File size: 17,173 Bytes
e545bf5 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 | """Safe provider-agnostic SQL retrieval agent for the complete SWMM SQLite store.
The language model produces a constrained JSON retrieval plan. This module validates
that plan and executes read-only SELECT statements only. No model-generated SQL is
executed directly.
"""
from __future__ import annotations
import json
import math
import re
import sqlite3
from dataclasses import dataclass
from typing import Any
import pandas as pd
_ALLOWED_FILTER_OPS = {
"eq": "=", "ne": "!=", "gt": ">", "gte": ">=", "lt": "<", "lte": "<=",
"like": "LIKE", "in": "IN", "between": "BETWEEN",
"is_null": "IS NULL", "not_null": "IS NOT NULL",
}
_ALLOWED_AGGS = {"MAX", "MIN", "AVG", "SUM", "COUNT"}
_MAX_ACTIONS = 6
_MAX_ROWS_PER_ACTION = 120
_MAX_CONTEXT_CHARS = 45_000
@dataclass
class RetrievalResult:
context: str
audit: list[dict[str, Any]]
plan: dict[str, Any]
class SafeSQLAgent:
"""Validate and execute bounded, read-only retrieval plans against a ResultDatabase."""
def __init__(self, result_db: Any) -> None:
self.db = result_db
self.conn: sqlite3.Connection = result_db.connection
self._tables = self._read_tables()
self._columns = {table: self._read_columns(table) for table in self._tables}
@staticmethod
def _quote_identifier(name: str) -> str:
return '"' + name.replace('"', '""') + '"'
def _read_tables(self) -> list[str]:
rows = self.conn.execute(
"SELECT name FROM sqlite_master WHERE type='table' AND name NOT LIKE 'sqlite_%' ORDER BY name"
).fetchall()
return [str(r[0]) for r in rows]
def _read_columns(self, table: str) -> list[str]:
if table not in self._tables:
return []
rows = self.conn.execute(f"PRAGMA table_info({self._quote_identifier(table)})").fetchall()
return [str(r[1]) for r in rows]
def _row_count(self, table: str) -> int:
try:
return int(self.conn.execute(
f"SELECT COUNT(*) FROM {self._quote_identifier(table)}"
).fetchone()[0])
except Exception:
return 0
@staticmethod
def _tokens(text: str) -> set[str]:
return {t.lower() for t in re.findall(r"[A-Za-z0-9_:.\-]+", text) if len(t) > 1}
def schema_context(self, question: str, max_tables: int = 16) -> str:
"""Return a compact, question-ranked schema catalogue for the planning call."""
q_tokens = self._tokens(question)
synonyms = {
"node": {"node", "junction", "outfall", "storage", "manhole", "flood", "head", "depth"},
"link": {"link", "conduit", "pipe", "pump", "weir", "orifice", "outlet", "flow", "velocity", "capacity"},
"subcatchment": {"subcatchment", "catchment", "runoff", "rainfall", "infiltration", "hydrology"},
"input": {"input", "option", "roughness", "diameter", "length", "invert", "curve", "pattern", "control"},
"timeseries": {"time", "timeseries", "duration", "when", "peak", "first", "last", "hour"},
}
expanded = set(q_tokens)
for _, words in synonyms.items():
if q_tokens & words:
expanded |= words
always = {
"simulation_metadata", "simulation_warnings", "node_summary", "link_summary",
"subcatchment_summary", "model_input_section_catalog",
}
scored: list[tuple[float, str]] = []
for table in self._tables:
hay = self._tokens(table + " " + " ".join(self._columns[table]))
score = float(len(expanded & hay) * 5)
for token in expanded:
if token in table.lower():
score += 3
score += sum(1 for c in self._columns[table] if token in c.lower()) * 0.5
if table in always:
score += 4
if table.endswith("_timeseries") and (expanded & synonyms["timeseries"]):
score += 5
if table.startswith("inp_") and (expanded & synonyms["input"]):
score += 5
scored.append((score, table))
selected = [t for _, t in sorted(scored, reverse=True)[:max_tables]]
for table in always:
if table in self._tables and table not in selected:
selected.append(table)
selected = selected[:max_tables]
lines = ["AVAILABLE SQLITE TABLES (question-ranked):"]
for table in selected:
cols = self._columns[table]
count = self._row_count(table)
lines.append(f"- {table} ({count} rows): {', '.join(cols)}")
lines.append(
"The complete database may contain additional inp_<section> tables. "
"Use model_input_lines to search exact source text when a section is not listed."
)
return "\n".join(lines)
def planner_system_prompt(self, schema_context: str) -> str:
return f"""You are a retrieval planner for an EPA SWMM SQLite database.
Return ONLY valid JSON. Do not answer the engineering question.
Create the smallest read-only retrieval plan that supplies enough evidence for a later engineering answer.
Never emit SQL. Use only the operations and fields below.
{schema_context}
JSON format:
{{
"reasoning_summary": "brief retrieval rationale",
"actions": [
{{
"operation": "select",
"table": "table_name",
"columns": ["column1", "column2"],
"filters": [{{"column":"column1","op":"eq|ne|gt|gte|lt|lte|like|in|between|is_null|not_null","value": "value or list"}}],
"order_by": [{{"column":"column1","direction":"asc|desc"}}],
"limit": 40,
"label": "descriptive result label"
}},
{{
"operation": "aggregate",
"table": "table_name",
"group_by": ["optional_column"],
"metrics": [{{"function":"MAX|MIN|AVG|SUM|COUNT","column":"column_or_*","alias":"metric_name"}}],
"filters": [],
"order_by": [{{"column":"metric_alias_or_group_column","direction":"desc"}}],
"limit": 40,
"label": "descriptive result label"
}},
{{
"operation": "search_input",
"term": "text to find in raw SWMM input lines",
"section": "optional SWMM section name without brackets",
"limit": 40,
"label": "descriptive result label"
}},
{{
"operation": "describe_table",
"table": "table_name",
"label": "table structure"
}}
]
}}
Rules:
- Maximum 6 actions and normally 1-4 actions.
- Use exact asset IDs appearing in the user's question as equality filters.
- Prefer summary/aggregate retrieval for broad questions and detailed time-series only for named assets or explicit temporal questions.
- For first/last/peak questions, use ordering and a small limit or an aggregate.
- Retrieve model inputs from inp_* tables or model_input_lines.
- Never request full tables. Keep each limit at or below 120.
- Use engineering evidence from both input and output tables when the question asks for diagnosis or recommendations.
"""
@staticmethod
def parse_plan(text: str) -> dict[str, Any]:
cleaned = text.strip()
cleaned = re.sub(r"^```(?:json)?\s*", "", cleaned, flags=re.IGNORECASE)
cleaned = re.sub(r"\s*```$", "", cleaned)
start, end = cleaned.find("{"), cleaned.rfind("}")
if start >= 0 and end > start:
cleaned = cleaned[start:end + 1]
plan = json.loads(cleaned)
if not isinstance(plan, dict):
raise ValueError("Planner response must be a JSON object")
actions = plan.get("actions", [])
if not isinstance(actions, list):
raise ValueError("Planner actions must be a list")
plan["actions"] = actions[:_MAX_ACTIONS]
return plan
def _validate_table(self, table: str) -> str:
if table not in self._tables:
raise ValueError(f"Unknown table: {table}")
return table
def _validate_column(self, table: str, column: str, *, allow_star: bool = False) -> str:
if allow_star and column == "*":
return column
if column not in self._columns[table]:
raise ValueError(f"Unknown column {column!r} in {table}")
return column
def _build_filters(self, table: str, filters: Any) -> tuple[str, list[Any]]:
if not filters:
return "", []
clauses: list[str] = []
params: list[Any] = []
for item in list(filters)[:12]:
if not isinstance(item, dict):
continue
col = self._validate_column(table, str(item.get("column", "")))
op_key = str(item.get("op", "eq")).lower()
if op_key not in _ALLOWED_FILTER_OPS:
raise ValueError(f"Unsupported filter operation: {op_key}")
sql_op = _ALLOWED_FILTER_OPS[op_key]
qcol = self._quote_identifier(col)
value = item.get("value")
if op_key in {"is_null", "not_null"}:
clauses.append(f"{qcol} {sql_op}")
elif op_key == "in":
values = value if isinstance(value, list) else [value]
values = values[:50]
if not values:
clauses.append("1=0")
else:
clauses.append(f"{qcol} IN ({','.join('?' for _ in values)})")
params.extend(values)
elif op_key == "between":
values = value if isinstance(value, list) else []
if len(values) != 2:
raise ValueError("between requires exactly two values")
clauses.append(f"{qcol} BETWEEN ? AND ?")
params.extend(values)
else:
clauses.append(f"{qcol} {sql_op} ?")
params.append(value)
return (" WHERE " + " AND ".join(clauses)) if clauses else "", params
def _build_order(self, table: str, order_by: Any, aliases: set[str] | None = None) -> str:
if not order_by:
return ""
aliases = aliases or set()
parts: list[str] = []
for item in list(order_by)[:4]:
if not isinstance(item, dict):
continue
col = str(item.get("column", ""))
if col not in aliases:
self._validate_column(table, col)
direction = "DESC" if str(item.get("direction", "asc")).lower() == "desc" else "ASC"
parts.append(f"{self._quote_identifier(col)} {direction}")
return " ORDER BY " + ", ".join(parts) if parts else ""
@staticmethod
def _clean_frame(df: pd.DataFrame) -> pd.DataFrame:
out = df.copy()
for col in out.columns:
out[col] = out[col].map(
lambda v: None if isinstance(v, float) and (math.isnan(v) or math.isinf(v)) else v
)
return out
def _execute_select(self, action: dict[str, Any]) -> tuple[pd.DataFrame, str]:
table = self._validate_table(str(action.get("table", "")))
requested = action.get("columns") or self._columns[table]
columns = [self._validate_column(table, str(c)) for c in requested]
columns = columns[:24]
where_sql, params = self._build_filters(table, action.get("filters"))
order_sql = self._build_order(table, action.get("order_by"))
limit = min(max(int(action.get("limit", 40)), 1), _MAX_ROWS_PER_ACTION)
sql = (
"SELECT " + ", ".join(self._quote_identifier(c) for c in columns) +
f" FROM {self._quote_identifier(table)}" + where_sql + order_sql + " LIMIT ?"
)
df = pd.read_sql_query(sql, self.conn, params=(*params, limit))
return self._clean_frame(df), sql
def _execute_aggregate(self, action: dict[str, Any]) -> tuple[pd.DataFrame, str]:
table = self._validate_table(str(action.get("table", "")))
group_by = [self._validate_column(table, str(c)) for c in (action.get("group_by") or [])][:6]
metric_exprs: list[str] = []
aliases: set[str] = set()
for metric in (action.get("metrics") or [])[:12]:
if not isinstance(metric, dict):
continue
fn = str(metric.get("function", "")).upper()
if fn not in _ALLOWED_AGGS:
raise ValueError(f"Unsupported aggregate: {fn}")
col = str(metric.get("column", "*"))
self._validate_column(table, col, allow_star=(fn == "COUNT"))
alias = re.sub(r"[^A-Za-z0-9_]+", "_", str(metric.get("alias") or f"{fn.lower()}_{col}"))
aliases.add(alias)
expr_col = "*" if col == "*" else self._quote_identifier(col)
metric_exprs.append(f"{fn}({expr_col}) AS {self._quote_identifier(alias)}")
if not metric_exprs:
metric_exprs = ['COUNT(*) AS "row_count"']
aliases.add("row_count")
select_parts = [self._quote_identifier(c) for c in group_by] + metric_exprs
where_sql, params = self._build_filters(table, action.get("filters"))
group_sql = " GROUP BY " + ", ".join(self._quote_identifier(c) for c in group_by) if group_by else ""
order_sql = self._build_order(table, action.get("order_by"), aliases=aliases)
limit = min(max(int(action.get("limit", 40)), 1), _MAX_ROWS_PER_ACTION)
sql = (
"SELECT " + ", ".join(select_parts) + f" FROM {self._quote_identifier(table)}" +
where_sql + group_sql + order_sql + " LIMIT ?"
)
df = pd.read_sql_query(sql, self.conn, params=(*params, limit))
return self._clean_frame(df), sql
def _execute_search_input(self, action: dict[str, Any]) -> tuple[pd.DataFrame, str]:
term = str(action.get("term", "")).strip()
if not term:
raise ValueError("search_input requires a term")
section = str(action.get("section", "")).strip().upper()
limit = min(max(int(action.get("limit", 40)), 1), _MAX_ROWS_PER_ACTION)
sql = (
"SELECT line_no, section_name, section_row_no, raw_text FROM model_input_lines "
"WHERE lower(raw_text) LIKE lower(?)"
)
params: list[Any] = [f"%{term}%"]
if section:
sql += " AND upper(section_name)=?"
params.append(section)
sql += " ORDER BY line_no LIMIT ?"
params.append(limit)
df = pd.read_sql_query(sql, self.conn, params=params)
return self._clean_frame(df), sql
def _execute_describe(self, action: dict[str, Any]) -> tuple[pd.DataFrame, str]:
table = self._validate_table(str(action.get("table", "")))
rows = self.conn.execute(f"PRAGMA table_info({self._quote_identifier(table)})").fetchall()
df = pd.DataFrame(rows, columns=["cid", "name", "type", "notnull", "default_value", "primary_key"])
df["row_count"] = self._row_count(table)
return df, f"PRAGMA table_info({table})"
def execute_plan(self, plan: dict[str, Any]) -> RetrievalResult:
contexts: list[str] = []
audit: list[dict[str, Any]] = []
used_chars = 0
for index, action in enumerate(plan.get("actions", [])[:_MAX_ACTIONS], start=1):
if not isinstance(action, dict):
continue
operation = str(action.get("operation", "select")).lower()
label = str(action.get("label") or f"Retrieval {index}")
try:
if operation == "select":
df, sql = self._execute_select(action)
elif operation == "aggregate":
df, sql = self._execute_aggregate(action)
elif operation == "search_input":
df, sql = self._execute_search_input(action)
elif operation == "describe_table":
df, sql = self._execute_describe(action)
else:
raise ValueError(f"Unsupported operation: {operation}")
csv_text = df.to_csv(index=False)
block = f"=== {label} ===\n{csv_text}"
remaining = _MAX_CONTEXT_CHARS - used_chars
if remaining <= 0:
break
if len(block) > remaining:
block = block[:remaining] + "\n[retrieval context truncated]"
contexts.append(block)
used_chars += len(block)
audit.append({
"action": index,
"label": label,
"operation": operation,
"table": action.get("table", "model_input_lines" if operation == "search_input" else ""),
"rows_returned": int(len(df)),
"status": "ok",
})
except Exception as exc:
audit.append({
"action": index,
"label": label,
"operation": operation,
"table": action.get("table", ""),
"rows_returned": 0,
"status": f"error: {exc}",
})
return RetrievalResult(context="\n\n".join(contexts), audit=audit, plan=plan)
|