['age', 'count', 'number', 'price', 'salary', 'amount']):
schema_info["has_numeric"] = True
schema_info["column_types"][header] = "numeric"
if any(word in header_lower for word in ['name', 'title', 'description', 'text', 'comment']):
schema_info["has_text"] = True
schema_info["column_types"][header] = "text"
if any(word in header_lower for word in ['date', 'time', 'created', 'updated', 'birth']):
schema_info["has_date"] = True
schema_info["column_types"][header] = "date"
if any(word in header_lower for word in ['is_', 'has_', 'active', 'enabled', 'status']):
schema_info["has_boolean"] = True
schema_info["column_types"][header] = "boolean"
return schema_info
def _calculate_schema_similarity(self,
schema1: Dict[str, Any],
schema2: Dict[str, Any]) -> float:
"""Calculate similarity between two table schemas."""
if not schema1 or not schema2:
return 0.0
# Column count similarity
count_diff = abs(schema1.get("column_count", 0) - schema2.get("column_count", 0))
count_similarity = max(0, 1 - (count_diff / max(schema1.get("column_count", 1), 1)))
# Data type similarity
type_similarity = 0.0
if schema1.get("has_numeric") == schema2.get("has_numeric"):
type_similarity += 0.25
if schema1.get("has_text") == schema2.get("has_text"):
type_similarity += 0.25
if schema1.get("has_date") == schema2.get("has_date"):
type_similarity += 0.25
if schema1.get("has_boolean") == schema2.get("has_boolean"):
type_similarity += 0.25
# Primary key similarity
pk_similarity = 0.0
if (schema1.get("primary_key_candidates") and
schema2.get("primary_key_candidates")):
pk_similarity = 0.2
# Weighted combination
final_similarity = (
count_similarity * 0.4 +
type_similarity * 0.4 +
pk_similarity * 0.2
)
return final_similarity
def _enhance_with_question_analysis(self,
results: List[Dict[str, Any]],
question: str,
table_headers: List[str]) -> List[Dict[str, Any]]:
"""Enhance results with question type analysis."""
# Analyze question type
question_type = self._classify_question_type(question)
for result in results:
# Boost examples that match question type
if question_type in result.get("category", "").lower():
result["enhanced_score"] *= 1.2
# Boost examples with similar complexity
question_complexity = self._assess_question_complexity(question)
example_complexity = self._assess_question_complexity(result["question"])
complexity_match = 1 - abs(question_complexity - example_complexity) / max(question_complexity, 1)
result["enhanced_score"] *= (0.9 + complexity_match * 0.1)
return results
def _classify_question_type(self, question: str) -> str:
"""Classify the type of SQL question."""
question_lower = question.lower()
if any(word in question_lower for word in ['count', 'how many', 'number of']):
return "aggregation"
elif any(word in question_lower for word in ['average', 'mean', 'sum', 'total']):
return "aggregation"
elif any(word in question_lower for word in ['group by', 'grouped', 'by department', 'by category']):
return "grouping"
elif any(word in question_lower for word in ['join', 'combine', 'merge', 'connect']):
return "join"
elif any(word in question_lower for word in ['order by', 'sort', 'rank', 'top', 'highest', 'lowest']):
return "sorting"
elif any(word in question_lower for word in ['where', 'filter', 'condition']):
return "filtering"
else:
return "general"
def _assess_question_complexity(self, question: str) -> float:
"""Assess the complexity of a question (0-1 scale)."""
complexity_score = 0.0
# Length complexity
if len(question.split()) > 20:
complexity_score += 0.3
elif len(question.split()) > 10:
complexity_score += 0.2
# Keyword complexity
complex_keywords = ['join', 'group by', 'having', 'subquery', 'union', 'intersect']
for keyword in complex_keywords:
if keyword in question.lower():
complexity_score += 0.15
# Question type complexity
if '?' in question:
complexity_score += 0.1
return min(1.0, complexity_score)
def _final_ranking(self,
results: List[Dict[str, Any]],
question: str,
table_headers: List[str],
top_k: int) -> List[Dict[str, Any]]:
"""Final ranking and selection of examples."""
if not results:
return []
# Sort by enhanced score
results.sort(key=lambda x: x.get("enhanced_score", 0), reverse=True)
# Ensure diversity in results
diverse_results = []
seen_categories = set()
for result in results:
if len(diverse_results) >= top_k:
break
category = result.get("category", "general")
if category not in seen_categories or len(diverse_results) < top_k // 2:
diverse_results.append(result)
seen_categories.add(category)
# Fill remaining slots with highest scoring examples
remaining_slots = top_k - len(diverse_results)
if remaining_slots > 0:
for result in results:
if result not in diverse_results and len(diverse_results) < top_k:
diverse_results.append(result)
# Final formatting
for result in diverse_results:
result["final_score"] = result.get("enhanced_score", result.get("similarity_score", 0))
# Remove internal scoring fields
result.pop("enhanced_score", None)
result.pop("schema_similarity", None)
return diverse_results[:top_k]
def get_retrieval_stats(self) -> Dict[str, Any]:
"""Get statistics about the retrieval system."""
vector_stats = self.vector_store.get_statistics()
return {
"vector_store_stats": vector_stats,
"schema_cache_size": len(self.schema_cache),
"retrieval_strategies": [
"vector_similarity",
"schema_filtering",
"question_analysis",
"diversity_ranking"
]
}