feat: Add Clickzetta Lakehouse vector database integration (#22551)

Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
yunqiqiliang
2025-08-07 14:21:46 +08:00
committed by GitHub
parent 2931c891a7
commit e01510e2a6
25 changed files with 4788 additions and 9 deletions

View File

@@ -0,0 +1,190 @@
# Clickzetta Vector Database Integration
This module provides integration with Clickzetta Lakehouse as a vector database for Dify.
## Features
- **Vector Storage**: Store and retrieve high-dimensional vectors using Clickzetta's native VECTOR type
- **Vector Search**: Efficient similarity search using HNSW algorithm
- **Full-Text Search**: Leverage Clickzetta's inverted index for powerful text search capabilities
- **Hybrid Search**: Combine vector similarity and full-text search for better results
- **Multi-language Support**: Built-in support for Chinese, English, and Unicode text processing
- **Scalable**: Leverage Clickzetta's distributed architecture for large-scale deployments
## Configuration
### Required Environment Variables
All seven configuration parameters are required:
```bash
# Authentication
CLICKZETTA_USERNAME=your_username
CLICKZETTA_PASSWORD=your_password
# Instance configuration
CLICKZETTA_INSTANCE=your_instance_id
CLICKZETTA_SERVICE=api.clickzetta.com
CLICKZETTA_WORKSPACE=your_workspace
CLICKZETTA_VCLUSTER=your_vcluster
CLICKZETTA_SCHEMA=your_schema
```
### Optional Configuration
```bash
# Batch processing
CLICKZETTA_BATCH_SIZE=100
# Full-text search configuration
CLICKZETTA_ENABLE_INVERTED_INDEX=true
CLICKZETTA_ANALYZER_TYPE=chinese # Options: keyword, english, chinese, unicode
CLICKZETTA_ANALYZER_MODE=smart # Options: max_word, smart
# Vector search configuration
CLICKZETTA_VECTOR_DISTANCE_FUNCTION=cosine_distance # Options: l2_distance, cosine_distance
```
## Usage
### 1. Set Clickzetta as the Vector Store
In your Dify configuration, set:
```bash
VECTOR_STORE=clickzetta
```
### 2. Table Structure
Clickzetta will automatically create tables with the following structure:
```sql
CREATE TABLE <collection_name> (
id STRING NOT NULL,
content STRING NOT NULL,
metadata JSON,
vector VECTOR(FLOAT, <dimension>) NOT NULL,
PRIMARY KEY (id)
);
-- Vector index for similarity search
CREATE VECTOR INDEX idx_<collection_name>_vec
ON TABLE <schema>.<collection_name>(vector)
PROPERTIES (
"distance.function" = "cosine_distance",
"scalar.type" = "f32"
);
-- Inverted index for full-text search (if enabled)
CREATE INVERTED INDEX idx_<collection_name>_text
ON <schema>.<collection_name>(content)
PROPERTIES (
"analyzer" = "chinese",
"mode" = "smart"
);
```
## Full-Text Search Capabilities
Clickzetta supports advanced full-text search with multiple analyzers:
### Analyzer Types
1. **keyword**: No tokenization, treats the entire string as a single token
- Best for: Exact matching, IDs, codes
2. **english**: Designed for English text
- Features: Recognizes ASCII letters and numbers, converts to lowercase
- Best for: English content
3. **chinese**: Chinese text tokenizer
- Features: Recognizes Chinese and English characters, removes punctuation
- Best for: Chinese or mixed Chinese-English content
4. **unicode**: Multi-language tokenizer based on Unicode
- Features: Recognizes text boundaries in multiple languages
- Best for: Multi-language content
### Analyzer Modes
- **max_word**: Fine-grained tokenization (more tokens)
- **smart**: Intelligent tokenization (balanced)
### Full-Text Search Functions
- `MATCH_ALL(column, query)`: All terms must be present
- `MATCH_ANY(column, query)`: At least one term must be present
- `MATCH_PHRASE(column, query)`: Exact phrase matching
- `MATCH_PHRASE_PREFIX(column, query)`: Phrase prefix matching
- `MATCH_REGEXP(column, pattern)`: Regular expression matching
## Performance Optimization
### Vector Search
1. **Adjust exploration factor** for accuracy vs speed trade-off:
```sql
SET cz.vector.index.search.ef=64;
```
2. **Use appropriate distance functions**:
- `cosine_distance`: Best for normalized embeddings (e.g., from language models)
- `l2_distance`: Best for raw feature vectors
### Full-Text Search
1. **Choose the right analyzer**:
- Use `keyword` for exact matching
- Use language-specific analyzers for better tokenization
2. **Combine with vector search**:
- Pre-filter with full-text search for better performance
- Use hybrid search for improved relevance
## Troubleshooting
### Connection Issues
1. Verify all 7 required configuration parameters are set
2. Check network connectivity to Clickzetta service
3. Ensure the user has proper permissions on the schema
### Search Performance
1. Verify vector index exists:
```sql
SHOW INDEX FROM <schema>.<table_name>;
```
2. Check if vector index is being used:
```sql
EXPLAIN SELECT ... WHERE l2_distance(...) < threshold;
```
Look for `vector_index_search_type` in the execution plan.
### Full-Text Search Not Working
1. Verify inverted index is created
2. Check analyzer configuration matches your content language
3. Use `TOKENIZE()` function to test tokenization:
```sql
SELECT TOKENIZE('your text', map('analyzer', 'chinese', 'mode', 'smart'));
```
## Limitations
1. Vector operations don't support `ORDER BY` or `GROUP BY` directly on vector columns
2. Full-text search relevance scores are not provided by Clickzetta
3. Inverted index creation may fail for very large existing tables (continue without error)
4. Index naming constraints:
- Index names must be unique within a schema
- Only one vector index can be created per column
- The implementation uses timestamps to ensure unique index names
5. A column can only have one vector index at a time
## References
- [Clickzetta Vector Search Documentation](../../../../../../../yunqidoc/cn_markdown_20250526/vector-search.md)
- [Clickzetta Inverted Index Documentation](../../../../../../../yunqidoc/cn_markdown_20250526/inverted-index.md)
- [Clickzetta SQL Functions](../../../../../../../yunqidoc/cn_markdown_20250526/sql_functions/)

View File

@@ -0,0 +1 @@
# Clickzetta Vector Database Integration for Dify

View File

@@ -0,0 +1,834 @@
import json
import logging
import queue
import threading
import uuid
from typing import Any, Optional, TYPE_CHECKING
import clickzetta # type: ignore
from pydantic import BaseModel, model_validator
if TYPE_CHECKING:
from clickzetta import Connection
from configs import dify_config
from core.rag.datasource.vdb.field import Field
from core.rag.datasource.vdb.vector_base import BaseVector
from core.rag.datasource.vdb.vector_factory import AbstractVectorFactory
from core.rag.embedding.embedding_base import Embeddings
from core.rag.models.document import Document
from models.dataset import Dataset
logger = logging.getLogger(__name__)
# ClickZetta Lakehouse Vector Database Configuration
class ClickzettaConfig(BaseModel):
"""
Configuration class for Clickzetta connection.
"""
username: str
password: str
instance: str
service: str = "api.clickzetta.com"
workspace: str = "quick_start"
vcluster: str = "default_ap"
schema_name: str = "dify" # Renamed to avoid shadowing BaseModel.schema
# Advanced settings
batch_size: int = 20 # Reduced batch size to avoid large SQL statements
enable_inverted_index: bool = True # Enable inverted index for full-text search
analyzer_type: str = "chinese" # Analyzer type for full-text search: keyword, english, chinese, unicode
analyzer_mode: str = "smart" # Analyzer mode: max_word, smart
vector_distance_function: str = "cosine_distance" # l2_distance or cosine_distance
@model_validator(mode="before")
@classmethod
def validate_config(cls, values: dict) -> dict:
"""
Validate the configuration values.
"""
if not values.get("username"):
raise ValueError("config CLICKZETTA_USERNAME is required")
if not values.get("password"):
raise ValueError("config CLICKZETTA_PASSWORD is required")
if not values.get("instance"):
raise ValueError("config CLICKZETTA_INSTANCE is required")
if not values.get("service"):
raise ValueError("config CLICKZETTA_SERVICE is required")
if not values.get("workspace"):
raise ValueError("config CLICKZETTA_WORKSPACE is required")
if not values.get("vcluster"):
raise ValueError("config CLICKZETTA_VCLUSTER is required")
if not values.get("schema_name"):
raise ValueError("config CLICKZETTA_SCHEMA is required")
return values
class ClickzettaVector(BaseVector):
"""
Clickzetta vector storage implementation.
"""
# Class-level write queue and lock for serializing writes
_write_queue: Optional[queue.Queue] = None
_write_thread: Optional[threading.Thread] = None
_write_lock = threading.Lock()
_shutdown = False
def __init__(self, collection_name: str, config: ClickzettaConfig):
super().__init__(collection_name)
self._config = config
self._table_name = collection_name.replace("-", "_").lower() # Ensure valid table name
self._connection: Optional["Connection"] = None
self._init_connection()
self._init_write_queue()
def _init_connection(self):
"""Initialize Clickzetta connection."""
self._connection = clickzetta.connect(
username=self._config.username,
password=self._config.password,
instance=self._config.instance,
service=self._config.service,
workspace=self._config.workspace,
vcluster=self._config.vcluster,
schema=self._config.schema_name
)
# Set session parameters for better string handling and performance optimization
if self._connection is not None:
with self._connection.cursor() as cursor:
# Use quote mode for string literal escaping to handle quotes better
cursor.execute("SET cz.sql.string.literal.escape.mode = 'quote'")
logger.info("Set string literal escape mode to 'quote' for better quote handling")
# Performance optimization hints for vector operations
self._set_performance_hints(cursor)
def _set_performance_hints(self, cursor):
"""Set ClickZetta performance optimization hints for vector operations."""
try:
# Performance optimization hints for vector operations and query processing
performance_hints = [
# Vector index optimization
"SET cz.storage.parquet.vector.index.read.memory.cache = true",
"SET cz.storage.parquet.vector.index.read.local.cache = false",
# Query optimization
"SET cz.sql.table.scan.push.down.filter = true",
"SET cz.sql.table.scan.enable.ensure.filter = true",
"SET cz.storage.always.prefetch.internal = true",
"SET cz.optimizer.generate.columns.always.valid = true",
"SET cz.sql.index.prewhere.enabled = true",
# Storage optimization
"SET cz.storage.parquet.enable.io.prefetch = false",
"SET cz.optimizer.enable.mv.rewrite = false",
"SET cz.sql.dump.as.lz4 = true",
"SET cz.optimizer.limited.optimization.naive.query = true",
"SET cz.sql.table.scan.enable.push.down.log = false",
"SET cz.storage.use.file.format.local.stats = false",
"SET cz.storage.local.file.object.cache.level = all",
# Job execution optimization
"SET cz.sql.job.fast.mode = true",
"SET cz.storage.parquet.non.contiguous.read = true",
"SET cz.sql.compaction.after.commit = true"
]
for hint in performance_hints:
cursor.execute(hint)
logger.info("Applied %d performance optimization hints for ClickZetta vector operations", len(performance_hints))
except Exception:
# Catch any errors setting performance hints but continue with defaults
logger.exception("Failed to set some performance hints, continuing with default settings")
@classmethod
def _init_write_queue(cls):
"""Initialize the write queue and worker thread."""
with cls._write_lock:
if cls._write_queue is None:
cls._write_queue = queue.Queue()
cls._write_thread = threading.Thread(target=cls._write_worker, daemon=True)
cls._write_thread.start()
logger.info("Started Clickzetta write worker thread")
@classmethod
def _write_worker(cls):
"""Worker thread that processes write tasks sequentially."""
while not cls._shutdown:
try:
# Get task from queue with timeout
if cls._write_queue is not None:
task = cls._write_queue.get(timeout=1)
if task is None: # Shutdown signal
break
# Execute the write task
func, args, kwargs, result_queue = task
try:
result = func(*args, **kwargs)
result_queue.put((True, result))
except (RuntimeError, ValueError, TypeError, ConnectionError) as e:
logger.exception("Write task failed")
result_queue.put((False, e))
finally:
cls._write_queue.task_done()
else:
break
except queue.Empty:
continue
except (RuntimeError, ValueError, TypeError, ConnectionError) as e:
logger.exception("Write worker error")
def _execute_write(self, func, *args, **kwargs):
"""Execute a write operation through the queue."""
if ClickzettaVector._write_queue is None:
raise RuntimeError("Write queue not initialized")
result_queue: queue.Queue[tuple[bool, Any]] = queue.Queue()
ClickzettaVector._write_queue.put((func, args, kwargs, result_queue))
# Wait for result
success, result = result_queue.get()
if not success:
raise result
return result
def get_type(self) -> str:
"""Return the vector database type."""
return "clickzetta"
def _ensure_connection(self) -> "Connection":
"""Ensure connection is available and return it."""
if self._connection is None:
raise RuntimeError("Database connection not initialized")
return self._connection
def _table_exists(self) -> bool:
"""Check if the table exists."""
try:
connection = self._ensure_connection()
with connection.cursor() as cursor:
cursor.execute(f"DESC {self._config.schema_name}.{self._table_name}")
return True
except (RuntimeError, ValueError) as e:
if "table or view not found" in str(e).lower():
return False
else:
# Re-raise if it's a different error
raise
def create(self, texts: list[Document], embeddings: list[list[float]], **kwargs):
"""Create the collection and add initial documents."""
# Execute table creation through write queue to avoid concurrent conflicts
self._execute_write(self._create_table_and_indexes, embeddings)
# Add initial texts
if texts:
self.add_texts(texts, embeddings, **kwargs)
def _create_table_and_indexes(self, embeddings: list[list[float]]):
"""Create table and indexes (executed in write worker thread)."""
# Check if table already exists to avoid unnecessary index creation
if self._table_exists():
logger.info("Table %s.%s already exists, skipping creation", self._config.schema_name, self._table_name)
return
# Create table with vector and metadata columns
dimension = len(embeddings[0]) if embeddings else 768
create_table_sql = f"""
CREATE TABLE IF NOT EXISTS {self._config.schema_name}.{self._table_name} (
id STRING NOT NULL COMMENT 'Unique document identifier',
{Field.CONTENT_KEY.value} STRING NOT NULL COMMENT 'Document text content for search and retrieval',
{Field.METADATA_KEY.value} JSON COMMENT 'Document metadata including source, type, and other attributes',
{Field.VECTOR.value} VECTOR(FLOAT, {dimension}) NOT NULL COMMENT
'High-dimensional embedding vector for semantic similarity search',
PRIMARY KEY (id)
) COMMENT 'Dify RAG knowledge base vector storage table for document embeddings and content'
"""
connection = self._ensure_connection()
with connection.cursor() as cursor:
cursor.execute(create_table_sql)
logger.info("Created table %s.%s", self._config.schema_name, self._table_name)
# Create vector index
self._create_vector_index(cursor)
# Create inverted index for full-text search if enabled
if self._config.enable_inverted_index:
self._create_inverted_index(cursor)
def _create_vector_index(self, cursor):
"""Create HNSW vector index for similarity search."""
# Use a fixed index name based on table and column name
index_name = f"idx_{self._table_name}_vector"
# First check if an index already exists on this column
try:
cursor.execute(f"SHOW INDEX FROM {self._config.schema_name}.{self._table_name}")
existing_indexes = cursor.fetchall()
for idx in existing_indexes:
# Check if vector index already exists on the embedding column
if Field.VECTOR.value in str(idx).lower():
logger.info("Vector index already exists on column %s", Field.VECTOR.value)
return
except (RuntimeError, ValueError) as e:
logger.warning("Failed to check existing indexes: %s", e)
index_sql = f"""
CREATE VECTOR INDEX IF NOT EXISTS {index_name}
ON TABLE {self._config.schema_name}.{self._table_name}({Field.VECTOR.value})
PROPERTIES (
"distance.function" = "{self._config.vector_distance_function}",
"scalar.type" = "f32",
"m" = "16",
"ef.construction" = "128"
)
"""
try:
cursor.execute(index_sql)
logger.info("Created vector index: %s", index_name)
except (RuntimeError, ValueError) as e:
error_msg = str(e).lower()
if ("already exists" in error_msg or
"already has index" in error_msg or
"with the same type" in error_msg):
logger.info("Vector index already exists: %s", e)
else:
logger.exception("Failed to create vector index")
raise
def _create_inverted_index(self, cursor):
"""Create inverted index for full-text search."""
# Use a fixed index name based on table name to avoid duplicates
index_name = f"idx_{self._table_name}_text"
# Check if an inverted index already exists on this column
try:
cursor.execute(f"SHOW INDEX FROM {self._config.schema_name}.{self._table_name}")
existing_indexes = cursor.fetchall()
for idx in existing_indexes:
idx_str = str(idx).lower()
# More precise check: look for inverted index specifically on the content column
if ("inverted" in idx_str and
Field.CONTENT_KEY.value.lower() in idx_str and
(index_name.lower() in idx_str or f"idx_{self._table_name}_text" in idx_str)):
logger.info("Inverted index already exists on column %s: %s", Field.CONTENT_KEY.value, idx)
return
except (RuntimeError, ValueError) as e:
logger.warning("Failed to check existing indexes: %s", e)
index_sql = f"""
CREATE INVERTED INDEX IF NOT EXISTS {index_name}
ON TABLE {self._config.schema_name}.{self._table_name} ({Field.CONTENT_KEY.value})
PROPERTIES (
"analyzer" = "{self._config.analyzer_type}",
"mode" = "{self._config.analyzer_mode}"
)
"""
try:
cursor.execute(index_sql)
logger.info("Created inverted index: %s", index_name)
except (RuntimeError, ValueError) as e:
error_msg = str(e).lower()
# Handle ClickZetta specific error messages
if (("already exists" in error_msg or
"already has index" in error_msg or
"with the same type" in error_msg or
"cannot create inverted index" in error_msg) and
"already has index" in error_msg):
logger.info("Inverted index already exists on column %s", Field.CONTENT_KEY.value)
# Try to get the existing index name for logging
try:
cursor.execute(f"SHOW INDEX FROM {self._config.schema_name}.{self._table_name}")
existing_indexes = cursor.fetchall()
for idx in existing_indexes:
if "inverted" in str(idx).lower() and Field.CONTENT_KEY.value.lower() in str(idx).lower():
logger.info("Found existing inverted index: %s", idx)
break
except (RuntimeError, ValueError):
pass
else:
logger.warning("Failed to create inverted index: %s", e)
# Continue without inverted index - full-text search will fall back to LIKE
def add_texts(self, documents: list[Document], embeddings: list[list[float]], **kwargs):
"""Add documents with embeddings to the collection."""
if not documents:
return
batch_size = self._config.batch_size
total_batches = (len(documents) + batch_size - 1) // batch_size
for i in range(0, len(documents), batch_size):
batch_docs = documents[i:i + batch_size]
batch_embeddings = embeddings[i:i + batch_size]
# Execute batch insert through write queue
self._execute_write(self._insert_batch, batch_docs, batch_embeddings, i, batch_size, total_batches)
def _insert_batch(self, batch_docs: list[Document], batch_embeddings: list[list[float]],
batch_index: int, batch_size: int, total_batches: int):
"""Insert a batch of documents using parameterized queries (executed in write worker thread)."""
if not batch_docs or not batch_embeddings:
logger.warning("Empty batch provided, skipping insertion")
return
if len(batch_docs) != len(batch_embeddings):
logger.error("Mismatch between docs (%d) and embeddings (%d)", len(batch_docs), len(batch_embeddings))
return
# Prepare data for parameterized insertion
data_rows = []
vector_dimension = len(batch_embeddings[0]) if batch_embeddings and batch_embeddings[0] else 768
for doc, embedding in zip(batch_docs, batch_embeddings):
# Optimized: minimal checks for common case, fallback for edge cases
metadata = doc.metadata if doc.metadata else {}
if not isinstance(metadata, dict):
metadata = {}
doc_id = self._safe_doc_id(metadata.get("doc_id", str(uuid.uuid4())))
# Fast path for JSON serialization
try:
metadata_json = json.dumps(metadata, ensure_ascii=True)
except (TypeError, ValueError):
logger.warning("JSON serialization failed, using empty dict")
metadata_json = "{}"
content = doc.page_content or ""
# According to ClickZetta docs, vector should be formatted as array string
# for external systems: '[1.0, 2.0, 3.0]'
vector_str = '[' + ','.join(map(str, embedding)) + ']'
data_rows.append([doc_id, content, metadata_json, vector_str])
# Check if we have any valid data to insert
if not data_rows:
logger.warning("No valid documents to insert in batch %d/%d", batch_index // batch_size + 1, total_batches)
return
# Use parameterized INSERT with executemany for better performance and security
# Cast JSON and VECTOR in SQL, pass raw data as parameters
columns = f"id, {Field.CONTENT_KEY.value}, {Field.METADATA_KEY.value}, {Field.VECTOR.value}"
insert_sql = (
f"INSERT INTO {self._config.schema_name}.{self._table_name} ({columns}) "
f"VALUES (?, ?, CAST(? AS JSON), CAST(? AS VECTOR({vector_dimension})))"
)
connection = self._ensure_connection()
with connection.cursor() as cursor:
try:
# Set session-level hints for batch insert operations
# Note: executemany doesn't support hints parameter, so we set them as session variables
cursor.execute("SET cz.sql.job.fast.mode = true")
cursor.execute("SET cz.sql.compaction.after.commit = true")
cursor.execute("SET cz.storage.always.prefetch.internal = true")
cursor.executemany(insert_sql, data_rows)
logger.info(
f"Inserted batch {batch_index // batch_size + 1}/{total_batches} "
f"({len(data_rows)} valid docs using parameterized query with VECTOR({vector_dimension}) cast)"
)
except (RuntimeError, ValueError, TypeError, ConnectionError) as e:
logger.exception("Parameterized SQL execution failed for %d documents: %s", len(data_rows), e)
logger.exception("SQL template: %s", insert_sql)
logger.exception("Sample data row: %s", data_rows[0] if data_rows else 'None')
raise
def text_exists(self, id: str) -> bool:
"""Check if a document exists by ID."""
safe_id = self._safe_doc_id(id)
connection = self._ensure_connection()
with connection.cursor() as cursor:
cursor.execute(
f"SELECT COUNT(*) FROM {self._config.schema_name}.{self._table_name} WHERE id = ?",
[safe_id]
)
result = cursor.fetchone()
return result[0] > 0 if result else False
def delete_by_ids(self, ids: list[str]) -> None:
"""Delete documents by IDs."""
if not ids:
return
# Check if table exists before attempting delete
if not self._table_exists():
logger.warning("Table %s.%s does not exist, skipping delete", self._config.schema_name, self._table_name)
return
# Execute delete through write queue
self._execute_write(self._delete_by_ids_impl, ids)
def _delete_by_ids_impl(self, ids: list[str]) -> None:
"""Implementation of delete by IDs (executed in write worker thread)."""
safe_ids = [self._safe_doc_id(id) for id in ids]
# Create properly escaped string literals for SQL
id_list = ",".join(f"'{id}'" for id in safe_ids)
sql = f"DELETE FROM {self._config.schema_name}.{self._table_name} WHERE id IN ({id_list})"
connection = self._ensure_connection()
with connection.cursor() as cursor:
cursor.execute(sql)
def delete_by_metadata_field(self, key: str, value: str) -> None:
"""Delete documents by metadata field."""
# Check if table exists before attempting delete
if not self._table_exists():
logger.warning("Table %s.%s does not exist, skipping delete", self._config.schema_name, self._table_name)
return
# Execute delete through write queue
self._execute_write(self._delete_by_metadata_field_impl, key, value)
def _delete_by_metadata_field_impl(self, key: str, value: str) -> None:
"""Implementation of delete by metadata field (executed in write worker thread)."""
connection = self._ensure_connection()
with connection.cursor() as cursor:
# Using JSON path to filter with parameterized query
# Note: JSON path requires literal key name, cannot be parameterized
# Use json_extract_string function for ClickZetta compatibility
sql = (f"DELETE FROM {self._config.schema_name}.{self._table_name} "
f"WHERE json_extract_string({Field.METADATA_KEY.value}, '$.{key}') = ?")
cursor.execute(sql, [value])
def search_by_vector(self, query_vector: list[float], **kwargs: Any) -> list[Document]:
"""Search for documents by vector similarity."""
top_k = kwargs.get("top_k", 10)
score_threshold = kwargs.get("score_threshold", 0.0)
document_ids_filter = kwargs.get("document_ids_filter")
# Handle filter parameter from canvas (workflow)
filter_param = kwargs.get("filter", {})
# Build filter clause
filter_clauses = []
if document_ids_filter:
safe_doc_ids = [str(id).replace("'", "''") for id in document_ids_filter]
doc_ids_str = ",".join(f"'{id}'" for id in safe_doc_ids)
# Use json_extract_string function for ClickZetta compatibility
filter_clauses.append(
f"json_extract_string({Field.METADATA_KEY.value}, '$.document_id') IN ({doc_ids_str})"
)
# No need for dataset_id filter since each dataset has its own table
# Add distance threshold based on distance function
vector_dimension = len(query_vector)
if self._config.vector_distance_function == "cosine_distance":
# For cosine distance, smaller is better (0 = identical, 2 = opposite)
distance_func = "COSINE_DISTANCE"
if score_threshold > 0:
query_vector_str = f"CAST('[{self._format_vector_simple(query_vector)}]' AS VECTOR({vector_dimension}))"
filter_clauses.append(f"{distance_func}({Field.VECTOR.value}, "
f"{query_vector_str}) < {2 - score_threshold}")
else:
# For L2 distance, smaller is better
distance_func = "L2_DISTANCE"
if score_threshold > 0:
query_vector_str = f"CAST('[{self._format_vector_simple(query_vector)}]' AS VECTOR({vector_dimension}))"
filter_clauses.append(f"{distance_func}({Field.VECTOR.value}, "
f"{query_vector_str}) < {score_threshold}")
where_clause = " AND ".join(filter_clauses) if filter_clauses else "1=1"
# Execute vector search query
query_vector_str = f"CAST('[{self._format_vector_simple(query_vector)}]' AS VECTOR({vector_dimension}))"
search_sql = f"""
SELECT id, {Field.CONTENT_KEY.value}, {Field.METADATA_KEY.value},
{distance_func}({Field.VECTOR.value}, {query_vector_str}) AS distance
FROM {self._config.schema_name}.{self._table_name}
WHERE {where_clause}
ORDER BY distance
LIMIT {top_k}
"""
documents = []
connection = self._ensure_connection()
with connection.cursor() as cursor:
# Use hints parameter for vector search optimization
search_hints = {
'hints': {
'sdk.job.timeout': 60, # Increase timeout for vector search
'cz.sql.job.fast.mode': True,
'cz.storage.parquet.vector.index.read.memory.cache': True
}
}
cursor.execute(search_sql, parameters=search_hints)
results = cursor.fetchall()
for row in results:
# Parse metadata from JSON string (may be double-encoded)
try:
if row[2]:
metadata = json.loads(row[2])
# If result is a string, it's double-encoded JSON - parse again
if isinstance(metadata, str):
metadata = json.loads(metadata)
if not isinstance(metadata, dict):
metadata = {}
else:
metadata = {}
except (json.JSONDecodeError, TypeError) as e:
logger.error("JSON parsing failed: %s", e)
# Fallback: extract document_id with regex
import re
doc_id_match = re.search(r'"document_id":\s*"([^"]+)"', str(row[2] or ''))
metadata = {"document_id": doc_id_match.group(1)} if doc_id_match else {}
# Ensure required fields are set
metadata["doc_id"] = row[0] # segment id
# Ensure document_id exists (critical for Dify's format_retrieval_documents)
if "document_id" not in metadata:
metadata["document_id"] = row[0] # fallback to segment id
# Add score based on distance
if self._config.vector_distance_function == "cosine_distance":
metadata["score"] = 1 - (row[3] / 2)
else:
metadata["score"] = 1 / (1 + row[3])
doc = Document(page_content=row[1], metadata=metadata)
documents.append(doc)
return documents
def search_by_full_text(self, query: str, **kwargs: Any) -> list[Document]:
"""Search for documents using full-text search with inverted index."""
if not self._config.enable_inverted_index:
logger.warning("Full-text search is not enabled. Enable inverted index in config.")
return []
top_k = kwargs.get("top_k", 10)
document_ids_filter = kwargs.get("document_ids_filter")
# Handle filter parameter from canvas (workflow)
filter_param = kwargs.get("filter", {})
# Build filter clause
filter_clauses = []
if document_ids_filter:
safe_doc_ids = [str(id).replace("'", "''") for id in document_ids_filter]
doc_ids_str = ",".join(f"'{id}'" for id in safe_doc_ids)
# Use json_extract_string function for ClickZetta compatibility
filter_clauses.append(
f"json_extract_string({Field.METADATA_KEY.value}, '$.document_id') IN ({doc_ids_str})"
)
# No need for dataset_id filter since each dataset has its own table
# Use match_all function for full-text search
# match_all requires all terms to be present
# Use simple quote escaping for MATCH_ALL since it needs to be in the WHERE clause
escaped_query = query.replace("'", "''")
filter_clauses.append(f"MATCH_ALL({Field.CONTENT_KEY.value}, '{escaped_query}')")
where_clause = " AND ".join(filter_clauses)
# Execute full-text search query
search_sql = f"""
SELECT id, {Field.CONTENT_KEY.value}, {Field.METADATA_KEY.value}
FROM {self._config.schema_name}.{self._table_name}
WHERE {where_clause}
LIMIT {top_k}
"""
documents = []
connection = self._ensure_connection()
with connection.cursor() as cursor:
try:
# Use hints parameter for full-text search optimization
fulltext_hints = {
'hints': {
'sdk.job.timeout': 30, # Timeout for full-text search
'cz.sql.job.fast.mode': True,
'cz.sql.index.prewhere.enabled': True
}
}
cursor.execute(search_sql, parameters=fulltext_hints)
results = cursor.fetchall()
for row in results:
# Parse metadata from JSON string (may be double-encoded)
try:
if row[2]:
metadata = json.loads(row[2])
# If result is a string, it's double-encoded JSON - parse again
if isinstance(metadata, str):
metadata = json.loads(metadata)
if not isinstance(metadata, dict):
metadata = {}
else:
metadata = {}
except (json.JSONDecodeError, TypeError) as e:
logger.error("JSON parsing failed: %s", e)
# Fallback: extract document_id with regex
import re
doc_id_match = re.search(r'"document_id":\s*"([^"]+)"', str(row[2] or ''))
metadata = {"document_id": doc_id_match.group(1)} if doc_id_match else {}
# Ensure required fields are set
metadata["doc_id"] = row[0] # segment id
# Ensure document_id exists (critical for Dify's format_retrieval_documents)
if "document_id" not in metadata:
metadata["document_id"] = row[0] # fallback to segment id
# Add a relevance score for full-text search
metadata["score"] = 1.0 # Clickzetta doesn't provide relevance scores
doc = Document(page_content=row[1], metadata=metadata)
documents.append(doc)
except (RuntimeError, ValueError, TypeError, ConnectionError) as e:
logger.exception("Full-text search failed")
# Fallback to LIKE search if full-text search fails
return self._search_by_like(query, **kwargs)
return documents
def _search_by_like(self, query: str, **kwargs: Any) -> list[Document]:
"""Fallback search using LIKE operator."""
top_k = kwargs.get("top_k", 10)
document_ids_filter = kwargs.get("document_ids_filter")
# Handle filter parameter from canvas (workflow)
filter_param = kwargs.get("filter", {})
# Build filter clause
filter_clauses = []
if document_ids_filter:
safe_doc_ids = [str(id).replace("'", "''") for id in document_ids_filter]
doc_ids_str = ",".join(f"'{id}'" for id in safe_doc_ids)
# Use json_extract_string function for ClickZetta compatibility
filter_clauses.append(
f"json_extract_string({Field.METADATA_KEY.value}, '$.document_id') IN ({doc_ids_str})"
)
# No need for dataset_id filter since each dataset has its own table
# Use simple quote escaping for LIKE clause
escaped_query = query.replace("'", "''")
filter_clauses.append(f"{Field.CONTENT_KEY.value} LIKE '%{escaped_query}%'")
where_clause = " AND ".join(filter_clauses)
search_sql = f"""
SELECT id, {Field.CONTENT_KEY.value}, {Field.METADATA_KEY.value}
FROM {self._config.schema_name}.{self._table_name}
WHERE {where_clause}
LIMIT {top_k}
"""
documents = []
connection = self._ensure_connection()
with connection.cursor() as cursor:
# Use hints parameter for LIKE search optimization
like_hints = {
'hints': {
'sdk.job.timeout': 20, # Timeout for LIKE search
'cz.sql.job.fast.mode': True
}
}
cursor.execute(search_sql, parameters=like_hints)
results = cursor.fetchall()
for row in results:
# Parse metadata from JSON string (may be double-encoded)
try:
if row[2]:
metadata = json.loads(row[2])
# If result is a string, it's double-encoded JSON - parse again
if isinstance(metadata, str):
metadata = json.loads(metadata)
if not isinstance(metadata, dict):
metadata = {}
else:
metadata = {}
except (json.JSONDecodeError, TypeError) as e:
logger.error("JSON parsing failed: %s", e)
# Fallback: extract document_id with regex
import re
doc_id_match = re.search(r'"document_id":\s*"([^"]+)"', str(row[2] or ''))
metadata = {"document_id": doc_id_match.group(1)} if doc_id_match else {}
# Ensure required fields are set
metadata["doc_id"] = row[0] # segment id
# Ensure document_id exists (critical for Dify's format_retrieval_documents)
if "document_id" not in metadata:
metadata["document_id"] = row[0] # fallback to segment id
metadata["score"] = 0.5 # Lower score for LIKE search
doc = Document(page_content=row[1], metadata=metadata)
documents.append(doc)
return documents
def delete(self) -> None:
"""Delete the entire collection."""
connection = self._ensure_connection()
with connection.cursor() as cursor:
cursor.execute(f"DROP TABLE IF EXISTS {self._config.schema_name}.{self._table_name}")
def _format_vector_simple(self, vector: list[float]) -> str:
"""Simple vector formatting for SQL queries."""
return ','.join(map(str, vector))
def _safe_doc_id(self, doc_id: str) -> str:
"""Ensure doc_id is safe for SQL and doesn't contain special characters."""
if not doc_id:
return str(uuid.uuid4())
# Remove or replace potentially problematic characters
safe_id = str(doc_id)
# Only allow alphanumeric, hyphens, underscores
safe_id = ''.join(c for c in safe_id if c.isalnum() or c in '-_')
if not safe_id: # If all characters were removed
return str(uuid.uuid4())
return safe_id[:255] # Limit length
class ClickzettaVectorFactory(AbstractVectorFactory):
"""Factory for creating Clickzetta vector instances."""
def init_vector(self, dataset: Dataset, attributes: list, embeddings: Embeddings) -> BaseVector:
"""Initialize a Clickzetta vector instance."""
# Get configuration from environment variables or dataset config
config = ClickzettaConfig(
username=dify_config.CLICKZETTA_USERNAME or "",
password=dify_config.CLICKZETTA_PASSWORD or "",
instance=dify_config.CLICKZETTA_INSTANCE or "",
service=dify_config.CLICKZETTA_SERVICE or "api.clickzetta.com",
workspace=dify_config.CLICKZETTA_WORKSPACE or "quick_start",
vcluster=dify_config.CLICKZETTA_VCLUSTER or "default_ap",
schema_name=dify_config.CLICKZETTA_SCHEMA or "dify",
batch_size=dify_config.CLICKZETTA_BATCH_SIZE or 100,
enable_inverted_index=dify_config.CLICKZETTA_ENABLE_INVERTED_INDEX or True,
analyzer_type=dify_config.CLICKZETTA_ANALYZER_TYPE or "chinese",
analyzer_mode=dify_config.CLICKZETTA_ANALYZER_MODE or "smart",
vector_distance_function=dify_config.CLICKZETTA_VECTOR_DISTANCE_FUNCTION or "cosine_distance",
)
# Use dataset collection name as table name
collection_name = Dataset.gen_collection_name_by_id(dataset.id).lower()
return ClickzettaVector(collection_name=collection_name, config=config)

View File

@@ -172,6 +172,10 @@ class Vector:
from core.rag.datasource.vdb.matrixone.matrixone_vector import MatrixoneVectorFactory
return MatrixoneVectorFactory
case VectorType.CLICKZETTA:
from core.rag.datasource.vdb.clickzetta.clickzetta_vector import ClickzettaVectorFactory
return ClickzettaVectorFactory
case _:
raise ValueError(f"Vector store {vector_type} is not supported.")

View File

@@ -30,3 +30,4 @@ class VectorType(StrEnum):
TABLESTORE = "tablestore"
HUAWEI_CLOUD = "huawei_cloud"
MATRIXONE = "matrixone"
CLICKZETTA = "clickzetta"