Issue 4130 similaritySearch throwing Exception due to schema name fix

[wilocu]

#4130

Problem
When using PgVectorStore with schema names containing hyphens (e.g., demo-1998), similaritySearch() operations would throw SQL grammar exceptions due to
unquoted PostgreSQL identifiers in generated SQL.

Solution

• Added proper PostgreSQL identifier quoting in getFullyQualifiedTableName() to generate "schema"."table" format
• Updated schema creation DDL to use quoted schema names
• Enhanced schema validator to accept hyphenated names while maintaining security against SQL injection
• Updated error messages to show proper quoted syntax in help text

Changes

• PgVectorStore.java: Fixed SQL generation to use quoted identifiers
• PgVectorSchemaValidator.java: Updated validation logic and error messages
• PgVectorStoreSchemaQuotingTest.java: Added test coverage for hyphenated schema names

Example

Before (❌ SQL error):
SELECT * FROM demo-1998.vector_store -- Syntax error

After (✅ Valid PostgreSQL):
SELECT * FROM "demo-1998"."vector_store" -- Works correctly

Testing

• All existing tests pass
• New test verifies proper quoting for hyphenated schema names
• Verified SQL generation produces valid PostgreSQL syntax

Signed-off-by: Mattia Pasetto [email protected]

I am still pretty new to contributions, please double check and let me know of any errors.

[markpollack]

Thanks! There is certainly a bug to fix.

I am not an expert on these database best practices, so I turned to some investigation on chatgpt.

Here is the reply. would like to know your thoughts @wilocu

This is a real bug, and the core fix (always quoting identifiers) is right. The piece to adjust is the validator.

---

Quoting identifiers in getFullyQualifiedTableName() and DDL is the correct fix for the hyphenated-schema bug. A couple of tweaks will align this with PostgreSQL rules and avoid false rejections:

1. Identifier rules

• PostgreSQL limits identifiers to 63 bytes (UTF-8), not 64 chars.
• Numeric names, reserved words, and symbols (e.g., -, .) are legal when quoted.
• Double quotes inside names must be escaped as "".

2. Validator

• Drop bans on numeric-only, keywords, ;, --, etc. If we always quote+escape, those are safe.
• Keep: non-null, non-empty, ≤63 bytes, and no NUL/control chars.
• Existence check: use SELECT to_regclass(quote_ident($1)||'.'||quote_ident($2)) IS NOT NULL to respect case/quoting instead of information_schema.

3. Quoting helper
   Ensure the FQN builder escapes quotes:

private static String q(String s) { return "\"" + s.replace("\"", "\"\"") + "\""; }
// usage: q(schema) + "." + q(table)

4. Tests
   Add cases for "demo-1998", "123", "select", a"b, non-ASCII, and 63-byte boundary.

With those adjustments, the quoting fix resolves the bug without over-restricting valid PostgreSQL identifiers.