Spaces:
Sleeping
Sleeping
File size: 13,100 Bytes
d96b38a 1ca0f60 d96b38a 1ca0f60 d96b38a 8a40a5a d96b38a 8a40a5a d96b38a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 | """Read-only SQL tool.
Three jobs:
1. **Introspect the schema once at startup** so the LLM has table /
column types in its prompt without us hand-maintaining a catalog.
2. **Translate a natural-language intent into a single ``SELECT``** via
``ChatOpenAI`` with structured output. Refuses anything that isn't a
read query.
3. **Execute the SQL safely.** Defense in depth:
- The DB connection should already be a read-only role
(``adaptive_rag_ro`` in our seed script).
- Statement-level allowlist: only ``SELECT`` / ``WITH`` allowed.
- ``statement_timeout`` set per-session so runaway queries die fast.
- A ``LIMIT N`` is appended if the SQL doesn't already have one, so
accidentally returning a million rows can't OOM the UI.
This is intentionally a thin tool — it is *not* an agent. The dispatcher
calls it once with a NL intent, and we either return rows or raise.
"""
from __future__ import annotations
import logging
import re
from dataclasses import dataclass, field
from typing import Any
from langchain_openai import ChatOpenAI
from pydantic import BaseModel, Field
from sqlalchemy import create_engine, inspect, text
from sqlalchemy.engine import Engine
from sqlalchemy.exc import SQLAlchemyError
from src.config import settings
logger = logging.getLogger(__name__)
# Single-statement, must start with SELECT or WITH (CTEs that resolve to a
# SELECT). Reject anything that smells like a write or DDL even if it would
# also be blocked by the read-only role — defense in depth, and surfaces
# better error messages than letting Postgres reject mid-query.
_SELECT_PATTERN = re.compile(r"^\s*(?:WITH\b|SELECT\b)", re.IGNORECASE)
_FORBIDDEN_KEYWORDS = re.compile(
r"\b(?:INSERT|UPDATE|DELETE|MERGE|TRUNCATE|DROP|ALTER|CREATE|"
r"GRANT|REVOKE|COPY|VACUUM|ANALYZE|REINDEX|CLUSTER|"
r"COMMENT\s+ON|SECURITY\s+DEFINER|DO\s+\$\$)\b",
re.IGNORECASE,
)
_LIMIT_PATTERN = re.compile(r"\blimit\s+\d+\b", re.IGNORECASE)
_MULTI_STATEMENT = re.compile(r";\s*\S")
SQL_SYSTEM_PROMPT = """You translate natural-language data questions into a
single read-only PostgreSQL SELECT query.
Hard rules:
- Output exactly one statement.
- The statement MUST start with SELECT or WITH.
- NEVER write, modify or define schema (no INSERT, UPDATE, DELETE, CREATE,
DROP, ALTER, etc.). The connection is read-only and will reject them
anyway, but don't generate them.
- Use only the tables and columns shown in the schema below.
- When the question implies a time window like "last month" or "this year",
prefer ``WHERE col >= NOW() - INTERVAL 'N units'`` over hardcoded dates.
- Add an explicit ORDER BY when the question implies ranking ("top",
"most", "biggest").
- If the question is ambiguous or cannot be answered from this schema,
return a single SELECT that explains the gap, e.g.
``SELECT 'cannot answer: <reason>' AS error;``
- Output only the SQL — no commentary, no markdown fences.
"""
SQL_USER_TEMPLATE = """Schema:
{schema}
Question:
{intent}
"""
class _SqlOutput(BaseModel):
"""Structured output the SQL LLM is forced into."""
sql: str = Field(
...,
description=(
"A single PostgreSQL SELECT statement that answers the question. "
"No trailing semicolon, no markdown fences."
),
)
class SqlToolError(Exception):
"""Raised when the SQL tool can't produce or execute a safe query."""
@dataclass
class SqlResult:
"""One execution of one query."""
intent: str
sql: str
columns: list[str] = field(default_factory=list)
rows: list[dict[str, Any]] = field(default_factory=list)
truncated: bool = False
elapsed_ms: float = 0.0
@property
def row_count(self) -> int:
return len(self.rows)
class SqlTool:
"""Schema-aware NL\u2192SQL tool with read-only execution."""
def __init__(
self,
database_url: str | None = None,
*,
statement_timeout_sec: int | None = None,
row_limit: int | None = None,
translator_model: str | None = None,
) -> None:
url = database_url or settings.SQL_DATABASE_URL
if not url:
raise SqlToolError(
"SQL_DATABASE_URL is not set. The SQL tool needs a connection "
"string. Run scripts/seed_demo_data.py first or point at your "
"own Postgres."
)
self.database_url = url
self.statement_timeout_sec = (
statement_timeout_sec or settings.SQL_QUERY_TIMEOUT_SEC
)
self.row_limit = row_limit or settings.SQL_ROW_LIMIT
try:
self._engine: Engine = create_engine(url, future=True, pool_pre_ping=True)
except SQLAlchemyError as exc:
raise SqlToolError(f"Cannot create SQL engine: {exc}") from exc
# Cache the schema description — it doesn't change at runtime and
# we'd otherwise pay an introspection round-trip on every query.
self._schema_text = self._describe_schema()
self._llm = ChatOpenAI(
model=translator_model or settings.SQL_MODEL,
# Ignored by gpt-5.6 reasoning models (only temperature=1 is
# supported); langchain-openai strips it automatically.
temperature=0.0,
# Pro-mode reasoning tokens count toward this budget — 400 was
# enough for gpt-4.1-mini but truncates Luna Pro mid-reasoning.
max_tokens=8000,
reasoning=settings.sql_reasoning,
).with_structured_output(_SqlOutput)
logger.info(
f"SqlTool ready (db={self._safe_url()}, "
f"timeout={self.statement_timeout_sec}s, row_limit={self.row_limit})"
)
# ---- public API ---------------------------------------------------
@property
def schema_text(self) -> str:
"""Human-readable schema description (cached)."""
return self._schema_text
def schema_summary(self) -> str:
"""One-line-per-table summary, suitable for the router prompt."""
try:
inspector = inspect(self._engine)
lines: list[str] = []
for table in sorted(inspector.get_table_names(schema="public")):
cols = [c["name"] for c in inspector.get_columns(table, schema="public")]
preview = ", ".join(cols[:6])
if len(cols) > 6:
preview += ", \u2026"
lines.append(f"- {table}: {preview}")
return "\n".join(lines)
except SQLAlchemyError as exc:
logger.warning(f"Schema summary failed: {exc}")
return "(schema introspection failed)"
def answer(self, intent: str) -> SqlResult:
"""End-to-end: NL intent -> SQL -> rows."""
sql = self.translate(intent)
return self.execute(sql, intent=intent)
def translate(self, intent: str) -> str:
"""Ask the LLM for a single SELECT statement matching ``intent``."""
intent = (intent or "").strip()
if not intent:
raise SqlToolError("Empty intent — nothing to translate.")
messages = [
{"role": "system", "content": SQL_SYSTEM_PROMPT},
{
"role": "user",
"content": SQL_USER_TEMPLATE.format(
schema=self._schema_text,
intent=intent,
),
},
]
from src.observability import get_callback_handler
try:
output: _SqlOutput = self._llm.invoke(
messages,
config={
"callbacks": get_callback_handler(),
"run_name": "sql_tool.translate",
"metadata": {"langfuse_tags": ["sql_tool", "nl2sql"]},
},
)
except Exception as exc:
raise SqlToolError(f"LLM SQL translation failed: {exc}") from exc
sql = self._clean(output.sql)
self._validate(sql)
return sql
def execute(self, sql: str, *, intent: str = "") -> SqlResult:
"""Validate and run ``sql``, returning a :class:`SqlResult`."""
sql = self._clean(sql)
self._validate(sql)
sql_to_run = self._inject_limit(sql)
import time
t0 = time.perf_counter()
try:
with self._engine.connect() as conn:
# Per-session timeout so runaway plans die fast. Postgres
# accepts an integer string of milliseconds.
conn.exec_driver_sql(
f"SET statement_timeout = {self.statement_timeout_sec * 1000}"
)
# Force read-only at the transaction level too. With a RO
# role this is redundant, but it makes the intent explicit
# and protects against misconfigured connection strings.
conn.exec_driver_sql("SET TRANSACTION READ ONLY")
result = conn.execute(text(sql_to_run))
rows = result.mappings().all()
columns = list(result.keys())
except SQLAlchemyError as exc:
raise SqlToolError(f"SQL execution failed: {exc}") from exc
elapsed_ms = (time.perf_counter() - t0) * 1000
truncated = len(rows) >= self.row_limit and not _LIMIT_PATTERN.search(sql)
materialised = [dict(r) for r in rows]
logger.info(
f"SqlTool.execute: {len(materialised)} rows ({elapsed_ms:.0f}ms) "
f"truncated={truncated}"
)
return SqlResult(
intent=intent,
sql=sql,
columns=columns,
rows=materialised,
truncated=truncated,
elapsed_ms=elapsed_ms,
)
# ---- internals ----------------------------------------------------
@staticmethod
def _clean(sql: str) -> str:
sql = (sql or "").strip()
# Strip surrounding markdown fence in case the LLM ignored the prompt.
if sql.startswith("```"):
sql = sql.strip("`")
# After stripping backticks, drop a leading ``sql`` language tag.
sql = re.sub(r"^\s*sql\b", "", sql, flags=re.IGNORECASE).strip()
# Drop single trailing semicolon — we'll add LIMIT before it otherwise.
sql = sql.rstrip(";").strip()
return sql
@staticmethod
def _validate(sql: str) -> None:
if not sql:
raise SqlToolError("Empty SQL produced.")
if not _SELECT_PATTERN.match(sql):
raise SqlToolError(
"Only SELECT / WITH statements are allowed. "
f"Got: {sql.split()[0] if sql else '?'}\u2026"
)
if _MULTI_STATEMENT.search(sql):
raise SqlToolError("Multiple statements are not allowed.")
if _FORBIDDEN_KEYWORDS.search(sql):
raise SqlToolError(
"SQL contains a forbidden keyword (write / DDL operation)."
)
def _inject_limit(self, sql: str) -> str:
if _LIMIT_PATTERN.search(sql):
return sql
return f"{sql}\nLIMIT {self.row_limit}"
def _safe_url(self) -> str:
# Hide credentials in log lines.
url = self.database_url
if "://" in url and "@" in url:
scheme, rest = url.split("://", 1)
creds, host = rest.rsplit("@", 1)
if ":" in creds:
user, _ = creds.split(":", 1)
return f"{scheme}://{user}:****@{host}"
return url
def _describe_schema(self) -> str:
try:
inspector = inspect(self._engine)
tables = sorted(inspector.get_table_names(schema="public"))
except SQLAlchemyError as exc:
raise SqlToolError(f"Cannot introspect schema: {exc}") from exc
if not tables:
return "(no tables found in 'public' schema)"
chunks: list[str] = []
for table in tables:
cols = inspector.get_columns(table, schema="public")
pk = inspector.get_pk_constraint(table, schema="public").get(
"constrained_columns", []
) or []
fks = inspector.get_foreign_keys(table, schema="public") or []
col_lines: list[str] = []
for c in cols:
pk_marker = " PRIMARY KEY" if c["name"] in pk else ""
nullable = "" if c.get("nullable", True) else " NOT NULL"
col_lines.append(f" {c['name']} {c['type']}{nullable}{pk_marker}")
fk_lines: list[str] = []
for fk in fks:
local = ", ".join(fk["constrained_columns"])
remote_table = fk["referred_table"]
remote_cols = ", ".join(fk["referred_columns"])
fk_lines.append(f" FOREIGN KEY ({local}) -> {remote_table}({remote_cols})")
block = f"TABLE {table} (\n" + ",\n".join(col_lines)
if fk_lines:
block += "\n --\n" + "\n".join(fk_lines)
block += "\n)"
chunks.append(block)
return "\n\n".join(chunks)
|