""" Baseline inference script for the SQL Query Environment. Runs Qwen-72B across all task domains and difficulty levels, printing results in the [START]/[STEP]/[END] format required by the OpenEnv evaluation pipeline. Usage: export HF_TOKEN="your-token" python inference.py """ import asyncio import os import textwrap from typing import List, Optional from openai import OpenAI from sql_query_env import SqlQueryAction, SqlQueryEnv # --- Config --- IMAGE_NAME = os.getenv("IMAGE_NAME") API_KEY = os.getenv("HF_TOKEN") or os.getenv("API_KEY") API_BASE_URL = os.getenv("API_BASE_URL") or "https://router.huggingface.co/v1" MODEL_NAME = os.getenv("MODEL_NAME") or "Qwen/Qwen2.5-72B-Instruct" BENCHMARK = "sql_query_env" MAX_STEPS_PER_TASK = 6 TEMPERATURE = 0.3 MAX_TOKENS = 500 # one task from each domain at each difficulty level TASK_IDS = [ "company_easy_1", "hospital_easy_1", "ecommerce_easy_1", "company_medium_1", "hospital_medium_1", "ecommerce_medium_1", "company_hard_1", "hospital_hard_1", "ecommerce_hard_1", "company_hard_2", # self-join challenge "hospital_hard_2", # repeat visits "ecommerce_hard_2", # anti-join / dead stock ] SUCCESS_THRESHOLD = 0.5 # system prompt for the LLM agent SYSTEM_PROMPT = textwrap.dedent(""" You are an expert SQL query writer. You are given a database schema and a natural language question. Write a correct SQL query to answer it. Rules: 1. Write ONLY valid SQLite SQL SELECT queries 2. Do NOT use DROP, DELETE, INSERT, UPDATE, ALTER, CREATE 3. Pay attention to the schema - different databases have different tables 4. Use appropriate JOINs when data spans multiple tables 5. Match the expected column names in your output 6. Use table aliases for clarity (e.g., FROM employees e) 7. Use column aliases with AS (e.g., SUM(salary) AS total) 8. Handle NULLs properly with IS NULL / IS NOT NULL / COALESCE 9. Return ONLY the raw SQL query - no markdown, no explanations """).strip() # --- Logging (required format) --- def log_start(task: str, env: str, model: str) -> None: print(f"[START] task={task} env={env} model={model}", flush=True) def log_step(step: int, action: str, reward: float, done: bool, error: Optional[str]) -> None: action_clean = action.replace("\n", " ").replace("\r", "").strip() error_val = error if error else "null" done_val = str(done).lower() print( f"[STEP] step={step} action={action_clean} reward={reward:.2f} done={done_val} error={error_val}", flush=True, ) def log_end(success: bool, steps: int, score: float, rewards: List[float]) -> None: rewards_str = ",".join(f"{r:.2f}" for r in rewards) print(f"[END] success={str(success).lower()} steps={steps} score={score:.3f} rewards={rewards_str}", flush=True) # --- LLM interaction --- def build_user_prompt( question: str, schema: str, feedback: str, expected_columns: List[str], expected_row_count: int, step: int, history: List[dict], diagnostics: List[dict], database_domain: str, ) -> str: """Build the prompt we send to the model with all available context.""" history_block = "" if history: lines = [f" Step {h['step']}: score={h['score']:.2f} | {h['query']}" for h in history[-3:]] history_block = "Previous attempts:\n" + "\n".join(lines) diag_block = "" if diagnostics: lines = [f" [{d['type']}] {d['message']} -> {d['suggestion']}" for d in diagnostics[-3:]] diag_block = "Diagnostics from last attempt:\n" + "\n".join(lines) return textwrap.dedent(f""" DATABASE DOMAIN: {database_domain} DATABASE SCHEMA: {schema} QUESTION: {question} EXPECTED OUTPUT COLUMNS: {', '.join(expected_columns)} EXPECTED ROW COUNT: {expected_row_count} {f'FEEDBACK:{chr(10)}{feedback}' if step > 1 else ''} {diag_block} {history_block} Step {step}: Write a SQL query to answer the question. Return ONLY the raw SQL query, no markdown. """).strip() def extract_sql(text: str) -> str: """Clean up LLM output to get just the SQL.""" text = text.strip() if text.startswith("```sql"): text = text[6:] elif text.startswith("```"): text = text[3:] if text.endswith("```"): text = text[:-3] text = text.strip() if text.endswith(";"): text = text[:-1].strip() return text def get_sql_query( client: OpenAI, question: str, schema: str, feedback: str, expected_columns: List[str], expected_row_count: int, step: int, history: List[dict], diagnostics: List[dict], database_domain: str, ) -> str: """Ask the LLM for a SQL query.""" user_prompt = build_user_prompt( question, schema, feedback, expected_columns, expected_row_count, step, history, diagnostics, database_domain, ) try: completion = client.chat.completions.create( model=MODEL_NAME, messages=[ {"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": user_prompt}, ], temperature=TEMPERATURE, max_tokens=MAX_TOKENS, stream=False, ) raw = (completion.choices[0].message.content or "").strip() return extract_sql(raw) if raw else "SELECT 1" except Exception as exc: print(f"[DEBUG] Model request failed: {exc}", flush=True) return "SELECT 1" # --- Main evaluation loop --- async def run_task(client: OpenAI, env: SqlQueryEnv, task_id: str) -> float: """Run one task, return the best score achieved.""" rewards: List[float] = [] steps_taken = 0 score = 0.0 success = False log_start(task=task_id, env=BENCHMARK, model=MODEL_NAME) try: result = await env.reset() obs = result.observation feedback = obs.feedback diagnostics = obs.diagnostics for step in range(1, MAX_STEPS_PER_TASK + 1): if result.done: break sql_query = get_sql_query( client, obs.question, obs.schema_description, feedback, obs.expected_columns, obs.expected_row_count, step, obs.history, diagnostics, obs.database_domain, ) result = await env.step(SqlQueryAction(query=sql_query)) obs = result.observation reward = result.reward or 0.0 done = result.done error = obs.query_error rewards.append(reward) steps_taken = step feedback = obs.feedback diagnostics = obs.diagnostics log_step(step=step, action=sql_query, reward=reward, done=done, error=error) if done: break score = max(rewards) if rewards else 0.0 score = min(max(score, 0.0), 1.0) success = score >= SUCCESS_THRESHOLD except Exception as e: print(f"[DEBUG] Task {task_id} error: {e}", flush=True) finally: log_end(success=success, steps=steps_taken, score=score, rewards=rewards) return score async def main() -> None: """Run all tasks and print summary.""" client = OpenAI(base_url=API_BASE_URL, api_key=API_KEY) if IMAGE_NAME: env = await SqlQueryEnv.from_docker_image(IMAGE_NAME) else: base_url = os.getenv("ENV_BASE_URL", "http://localhost:8000") env = SqlQueryEnv(base_url=base_url) try: async with env: scores = [] for task_id in TASK_IDS: score = await run_task(client, env, task_id) scores.append(score) # print summary print(f"\n{'='*50}", flush=True) print(f"[SUMMARY] Model: {MODEL_NAME}", flush=True) print(f"[SUMMARY] Average score: {sum(scores)/len(scores):.2f}", flush=True) print(f"{'='*50}", flush=True) domains = {"company": [], "hospital": [], "ecommerce": []} for tid, sc in zip(TASK_IDS, scores): domain = tid.split("_")[0] domains[domain].append(sc) print(f" {tid}: {sc:.2f}", flush=True) print(f"\nPer-domain averages:", flush=True) for domain, dscores in domains.items(): avg = sum(dscores) / len(dscores) if dscores else 0 print(f" {domain}: {avg:.2f}", flush=True) except Exception as e: print(f"[DEBUG] Environment connection error: {e}", flush=True) if __name__ == "__main__": asyncio.run(main())