I3eka commented on code in PR #43237: URL: https://github.com/apache/superset/pull/43237#discussion_r3818942785
########## superset/ai/tools/base.py: ########## @@ -0,0 +1,624 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +""" +The tool contract and the registry that dispatches to it. + +A tool is a small, self-authorizing unit of work. "Self-authorizing" is the +important half: the registry does not check permissions on a tool's behalf, and +neither does the policy chain in :mod:`superset.ai.policy`, which answers the +coarser question of whether a *shape* of call should be attempted at all. Any +tool that returns or mutates a data-bearing object performs its own +``security_manager`` check, because that is the only place with enough context to +know which object is being touched. + +A tool returns two things. :attr:`ToolOutput.content` is what the model reads. +:attr:`ToolOutput.display` is a summary for the UI, so a user can expand what the +assistant did and see the SQL it ran and the rows it got back. Both are +size-bounded here rather than in each tool: ``display`` is persisted on the +message and shipped to the browser, so an unbounded one would be a second way to +blow up a response. +""" + +from __future__ import annotations + +import logging +import time +from abc import ABC, abstractmethod +from dataclasses import dataclass, field +from typing import Any, ClassVar + +from superset.ai.llm.base import ToolCall, ToolDefinition, ToolResult +from superset.mcp_service.utils.sanitization import ( + LLM_CONTEXT_CLOSE_DELIMITER, + LLM_CONTEXT_ESCAPED_CLOSE_DELIMITER, + LLM_CONTEXT_ESCAPED_OPEN_DELIMITER, + LLM_CONTEXT_OPEN_DELIMITER, +) +from superset.utils import json + +logger = logging.getLogger(__name__) + +#: Keys added to a payload whose budget was exceeded. Phrased for the model: it +#: says what was lost and what to do differently, because a model told only +#: "truncated" reissues the identical call. +TRUNCATION_KEY = "_truncated" +TRUNCATION_NOTE_KEY = "_truncation_note" + +#: Share of the response budget the UI summary may use. The model's copy is the +#: one that has to be complete enough to reason over; ``display`` only has to be +#: enough to render, and it is persisted, so it gets the smaller share. +_DISPLAY_BUDGET_FRACTION = 0.5 + +#: Fallback response budget for use outside an application context, matching the +#: shipped ``AI_AGENT_MAX_RESULT_BYTES`` default. +_DEFAULT_MAX_BYTES = 256 * 1024 + + +class ToolError(Exception): + """ + A failure that should be shown to the model rather than raised at the user. + + Tools raise this for conditions the model can act on — a database it may not + read, a column that does not resolve, SQL that will not parse. The registry + turns it into a :class:`~superset.ai.llm.base.ToolResult` with + ``is_error=True`` so the turn continues and the model can correct itself. + + The message is model-visible, so it must never carry a driver exception, a + connection string, or a stack trace. + """ + + +@dataclass(frozen=True) +class ToolOutput: + """ + What a tool returns. + + Build one with :meth:`of` rather than by hand, so that ``content`` and + ``payload`` cannot disagree. + """ + + #: Model-facing text. Becomes :attr:`ToolResult.content`. + content: str + + #: JSON-serialisable summary for the UI, or ``None`` when there is nothing + #: worth rendering. Must never carry credentials, a connection string, or a + #: full result set. + display: dict[str, Any] | None = None + + #: The structure ``content`` was serialised from. Retained so the registry + #: can shrink an oversized result by dropping rows rather than cutting JSON + #: mid-token. ``None`` when the tool supplied text directly. + payload: Any = None + + @classmethod + def of( + cls, + payload: Any, + display: dict[str, Any] | None = None, + ) -> ToolOutput: + """Serialise ``payload`` as the model-facing content.""" + return cls( + content=json.dumps(payload, default=str), + display=display, + payload=payload, + ) + + +@dataclass +class ToolInvocation: + """ + One completed dispatch, with everything a caller might need. + + Exists because two consumers want different things from the same call: the + provider needs a :class:`~superset.ai.llm.base.ToolResult`, while the event + stream and message persistence want the UI summary and the timing. + """ + + call_id: str + tool_name: str + result: ToolResult + display: dict[str, Any] | None = None + duration_ms: int = 0 + truncated: bool = False + arguments: dict[str, Any] = field(default_factory=dict) + + @property + def is_error(self) -> bool: + """Whether the call failed.""" + return self.result.is_error + + def to_tool_result(self) -> ToolResult: + """The provider-neutral result to feed back to the model.""" + return self.result + + +class AITool(ABC): + """ + One capability offered to the model. + + Subclasses set :attr:`name`, :attr:`description` and :attr:`input_schema`, + and implement :meth:`run`. Everything else — size capping, error + translation, timing — is the registry's job. + """ + + #: Stable identifier the model calls, and the key an operator types when + #: configuring which tools an agent profile may use. Renaming one is a + #: breaking change: it appears in stored conversation history and in + #: deployment configuration. + name: ClassVar[str] = "" + + #: Shown to the model verbatim. This is the tool's entire user manual, so it + #: should say when to reach for the tool and what it returns, not merely + #: what it is called. + description: ClassVar[str] = "" + + #: JSON Schema for the arguments object. Providers translate it into + #: whatever their API expects. + input_schema: ClassVar[dict[str, Any]] = {"type": "object", "properties": {}} + + @abstractmethod + def run(self, **kwargs: Any) -> ToolOutput: + """ + Perform the work. + + Raise :class:`ToolError` for anything the model should see and be able + to recover from. Any other exception is treated as a defect: it is + logged with a traceback and reported to the model as a generic failure, + so that an unexpected driver error cannot leak its message. + """ + + def definition(self) -> ToolDefinition: + """Provider-neutral description of this tool.""" + return ToolDefinition( + name=self.name, + description=self.description, + input_schema=self.input_schema, + ) + + +def truncate_payload(payload: Any, max_bytes: int) -> tuple[str, bool]: + """ + Serialise ``payload`` and bound it to ``max_bytes``. + + Returns valid JSON text and whether anything was dropped. Mapping fields are + shortened largest-first so useful smaller fields survive without cutting JSON + mid-token. + """ + text = json.dumps(payload, default=str) + if len(text.encode("utf-8")) <= max_bytes: + return text, False + + if isinstance(payload, dict): + candidate = dict(payload) + list_counts = [ + f"{len(value)} {key}" + for key, value in payload.items() + if isinstance(value, list) and value + ] + candidate[TRUNCATION_KEY] = True + candidate[TRUNCATION_NOTE_KEY] = ( + f"Payload exceeded the {max_bytes} byte response budget" + + (f" ({', '.join(list_counts)})" if list_counts else "") + + ". Large fields were shortened; narrow the request to see the rest." + ) + + while True: + text = json.dumps(candidate, default=str) + if len(text.encode("utf-8")) <= max_bytes: + return text, True + values = { + key: value + for key, value in candidate.items() + if key not in {TRUNCATION_KEY, TRUNCATION_NOTE_KEY} + } + if not values: + break + key = max( + values, + key=lambda item: len( + json.dumps(values[item], default=str).encode("utf-8") + ), + ) + value = values[key] + if isinstance(value, str) and len(value) > 1: Review Comment: Fixed in 3c2fb7992f. The truncation loop now drops a field when shrinking makes no progress, which guarantees termination. Added a tight-budget regression using the one-character-plus-ellipsis case. ########## superset/ai/tools/sql.py: ########## @@ -0,0 +1,540 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. +""" +Running and checking SQL. + +Both tools are reads. ``execute_sql`` decides that with Superset's own parser +rather than a keyword match, because a prefix match is defeated by a leading +comment, a CTE wrapping a mutation, and a second statement smuggled after a +legitimate ``SELECT`` — all of which the parser already handles for the rest of +Superset. + +Authorization is layered deliberately: + +1. ``DatabaseDAO.find_by_id`` applies ``DatabaseFilter``, so a database the user + has no grant on is indistinguishable from one that does not exist. +2. ``expose_in_sqllab`` is honoured, so an operator who withheld a connection + from ad-hoc querying has also withheld it here. +3. The database's ``allow_dml`` setting is checked. +4. Every statement must be non-mutating. +5. ``security_manager.raise_for_access`` with ``force_dataset_match=True`` — the + same strictness SQL Lab applies — so the tables referenced must resolve to + datasets the user may read. +""" + +from __future__ import annotations + +import logging +import time +from decimal import Decimal +from typing import Any, Callable, ClassVar + +from superset.ai.tools.base import AITool, ToolError, ToolOutput + +logger = logging.getLogger(__name__) + +#: Rows returned when the caller does not ask for a specific limit. Small on +#: purpose: an agent inspecting data needs a shape, not a dump, and the model +#: pays for every row in context. +DEFAULT_ROW_LIMIT = 100 + +#: Rows kept in the UI summary. The summary is persisted on the message and sent +#: to the browser, so it carries a sample rather than the result set. +DISPLAY_SAMPLE_ROWS = 20 + +#: Characters of executed SQL shown in the UI summary. +DISPLAY_SQL_CHARS = 4000 + + +def _database_or_refuse(database_id: Any) -> Any: + """ + Resolve a database the current user is allowed to query, or refuse. + + Uses the DAO so that ``DatabaseFilter`` — the same filter the database REST + API applies — decides visibility. A database the user cannot see is reported + as "not found" rather than "forbidden", so the tool cannot be used to probe + for the existence of connections. + """ + from superset.daos.database import DatabaseDAO + + if not isinstance(database_id, int) or isinstance(database_id, bool): + raise ToolError("'database_id' must be an integer. Use list_databases first.") + + database = DatabaseDAO.find_by_id(database_id) + if database is None: + raise ToolError( + f"No database with id {database_id} is available to you. " + f"Call list_databases to see the ones you can query." + ) + if not database.expose_in_sqllab: + raise ToolError( + f"Database {database.database_name!r} is not available for ad-hoc " + f"queries. Call list_databases to see the ones that are." + ) + return database + + +def _parse_or_refuse(sql: str, database: Any) -> Any: + """ + Parse ``sql`` for the database's engine, or refuse. + + Fails closed: SQL that will not parse cannot be shown to be read-only, so it + is refused rather than executed. The parser error is logged rather than + returned, because its message quotes the offending query back and that text + is not ours to echo. + """ + from superset.sql.parse import SQLScript + + if not isinstance(sql, str) or not sql.strip(): + raise ToolError("'sql' must be a non-empty string.") + + try: + return SQLScript(sql, database.db_engine_spec.engine) + except Exception: # pylint: disable=broad-except + logger.info("Refusing unparseable SQL for database %s", database.id) + raise ToolError( + "That SQL could not be parsed. Send a single, syntactically valid " + "read-only statement." + ) from None + + +def _assert_read_only(script: Any, database: Any) -> None: + """ + Refuse anything that is not a read. + + The ``allow_dml`` check comes first because it is the more specific refusal: + telling the model the *deployment* forbids writes on this connection is more + useful than a generic "read-only tool" message. The blanket mutation check + then applies even where ``allow_dml`` is enabled, since this tool is a read + regardless of what the connection would otherwise permit. + """ + has_mutation = script.has_mutation() + + if has_mutation and not database.allow_dml: + raise ToolError( + f"Writes are disabled on database " + f"{database.database_name!r} (allow_dml is off). Rewrite this as a " + f"SELECT." + ) + if has_mutation: + raise ToolError( + "This tool only runs read-only SQL. Rewrite this as a SELECT — " + "statements that modify data or schema are refused." + ) + # A statement the parser could not fully model has no enumerable table + # references, so neither the mutation check above nor the per-table + # authorization below can vouch for it. Refused rather than guessed at. + if script.has_unparseable_statement: + raise ToolError( + "That SQL contains a statement this tool cannot verify as " + "read-only. Send a plain SELECT." + ) + + +def _raise_for_sql_access( + database: Any, + sql: str, + catalog: str | None, + schema: str | None, +) -> None: + """ + Check the user may read every table the query touches. + + ``force_dataset_match=True`` matches what SQL Lab's own pre-execute + validator uses: each referenced table must resolve to a registered dataset + the user has access to, rather than falling through to a broader + catalog- or schema-level grant. + """ + from superset import security_manager + from superset.exceptions import SupersetSecurityException + + try: + security_manager.raise_for_access( + database=database, + sql=sql, + catalog=catalog, + schema=schema, + force_dataset_match=True, + ) + except SupersetSecurityException as ex: + # The exception message names the tables that were denied, which is + # exactly what lets the model pick a different source. + raise ToolError(str(ex.error.message)) from None + + +def _row_limit(requested: Any) -> int: + """ + Clamp the caller's limit to the configured ceiling. + + The model may not raise the cap by asking for more; ``limit`` narrows only. + """ + # A misconfigured or absent ceiling falls back to the default rather than + # becoming unbounded: sending a query with no limit is the one outcome this + # function exists to prevent. + ceiling = DEFAULT_ROW_LIMIT + try: + from flask import current_app + + configured = current_app.config.get("AI_AGENT_MAX_RESULT_ROWS") + if configured is not None: + ceiling = int(configured) + except Exception: # pylint: disable=broad-except + ceiling = DEFAULT_ROW_LIMIT + if ceiling < 1: + ceiling = DEFAULT_ROW_LIMIT + + if requested is None: + return min(DEFAULT_ROW_LIMIT, ceiling) + if not isinstance(requested, int) or isinstance(requested, bool) or requested < 1: + raise ToolError("'limit' must be a positive integer.") + return min(requested, ceiling) + + +def _columns_and_records( + data: Any, +) -> tuple[list[dict[str, str]], list[dict[str, Any]]]: + """ + Normalise a statement's result rows. + + ``Database.execute`` returns a ``DataFrame`` when a row limit was supplied + and a plain list of mappings when one was not, so both shapes are handled + rather than relying on the caller always producing the first. Assuming the + frame is how this tool previously raised ``AttributeError`` on a result it + had asked for perfectly legitimately. + """ + if hasattr(data, "columns") and hasattr(data, "to_dict"): + columns = [ + {"name": str(name), "type": str(data[name].dtype)} for name in data.columns + ] + return columns, list(data.to_dict(orient="records")) + + records = [dict(row) for row in (data or [])] + names: list[str] = [] + for record in records: + for key in record: + if key not in names: + names.append(key) + # Without a frame there are no dtypes to report; the values themselves still + # carry their types through ``_json_safe``. + return [{"name": str(name), "type": "unknown"} for name in names], records + + +def _json_safe(value: Any) -> Any: + """ + Coerce one warehouse value into something JSON can carry. + + ``Decimal`` becomes a float and binary becomes text (or hex when it is not + text at all); everything else exotic — dates, intervals, UUIDs, driver + types — becomes its string form. Lossy by design: the model reads these, it + does not compute on them. + """ + if isinstance(value, (str, int, float, bool)) or value is None: + return value + if isinstance(value, Decimal): + return float(value) + if isinstance(value, (bytes, memoryview)): + raw = bytes(value) + try: + return raw.decode("utf-8") + except UnicodeDecodeError: + return raw.hex() + if isinstance(value, (list, tuple)): + return [_json_safe(item) for item in value] + if isinstance(value, dict): + return {str(key): _json_safe(item) for key, item in value.items()} + return str(value) + + +def _execute_via_database( + database: Any, + sql: str, + catalog: str | None, + schema: str | None, + limit: int, +) -> Any: + """ + The single point at which a warehouse is touched. + + Isolated as a module-level function so a test can substitute it and exercise + every guard above without a live connection. ``Database.execute`` is the + same entry point SQL Lab and the MCP service use, so this inherits Jinja + rendering, ``SQL_QUERY_MUTATOR``, disallowed-function and disallowed-table + checks, row-level security, and the executor's own ``allow_dml`` gate. + """ + from superset_core.queries.types import QueryOptions + + return database.execute( + sql, + QueryOptions(catalog=catalog, schema=schema, limit=limit), + ) + + +def _result_to_payload(result: Any, limit: int) -> dict[str, Any]: + """ + Flatten a ``QueryResult`` into rows the model can read. + + Only the last data-bearing statement is returned. A read-only script with + several ``SELECT``s is unusual, and returning every result set multiplies + the context cost for a case the model can trivially split into two calls. + """ + from superset_core.queries.types import QueryStatus + + if result.status != QueryStatus.SUCCESS: + raise ToolError( + f"The query did not complete: {result.error_message or result.status}" + ) + + statement = next( + (item for item in reversed(result.statements) if item.data is not None), + None, + ) + if statement is None: + return { + "rows": [], + "row_count": 0, + "columns": [], + "note": "No rows returned.", + "executed_sql": None, + } + + columns, records = _columns_and_records(statement.data) + rows = [ + {str(key): _json_safe(value) for key, value in record.items()} + for record in records[:limit] + ] + + payload: dict[str, Any] = { + "rows": rows, + "row_count": len(rows), + "columns": columns, + # The SQL the warehouse actually ran, after the limit and any row-level + # security rewrite. This is what the user needs to see to trust the + # answer, and what they would paste into SQL Lab to check it. + "executed_sql": getattr(statement, "executed_sql", None), + } + if len(records) > limit: + payload["truncated"] = True + payload["note"] = ( + f"Showing {limit} of {len(records)} rows. Add a tighter WHERE clause " + f"or aggregate to see the rest." + ) + return payload + + +def _sql_display( + database: Any, + payload: dict[str, Any], + duration_ms: int, +) -> dict[str, Any]: + """ + Build the UI summary for one query. + + Deliberately not the model-facing payload: the row sample is smaller, and + only the connection's name and id appear — never its URI or credentials. + """ + executed = payload.get("executed_sql") or "" + return { + "kind": "sql_result", + "database_id": database.id, + "database_name": database.database_name, + "executed_sql": str(executed)[:DISPLAY_SQL_CHARS], + "executed_sql_truncated": len(str(executed)) > DISPLAY_SQL_CHARS, + "columns": [column["name"] for column in payload.get("columns", [])], + "rows": payload.get("rows", [])[:DISPLAY_SAMPLE_ROWS], + "row_count": payload.get("row_count", 0), + "sample_only": len(payload.get("rows", [])) > DISPLAY_SAMPLE_ROWS, + "truncated": bool(payload.get("truncated", False)), + "duration_ms": duration_ms, + } + + +class ExecuteSqlTool(AITool): + """Run a read-only query and return its rows.""" + + name: ClassVar[str] = "execute_sql" + description: ClassVar[str] = ( + "Run a read-only SQL query against a Superset database connection and " + "return the rows. Only SELECT-style statements are accepted; anything " + "that writes data or changes schema is refused. Results are capped, so " + "aggregate or filter in SQL rather than asking for everything. Call " + "list_databases for a database_id and get_schema for table and column " + "names before writing the query." + ) + input_schema: ClassVar[dict[str, Any]] = { + "type": "object", + "properties": { + "database_id": { + "type": "integer", + "description": "Database connection to query, from list_databases.", + }, + "sql": { + "type": "string", + "description": "A single read-only SQL statement.", + }, + "schema": { + "type": "string", + "description": ( + "Schema unqualified table names resolve to. Optional; the " + "database's default is used when omitted." + ), + }, + "catalog": { + "type": "string", + "description": "Catalog to query, for engines that have them.", + }, + "limit": { + "type": "integer", + "description": ( + "Maximum rows to return. Narrows the default only; it " + "cannot raise the configured ceiling." + ), + }, + }, + "required": ["database_id", "sql"], + } + + def __init__( + self, + executor: Callable[[Any, str, str | None, str | None, int], Any] | None = None, + ) -> None: + # Injectable so the guards are testable without a warehouse. + self._executor = executor or _execute_via_database + + def run( # pylint: disable=too-many-arguments + self, + database_id: Any = None, + sql: Any = None, + schema: Any = None, + catalog: Any = None, + limit: Any = None, + **_ignored: Any, + ) -> ToolOutput: + database = _database_or_refuse(database_id) + script = _parse_or_refuse(sql, database) + _assert_read_only(script, database) + + schema_name = str(schema) if schema else None + catalog_name = str(catalog) if catalog else None + _raise_for_sql_access(database, sql, catalog_name, schema_name) + + row_limit = _row_limit(limit) + started = time.monotonic() + result = self._executor(database, sql, catalog_name, schema_name, row_limit) Review Comment: Fixed in 3c2fb7992f. ExecuteSqlTool now checks can_execute_sql_query on SQLLab before database lookup or executor construction, with a regression test for the denied path. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
