NL-to-PL/SQL Basic (nl2plsql-basic)

This is the 基础版 (free) NL-to-PL/SQL skill. It handles basic tasks only.

NL-to-PL/SQL is translating a natural-language requirement into a procedural extension of SQL (broadly "PL/SQL"). This skill covers two dialects:

• Oracle PL/SQL
• PostgreSQL PL/pgSQL

What basic covers (free)

1. Read and inspect NL-to-PL/SQL seed data in the canonical JSON format.
2. Write a simple Oracle PL/SQL or PostgreSQL PL/pgSQL stored procedure or function that performs a single SQL statement: one SELECT ... INTO, one INSERT (values or INSERT ... SELECT), one UPDATE, or one DELETE, with basic IN/OUT/RETURN parameters.
3. Write basic IF ... THEN ... ELSIF ... ELSE ... END IF control flow around a single statement.
4. Apply the core Oracle↔PostgreSQL differences (see references/dialect-basics.md).

What basic does NOT cover (requires paid nl2plsql-pro)

Do not attempt these in the basic tier; tell the user they require the paid version nl2plsql-pro:

• Triggers (Oracle CREATE TRIGGER; PostgreSQL trigger functions + CREATE TRIGGER).
• Complex procedural logic: cursors, loops beyond a trivial IF, exception handling, dynamic SQL, transaction control.
• Multi-statement bodies beyond a single SQL operation.
• Data generation: template-based or multi-granularity (goal/requirement/implementation) NL-to-PL/SQL data generation.
• Prompt construction: schema pruning, metadata enrichment, chain-of-thought, skeleton prediction.
• Model training: incremental pre-training, SFT, reinforcement learning, error correction.
• Evaluation: execution-match (EX) / exact-match (EM) / execution-accuracy metrics, call suites.
• ProcArena scenarios (A1–A5, B1–B4) and Direct/Interactive modes.

If a request mixes basic and advanced parts, complete only the basic parts and clearly flag the advanced parts as requiring nl2plsql-pro.

Canonical seed-data format

The seed files are JSON objects with three top-level keys: procedure, function, trigger. Each value is a list of records:

{
  "ir": "<natural-language instruction>",
  "plsql": "<target Oracle PL/SQL or PostgreSQL PL/pgSQL code>",
  "database_name": "<snake_case database identifier>",
  "tables": ["<relevant table name>", "..."]
}

• ir is the NL instruction (input); plsql is the ground-truth code (output).
• tables lists only the table names relevant to that record; column-level schema lives elsewhere.
• Read references/data-format.md before loading these large (~48–57 MB) files.

References

• references/data-format.md — seed JSON format, scale, and reading tips.
• references/dialect-basics.md — minimal Oracle PL/SQL and PostgreSQL PL/pgSQL syntax for simple procedures/functions, plus core differences.

Quick utility

scripts/inspect_seed.py prints per-construct counts, distinct-database counts, average table count, and a few samples for any seed JSON.