I3eka commented on code in PR #43237: URL: https://github.com/apache/superset/pull/43237#discussion_r4384700375
########## 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: Rechecked on `ff25e7fbe7` after the Apache master merge. The loop removes a field when shrinking makes no progress; the one-character-plus-ellipsis regression still passes. The structured-output and original SQL-display tests pass in the full AI suite. ########## 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: Rechecked on `ff25e7fbe7`. The SQL Lab execution check remains the first operation in ExecuteSqlTool.run, before database lookup or executor creation. Its denied-path regression asserts that database lookup is never called. This preserves the SQL capability boundary for dataset readers described in SECURITY.md. ########## docs/admin_docs/configuration/ai-assistant.mdx: ########## @@ -0,0 +1,489 @@ +--- +title: AI Assistant +hide_title: true +sidebar_position: 17 +version: 1 +--- + +# AI Assistant + +The AI Assistant is a conversational interface for exploring your data. A user +asks a question in plain language; the assistant finds relevant datasets, +inspects their schema, writes and runs read-only SQL, and answers with both the +result and the query it used. + +Superset ships **no model provider and talks to no model vendor by default**. +The feature is disabled, and even when enabled it returns `404` until you point +it at a provider you control. Nothing is sent anywhere until you configure it. + +## Enabling it + +Two things are required: the feature flag, and a provider. + +```python +# superset_config.py +FEATURE_FLAGS = { + "AI_ASSISTANT": True, +} + +AI_LLM_PROVIDER_CLASS = "superset.ai.llm.anthropic.AnthropicProvider" +AI_LLM_PROVIDER_CONFIG = { + "api_key": os.environ["ANTHROPIC_API_KEY"], + "models": { + "default": "claude-sonnet-4-5", + "fast": "claude-haiku-4-5", + "reasoning": "claude-opus-4-1", + }, +} +``` + +Install the matching extra: + +```bash +pip install "apache-superset[ai-anthropic]" # or [ai-openai] +``` + +Then run `superset init` so the assistant's permissions are created and assigned +to roles. Without this the endpoints return `403`. + +Conversations are stored in Superset's metadata database, so no extra +infrastructure is needed for the default configuration. + +### Which roles get access + +`superset init` grants `can_read`/`can_write` on `AIAssistant` to **Admin** and +**Alpha** only. "Write" here means writing one's own conversation — the +assistant's tools are read-only and it cannot create or modify assets. + +**Gamma does not get it by default.** The assistant runs queries and costs +money per question, so it is granted deliberately rather than inherited. To +give it to Gamma users, add `can_read`/`can_write` on `AIAssistant` to Gamma or +to a custom role. + +Every query the assistant runs is subject to the *user's own* database and +dataset permissions. It cannot read anything the person chatting with it could +not read themselves. + +Because it is not in Gamma, it is also not inherited by the Public role when +`PUBLIC_ROLE_LIKE = "Gamma"` — an anonymous visitor cannot reach the assistant +unless you grant it explicitly. + +## Choosing a provider + +`AI_LLM_PROVIDER_CLASS` is a dotted path to a +`superset.ai.llm.base.BaseLLMProvider` subclass. Two are bundled: + +| Class | Use for | +| --- | --- | +| `superset.ai.llm.anthropic.AnthropicProvider` | The Anthropic Messages API | +| `superset.ai.llm.openai_compatible.OpenAICompatibleProvider` | OpenAI, and anything exposing an OpenAI-compatible endpoint — vLLM, Ollama, a private gateway | + +`AI_LLM_PROVIDER_CONFIG` is passed to the provider's constructor and its +contents are provider-defined. For the OpenAI-compatible provider, `base_url` +points it anywhere: + +```python +AI_LLM_PROVIDER_CLASS = "superset.ai.llm.openai_compatible.OpenAICompatibleProvider" +AI_LLM_PROVIDER_CONFIG = { + "base_url": "https://llm.internal.example.com/v1";, + "api_key": os.environ["MY_GATEWAY_KEY"], + "models": {"default": "our-hosted-model"}, +} +``` + +Everything vendor-specific — URLs, authentication, model naming — lives in the +provider. Superset core contains none of it, so a self-hosted model or a private +gateway needs configuration rather than a fork. + +### Model tiers and selection + +Profiles and prompts refer to capability *tiers* (`default`, `fast`, +`reasoning`), never to a vendor's model names. The provider maps tiers to +concrete models via the `models` dict. A tier you do not configure is an error +when requested, never a silent substitution — so cost and answer quality stay +attributable to the model actually used. + +Users may also pin a specific model per turn. Only models present in your +`models` mapping are accepted; anything else is rejected. + +## Agent profiles + +A profile bundles the decisions that differ between a quick answer and a careful +investigation: which tools are available, which model tier, and how many steps. +Two ship by default — `default` and `analyst`. + +**Which tools a model may invoke is a decision each deployment makes**, so +profiles are fully configurable. `AI_AGENT_PROFILES` maps a profile key to the +fields you want to override, leaving the rest alone: + +```python +AI_AGENT_PROFILES = { + # Let the assistant search and inspect, but never run SQL. + "default": {"tools": ["search_assets", "list_databases", "get_schema"]}, + + # Let the analyst profile think harder and longer. + "analyst": {"model_alias": "reasoning", "max_turns": 60}, + + # Add a profile only some users may select. + "deep": { + "name": "Deep analysis", + "description": "Slow, thorough, multi-step.", + "tools": ["search_assets", "get_schema", "execute_sql"], + "required_permission": ("can_write", "AIAssistant"), + }, +} +``` + +A tool name that does not exist is an error naming the typo and listing the +valid names, rather than an assistant that quietly lacks a capability. An empty +`tools` list is valid and means conversation with no data access. + +`required_permission` is enforced on both the listing *and* the run path, so a +profile a user cannot see is also one they cannot invoke by posting its key. + +### Available tools + +| Tool | What it does | +| --- | --- | +| `search_assets` | Finds datasets, charts and dashboards the user can see | +| `list_databases` | Lists database connections exposed to SQL Lab | +| `get_schema` | Lists schemas, tables and columns | +| `execute_sql` | Runs a **read-only** query | +| `validate_sql` | Checks a query without running it | +| `get_chart_context` | Reads a chart's definition | +| `get_dashboard_context` | Reads a dashboard's definition | + +## Customising the prompt + +Three levers, in increasing order of bluntness. + +**Add to it.** `AI_EXTRA_PROMPT_SECTIONS` appends your own sections. This is +where deployment-specific knowledge belongs — your table conventions, your +warehouse's dialect quirks, how your business defines a metric. The shipped +prompt is deliberately generic and mentions no particular database engine. + +**Remove from it.** `AI_DISABLED_PROMPT_SECTIONS` drops a shipped section by +key, for when you disagree with one. The safety section cannot be disabled. + +**Replace it.** `AI_SYSTEM_PROMPT` substitutes the whole thing. + +:::warning +Setting `AI_SYSTEM_PROMPT` discards the shipped safety and prompt-injection +rules along with everything else. Your deployment then owns them. +::: + +`AI_SYSTEM_PROMPT_MUTATOR` is a last-mile callable applied after assembly, +mirroring `SQL_QUERY_MUTATOR`. + +## Where turns execute + +`AI_ASSISTANT_EXECUTION_MODE` decides where the work happens. + +**`"inline"`** (default) runs the turn in the web process. Nothing extra to +deploy. + +**`"worker"`** hands it to Celery. Web workers stay free, and a browser that +loses its connection can rejoin a run in progress. It requires Celery and a +Redis event bus: + +```python +AI_ASSISTANT_EXECUTION_MODE = "worker" +AI_ASSISTANT_EVENT_BUS = "redis" +AI_ASSISTANT_EVENT_BUS_CACHE_CONFIG = { + "CACHE_TYPE": "RedisCache", + "CACHE_REDIS_HOST": "redis", + "CACHE_REDIS_PORT": 6379, + "CACHE_REDIS_DB": 0, +} + +class CeleryConfig: + imports = ( + # ... your existing imports ... + "superset.ai.tasks", + ) +``` + +Streams need Redis commands the general-purpose cache client does not expose, +which is why the bus is configured separately rather than reusing `CACHE_CONFIG`. + +Selecting `"worker"` with the in-memory event bus raises rather than leaving +every stream silently empty, and so does selecting the Redis bus without a +usable connection. + +A turn is deliberately **not** retried after a worker crash: inference costs +money, and re-running a turn the user may already have partly seen would charge +twice. The message records that it failed and the user can ask again. + +## Safety and limits + +Guards are applied before any tool runs, configured via +`AI_AGENT_TOOL_POLICIES`: + +- **Read-only SQL.** Enforced using Superset's own SQL parser, not pattern + matching — so a write hidden behind a comment, a CTE, a second statement, or + an unparseable construct is refused. `EXPLAIN`, `SHOW` and `DESCRIBE` are + permitted; everything the parser cannot vouch for is not. +- **Identifier safety.** Table and column names are resolved against metadata + the user may see rather than interpolated into SQL. + +These bound blast radius; they do not replace authorization. Every tool that +touches a data-bearing object performs the same permission check the REST API +does. + +Result sizes are capped by `AI_AGENT_MAX_RESULT_ROWS` and +`AI_AGENT_MAX_RESULT_BYTES`, and truncation is reported rather than hidden. Turn +length is bounded by `AI_AGENT_MAX_TURNS` and `AI_AGENT_TIMEOUT_SECONDS`; a run +that exhausts either answers with what it has. + +Content that arrives from your warehouse or asset metadata — table comments, +chart titles, column labels — is marked as untrusted in the prompt, because a +value in a database is data and not an instruction. + +### Cancellation + +Cancellation is cooperative: a run stops at its next step boundary. A run inside +a single long model call or a single long query will not stop until that call +returns. + +## Monitoring and tracing + +Superset bundles **no integration with any AI monitoring product**. Instead it +exposes a small sink interface, `AITelemetry`, and calls it once per run, once +per model round trip and once per tool call. Whatever you already use — +Braintrust, LangSmith, Langfuse, Arize Phoenix, an OpenTelemetry collector, a +self-hosted alternative, or a table in your own warehouse — you connect by +implementing that interface and listing it in `AI_TELEMETRY`. + +Entries are instances or dotted paths, exactly as for `EVENT_LOGGER` and +`STATS_LOGGER`. Two sinks ship in-tree and depend on nothing external: + +```python +# superset_config.py +import logging + +from superset.ai.telemetry import LoggingAITelemetry, StatsLoggerAITelemetry + +AI_TELEMETRY = [ + # One structured line per span, at the level you choose. + LoggingAITelemetry(level=logging.INFO), + # Counters and timings through your configured STATS_LOGGER. + StatsLoggerAITelemetry(), +] +``` + +`StatsLoggerAITelemetry` emits under a `superset.ai.` prefix: `run.start`, +`run.end`, `run.outcome.<outcome>`, `run.duration_ms`, `run.turns`, +`run.tokens.input`, `run.tokens.output`, `model_call`, +`model_call.duration_ms`, `model_call.error`, `error`, and per tool +`tool_call.<tool>`, `tool_call.<tool>.duration_ms`, `tool_call.<tool>.error` +and `tool_call.<tool>.truncated`. User, run and thread identifiers deliberately +never appear in a metric name — a metric per user is how a metrics backend gets +brought down. That detail belongs in a trace, which is what a custom sink is +for. + +### The content trade-off + +`AI_TELEMETRY_REDACT_CONTENT` defaults to `True`, and telemetry then carries +**structure and measurements only**: durations, token counts, model names, tool +names, outcomes, error classes, and the run, thread and user identifiers. No +question, no answer, no SQL, no row of data. Redaction is applied where the +trace is built, so a sink cannot receive content by accident even if it looks +for it. + +Setting it to `False` is what makes a trace genuinely useful for debugging +answer quality — you can read the prompt that produced a wrong answer and the +statement it ran. It also means the text of business questions and values from +your warehouse leave Superset for whichever service your sinks talk to. In many +organisations that is a decision for someone other than the person editing the +config file. `AI_TELEMETRY_MAX_CONTENT_CHARS` (default 10,000) caps any single +content field so one large result cannot dominate a payload. + +### A custom sink + +Every method has a no-op default, so implement only the ones you need — a sink +that only wants token counts overrides `on_model_call` and nothing else. + +```python +from superset.ai.telemetry import AITelemetry, ModelCallTrace, RunTrace + + +class TracingServiceTelemetry(AITelemetry): + """Forwards runs to an external tracing service.""" + + def __init__(self, client): + self._client = client + + def on_run_start(self, run: RunTrace) -> None: + self._client.start_span(run.run_id, name="superset.ai.run", attributes={ + "thread": run.thread_uuid, + "user": run.user_id, + }) + + def on_model_call(self, run: RunTrace, call: ModelCallTrace) -> None: + self._client.event(run.run_id, "model_call", { + "turn": call.turn, + "model": call.model, + "input_tokens": call.input_tokens, + "output_tokens": call.output_tokens, + # None unless you have turned redaction off. + "prompt": call.system_prompt, + }) + + def on_run_end(self, run: RunTrace) -> None: + self._client.end_span(run.run_id, status=str(run.outcome), attributes={ + "duration_ms": run.duration_ms, + "turns": run.turns, + "usage": run.usage, + }) + + +AI_TELEMETRY = [TracingServiceTelemetry(client=my_tracing_client)] +``` + +Three things to know before you write one: + +- **Sinks are called on the thread answering the user.** Anything that makes a + network call should hand off to a queue or a background thread; otherwise a + slow monitoring backend becomes slow answers. +- **A sink that raises cannot break a run.** Failures are logged once and + ignored, and the other configured sinks still receive everything. The same + applies to a dotted path that will not import: it is skipped with a warning + rather than taking the assistant down, because a missing observer loses the + record of a run and not the run itself. +- **`agent_key`, `model` and `question` are resolved after the run starts**, so + a `RunTrace` passed to `on_run_start` may carry less than the one passed to + the later hooks. Read those on `on_run_end`. + +## Connecting your own MCP servers + +The assistant's built-in tools cover Superset itself. To let it reach anything +else — your data catalog, a metrics service, a ticketing system — attach an +[MCP](https://modelcontextprotocol.io) server. Superset bundles no third-party +integration and connects to nothing by default; you name the servers. + +```bash +pip install "apache-superset[ai-mcp]" +``` + +```python +AI_AGENT_MCP_SERVERS = { + "acme_catalog": { + "url": "https://mcp.acme.internal/mcp";, + "transport": "streamable_http", # or "sse" + "headers": {"Authorization": f"Bearer {os.environ['ACME_MCP_TOKEN']}"}, + "timeout_seconds": 30, + "tool_allowlist": ["search_tables"], # omit to offer every tool + }, +} + +# Then let a profile use it. +AI_AGENT_PROFILES = { + "default": {"mcp_servers": ["acme_catalog"]}, +} +``` + +Its tools appear to the model as `mcp__acme_catalog__search_tables`. The +namespace means a foreign tool can never shadow a built-in one, and it is the +name to use in `tool_allowlist` and `tool_denylist`. + +### What Superset does to keep a foreign server contained + +A third-party server is untrusted input, and possibly untrusted intent: + +- **Everything it returns is marked as untrusted** before the model sees it, so + text in a tool result is treated as data rather than instructions. Tool + *descriptions* get the same treatment, since they enter the prompt every turn. +- **No Superset credential is ever forwarded.** Only the headers you configured + for that server are sent — never the user's session cookie, CSRF token, or an + inbound authorization header. +- **SQL execution through a foreign server is refused by default.** Superset's + read-only enforcement and per-dataset authorization cannot apply to a query + another system runs, so allowing it would silently bypass both. Set + `AI_AGENT_MCP_DENY_FOREIGN_SQL = False` to accept that trade deliberately. +- **Results obey the same size cap** as built-in tools, and the cap is applied + while reading, so a hostile server cannot exhaust memory before truncation. +- **A server being down does not break the assistant.** Discovery failure means + that server contributes no tools for the turn; the built-ins keep working. + +A profile naming a server you have not configured is an error, because a typo +there is indistinguishable at runtime from an agent that has quietly lost a +capability. Note that discovery happens per turn, so a slow server adds its +latency to every turn that uses it. + +## Retention + +Conversations are kept for `AI_ASSISTANT_MESSAGE_RETENTION_DAYS` (default 30). +Pruning is not automatic — schedule it if you want it enforced. Review Comment: Confirmed and fixed in this branch by importing the pruning implementation from #43135. ai.prune_conversations is registered in the default daily beat schedule; the DAO removes expired feedback, messages and threads together. Both the deletion and schedule regressions failed here before the import and pass on `ff25e7fbe7`. Operators with their own beat schedule must include the documented task. ########## superset/ai/eventbus.py: ########## @@ -0,0 +1,314 @@ +# 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. +""" +Carries streamed events from whatever produced them to the HTTP response. + +Two implementations, matching the two execution modes. Inline execution needs +nothing more than an in-process queue. Worker execution needs a shared, +*replayable* channel — replayable because a browser that loses its connection +must be able to rejoin a run already in progress, which rules out +publish/subscribe: a subscriber that was absent when an event was published +never sees it. + +The Redis implementation therefore uses streams, and reuses the cache backend +that Superset's async-query channel already configures rather than introducing +a second Redis client to operate. +""" + +from __future__ import annotations + +import logging +import queue +from abc import ABC, abstractmethod +from collections.abc import Iterator +from typing import Any + +from superset.ai.events import StreamEvent +from superset.ai.types import StreamEventType +from superset.utils import json + +logger = logging.getLogger(__name__) + +#: Yielded by :meth:`BaseEventBus.consume` when nothing arrived within the poll +#: interval, so a caller can emit a keep-alive rather than block indefinitely. +IDLE = None + +#: Terminal event types. Seeing one ends consumption, so a reader does not hang +#: waiting for a producer that has already finished. +_TERMINAL = frozenset( + {StreamEventType.DONE, StreamEventType.ERROR, StreamEventType.CANCELLED} +) + + +class BaseEventBus(ABC): + """A per-run channel of events.""" + + @abstractmethod + def publish(self, run_id: str, event: StreamEvent) -> None: + """Append an event to a run's channel.""" + + @abstractmethod + def consume( + self, + run_id: str, + timeout_seconds: float, + poll_seconds: float = 1.0, + ) -> Iterator[StreamEvent | None]: + """ + Yield a run's events until a terminal one arrives or time runs out. + + Yields :data:`IDLE` when a poll interval passes with nothing new, which + is the caller's cue to send a keep-alive frame. + """ + + @abstractmethod + def close(self, run_id: str) -> None: + """Release any resources held for a run.""" + + +class MemoryEventBus(BaseEventBus): + """ + An in-process queue per run. + + Correct only when the producer and the streaming request share a process. + Selecting this alongside worker execution would leave every stream silent, + which :func:`get_event_bus` refuses to allow. + """ + + def __init__(self) -> None: + self._queues: dict[str, queue.SimpleQueue[StreamEvent]] = {} + + def _queue_for(self, run_id: str) -> queue.SimpleQueue[StreamEvent]: + return self._queues.setdefault(run_id, queue.SimpleQueue()) + + def publish(self, run_id: str, event: StreamEvent) -> None: + self._queue_for(run_id).put(event) + + def consume( + self, + run_id: str, + timeout_seconds: float, + poll_seconds: float = 1.0, + ) -> Iterator[StreamEvent | None]: + import time + + # Deliberately not ``_queue_for``: reading must not create a channel. + # This bus lives for the life of the process, so a client polling + # unknown run identifiers would otherwise grow the dict without bound. + channel = self._queues.get(run_id) + deadline = time.monotonic() + timeout_seconds + + while True: + remaining = deadline - time.monotonic() + if remaining <= 0: + return + if channel is None: + # The producer may not have published yet; look again rather + # than deciding the run does not exist. Only report idle if it + # is still absent, so a channel that appeared during the wait + # is drained on this pass instead of costing an extra tick. + channel = self._queues.get(run_id) + if channel is None: + yield IDLE + time.sleep(min(poll_seconds, remaining)) + continue + try: + # Bounded by whichever is sooner, so a generous poll interval + # cannot overshoot the caller's deadline. + event = channel.get(timeout=min(poll_seconds, remaining)) + except queue.Empty: + yield IDLE + continue + yield event + if event.type in _TERMINAL: + return + + def close(self, run_id: str) -> None: + self._queues.pop(run_id, None) + + +class RedisStreamEventBus(BaseEventBus): + """ + A Redis stream per run. + + Replayable by construction: a reconnecting reader starts from the beginning + of the stream and catches up, which is what makes worker execution usable + from a browser on a flaky connection. + """ + + def __init__( + self, + cache: Any, + prefix: str = "ai-events-", + ttl_seconds: int = 900, + ) -> None: + self._cache = cache + self._prefix = prefix + self._ttl = ttl_seconds + + def _stream(self, run_id: str) -> str: + return f"{self._prefix}{run_id}" + + def publish(self, run_id: str, event: StreamEvent) -> None: + payload = { + "data": json.dumps({"type": event.type.value, "payload": event.payload}) + } + # A failure to publish must not kill the run that is producing useful + # work; the reader will time out and the answer is still persisted. + try: + self._cache.xadd(self._stream(run_id), payload, "*", 10_000) Review Comment: Confirmed and fixed here using the #43134 event-bus correction. Every successful publish sets/refreshes the configured stream TTL, including unfinished runs with no reader. Consumption also preserves ERROR/CANCELLED followed by DONE. Nine event-bus regressions failed before the import and pass on `ff25e7fbe7`; this uses a deterministic stream-cache fixture, not a live Redis test. ########## docs/admin_docs/configuration/ai-assistant.mdx: ########## @@ -0,0 +1,489 @@ +--- +title: AI Assistant +hide_title: true +sidebar_position: 17 +version: 1 +--- + +# AI Assistant + +The AI Assistant is a conversational interface for exploring your data. A user +asks a question in plain language; the assistant finds relevant datasets, +inspects their schema, writes and runs read-only SQL, and answers with both the +result and the query it used. + +Superset ships **no model provider and talks to no model vendor by default**. +The feature is disabled, and even when enabled it returns `404` until you point +it at a provider you control. Nothing is sent anywhere until you configure it. + +## Enabling it + +Two things are required: the feature flag, and a provider. + +```python +# superset_config.py +FEATURE_FLAGS = { + "AI_ASSISTANT": True, +} + +AI_LLM_PROVIDER_CLASS = "superset.ai.llm.anthropic.AnthropicProvider" +AI_LLM_PROVIDER_CONFIG = { + "api_key": os.environ["ANTHROPIC_API_KEY"], + "models": { + "default": "claude-sonnet-4-5", + "fast": "claude-haiku-4-5", + "reasoning": "claude-opus-4-1", + }, +} +``` + +Install the matching extra: + +```bash +pip install "apache-superset[ai-anthropic]" # or [ai-openai] +``` + +Then run `superset init` so the assistant's permissions are created and assigned Review Comment: Fixed in the setup guide on `ff25e7fbe7`: enable the assistant and extension UI, install the selected provider dependency before container startup, run superset db upgrade, then superset init. I checked the bootstrap dependency-loading path. No production migration or deployment was performed. ########## superset/ai/api.py: ########## @@ -0,0 +1,989 @@ +# 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. +""" +REST API for the AI assistant. + +Every route carries ``@protect()`` and is reached through ``@expose`` on a +``BaseSupersetApi`` subclass, which is what makes Flask-AppBuilder's +authorization actually run. Ownership is enforced a second time in the command +and DAO layers, so a conversation identifier is never on its own a capability. +""" + +from __future__ import annotations + +import logging +import time +from collections.abc import Generator +from typing import Any, cast + +from flask import current_app, request, Response, stream_with_context +from flask_appbuilder.api import expose, permission_name, protect, safe +from marshmallow import ValidationError + +from superset.ai.events import ( + error_event, + KEEPALIVE_FRAME, + KEEPALIVE_INTERVAL_SECONDS, +) +from superset.ai.schemas import ( + AgentResponseSchema, + CancelPostSchema, + FeedbackPostSchema, + MessagePostSchema, + RunAcceptedResponseSchema, + SuggestedPromptsPostSchema, + ThreadDetailResponseSchema, + ThreadPostSchema, + ThreadPutSchema, + ThreadResponseSchema, +) +from superset.ai.types import MessageRole, MessageStatus +from superset.commands.ai.exceptions import ( + AIChatMessageInvalidError, + AIChatMessageNotFoundError, + AIChatThreadInvalidError, + AIChatThreadNotFoundError, +) +from superset.extensions import event_logger +from superset.utils.core import get_user_id +from superset.utils.decorators import transaction +from superset.views.base_api import BaseSupersetApi, statsd_metrics + +logger = logging.getLogger(__name__) + +#: Upper bound on how long a client may hold a stream open, so an abandoned +#: browser tab cannot pin a worker indefinitely. +_STREAM_TIMEOUT_SECONDS = 900 + +#: How often a reader checks the event bus for new frames. +#: +#: Deliberately separate from ``KEEPALIVE_INTERVAL_SECONDS``. Passing the +#: keep-alive interval as the poll interval made the reader sleep fifteen seconds +#: between checks and then deliver everything that had accumulated in one batch — +#: so a worker-mode run showed no streaming at all: the answer and every tool call +#: appeared in fifteen-second lumps. One controls responsiveness, the other how +#: often an idle connection is reassured; they are not the same number. +_EVENT_POLL_SECONDS = 0.1 + + +class AIRestApi(BaseSupersetApi): + """Conversations with the AI assistant.""" + + resource_name = "ai" + openapi_spec_tag = "AI Assistant" + allow_browser_login = True + class_permission_name = "AIAssistant" + + openapi_spec_component_schemas = ( + AgentResponseSchema, + CancelPostSchema, + FeedbackPostSchema, + MessagePostSchema, + RunAcceptedResponseSchema, + SuggestedPromptsPostSchema, + ThreadDetailResponseSchema, + ThreadPostSchema, + ThreadPutSchema, + ThreadResponseSchema, + ) + + @expose("/agent/", methods=("GET",)) + @protect() + @safe + @statsd_metrics + @permission_name("read") + def agents(self) -> Response: + """List agent profiles the current user may select. + --- + get: + summary: List available agent profiles + responses: + 200: + description: Available profiles + content: + application/json: + schema: + type: object + properties: + result: + type: array + items: + $ref: '#/components/schemas/AgentResponseSchema' + 401: + $ref: '#/components/responses/401' + 403: + $ref: '#/components/responses/403' + 404: + $ref: '#/components/responses/404' + """ + if (unavailable := self._reject_if_unconfigured()) is not None: + return unavailable + + from superset.ai.factories import get_profiles + + profiles = get_profiles().visible_to_current_user() + return self.response(200, result=[p.to_public_dict() for p in profiles]) + + @expose("/model/", methods=("GET",)) + @protect() + @safe + @statsd_metrics + @permission_name("read") + def models(self) -> Response: + """List models this deployment has configured. + --- + get: + summary: List selectable models + responses: + 200: + description: Configured model identifiers + content: + application/json: + schema: + type: object + properties: + result: + type: array + items: + type: string + 401: + $ref: '#/components/responses/401' + 404: + $ref: '#/components/responses/404' + """ + if (unavailable := self._reject_if_unconfigured()) is not None: + return unavailable + + from superset.ai.factories import get_provider + + return self.response(200, result=get_provider().available_models()) + + @expose("/thread/", methods=("POST",)) + @protect() + @safe + @statsd_metrics + @permission_name("write") + @event_logger.log_this_with_context( + action=lambda self, *args, **kwargs: f"{self.__class__.__name__}.post_thread", + log_to_statsd=False, + ) + def post_thread(self) -> Response: + """Create a conversation. + --- + post: + summary: Create a conversation + requestBody: + content: + application/json: + schema: + $ref: '#/components/schemas/ThreadPostSchema' + responses: + 201: + description: Conversation created + content: + application/json: + schema: + type: object + properties: + result: + $ref: '#/components/schemas/ThreadResponseSchema' + 400: + $ref: '#/components/responses/400' + 401: + $ref: '#/components/responses/401' + 404: + $ref: '#/components/responses/404' + """ + if (unavailable := self._reject_if_unconfigured()) is not None: + return unavailable + + from superset.commands.ai import CreateAIChatThreadCommand + + try: + payload = ThreadPostSchema().load(request.json or {}) + except ValidationError as error: + return self.response_400(message=error.messages) + try: + thread = CreateAIChatThreadCommand( + user_id=self._user_id(), + title=payload.get("title"), + agent_key=payload.get("agent_key"), + ).run() + except AIChatThreadInvalidError as ex: + return self.response_422(message=str(ex)) + return self.response(201, result=_thread_dict(thread)) + + @expose("/thread/", methods=("GET",)) + @protect() + @safe + @statsd_metrics + @permission_name("read") + def get_threads(self) -> Response: + """List the current user's conversations. + --- + get: + summary: List conversations + parameters: + - in: query + name: limit + schema: + type: integer + - in: query + name: offset + schema: + type: integer + responses: + 200: + description: Conversations + content: + application/json: + schema: + type: object + properties: + count: + type: integer + result: + type: array + items: + $ref: '#/components/schemas/ThreadResponseSchema' + 401: + $ref: '#/components/responses/401' + 404: + $ref: '#/components/responses/404' + """ + if (unavailable := self._reject_if_unconfigured()) is not None: + return unavailable + + from superset.daos.ai import AIChatThreadDAO + + limit = request.args.get("limit", type=int) or 50 + offset = request.args.get("offset", type=int) or 0 + threads = AIChatThreadDAO.find_all_for_user( + self._user_id(), limit=limit, offset=offset + ) + return self.response( + 200, + count=len(threads), + result=[_thread_dict(thread) for thread in threads], + ) + + @expose("/thread/<thread_uuid>", methods=("GET",)) + @protect() + @safe + @statsd_metrics + @permission_name("read") + def get_thread(self, thread_uuid: str) -> Response: + """Fetch a conversation and its messages. + --- + get: + summary: Get a conversation + parameters: + - in: path + name: thread_uuid + required: true + schema: + type: string + format: uuid + responses: + 200: + description: Conversation with messages + content: + application/json: + schema: + type: object + properties: + result: + $ref: '#/components/schemas/ThreadDetailResponseSchema' + 401: + $ref: '#/components/responses/401' + 404: + $ref: '#/components/responses/404' + """ + if (unavailable := self._reject_if_unconfigured()) is not None: + return unavailable + + from superset.daos.ai import ( + AIChatFeedbackDAO, + AIChatMessageDAO, + AIChatThreadDAO, + ) + + user_id = self._user_id() + thread = AIChatThreadDAO.find_by_uuid_for_user(thread_uuid, user_id) + if thread is None: + return self.response_404() + + messages = AIChatMessageDAO.find_for_thread(thread) + # Resolved for the whole transcript at once so the panel can show which + # replies this user already rated; without it a reload loses the verdict + # and the message looks unrated. + verdicts = AIChatFeedbackDAO.find_verdicts_for_user( + [message.id for message in messages], user_id + ) + detail = _thread_dict(thread) + detail["messages"] = [ + _message_dict(message, liked=verdicts.get(message.id)) + for message in messages + ] + return self.response(200, result=detail) + + @expose("/thread/<thread_uuid>", methods=("PUT",)) + @protect() + @safe + @statsd_metrics + @permission_name("write") + @event_logger.log_this_with_context( + action=lambda self, *args, **kwargs: f"{self.__class__.__name__}.put_thread", + log_to_statsd=False, + ) + def put_thread(self, thread_uuid: str) -> Response: + """Rename or archive a conversation. + --- + put: + summary: Update a conversation + parameters: + - in: path + name: thread_uuid + required: true + schema: + type: string + format: uuid + requestBody: + content: + application/json: + schema: + $ref: '#/components/schemas/ThreadPutSchema' + responses: + 200: + description: Conversation updated + 401: + $ref: '#/components/responses/401' + 404: + $ref: '#/components/responses/404' + 422: + $ref: '#/components/responses/422' + """ + if (unavailable := self._reject_if_unconfigured()) is not None: + return unavailable + + from superset.commands.ai import UpdateAIChatThreadCommand + + try: + payload = ThreadPutSchema().load(request.json or {}) + except ValidationError as error: + return self.response_400(message=error.messages) + try: + thread = UpdateAIChatThreadCommand( + thread_uuid, + self._user_id(), + title=payload.get("title"), + status=payload.get("status"), + ).run() + except AIChatThreadNotFoundError: + return self.response_404() + except AIChatThreadInvalidError as ex: + return self.response_422(message=str(ex)) + return self.response(200, result=_thread_dict(thread)) + + @expose("/thread/<thread_uuid>", methods=("DELETE",)) + @protect() + @safe + @statsd_metrics + @permission_name("write") + @event_logger.log_this_with_context( + action=lambda self, *args, **kwargs: f"{self.__class__.__name__}.delete_thread", + log_to_statsd=False, + ) + def delete_thread(self, thread_uuid: str) -> Response: + """Delete a conversation and its messages. + --- + delete: + summary: Delete a conversation + parameters: + - in: path + name: thread_uuid + required: true + schema: + type: string + format: uuid + responses: + 200: + description: Conversation deleted + 401: + $ref: '#/components/responses/401' + 404: + $ref: '#/components/responses/404' + """ + if (unavailable := self._reject_if_unconfigured()) is not None: + return unavailable + + from superset.commands.ai import DeleteAIChatThreadCommand + + try: + DeleteAIChatThreadCommand(thread_uuid, self._user_id()).run() + except AIChatThreadNotFoundError: + return self.response_404() + return self.response(200, message="OK") + + @expose("/thread/<thread_uuid>/message", methods=("POST",)) + @protect() + @safe + @statsd_metrics + @permission_name("write") + @event_logger.log_this_with_context( + action=lambda self, *args, **kwargs: f"{self.__class__.__name__}.post_message", + log_to_statsd=False, + ) + def post_message(self, thread_uuid: str) -> Response: + """Post a user message and start a run. + --- + post: + summary: Post a message + description: > + Stores the user's message, creates a placeholder assistant message, + and starts a run. Returns immediately; consume the answer from the + stream endpoint using the returned run identifier. + parameters: + - in: path + name: thread_uuid + required: true + schema: + type: string + format: uuid + requestBody: + content: + application/json: + schema: + $ref: '#/components/schemas/MessagePostSchema' + responses: + 202: + description: Run accepted + content: + application/json: + schema: + type: object + properties: + result: + $ref: '#/components/schemas/RunAcceptedResponseSchema' + 400: + $ref: '#/components/responses/400' + 401: + $ref: '#/components/responses/401' + 404: + $ref: '#/components/responses/404' + 422: + $ref: '#/components/responses/422' + """ + if (unavailable := self._reject_if_unconfigured()) is not None: + return unavailable + + from superset.ai.orchestrator import new_run_id + from superset.commands.ai import AppendAIChatMessageCommand + + try: + payload = MessagePostSchema().load(request.json or {}) + except ValidationError as error: + return self.response_400(message=error.messages) + user_id = self._user_id() + + try: + user_message = AppendAIChatMessageCommand( + thread_uuid, + user_id, + MessageRole.USER, + payload["content"], + request_id=payload.get("request_id"), + ).run() + # Created up front so a client that reconnects before any token + # arrives still has a row to attach its stream to. + assistant_message = AppendAIChatMessageCommand( + thread_uuid, + user_id, + MessageRole.ASSISTANT, + "", + request_id=payload.get("request_id"), + status=MessageStatus.PENDING, + ).run() + except AIChatThreadNotFoundError: + return self.response_404() + except (AIChatMessageInvalidError, AIChatThreadInvalidError) as ex: + return self.response_422(message=str(ex)) + + run_id = new_run_id() + _record_run_context(assistant_message, run_id, payload) + Review Comment: Revalidated the imported #44561 fix on `ff25e7fbe7` together with the new profile propagation. Replays keep the accepted run ID, model, profile and page context, and the atomic pending claim still permits only one execution. A new regression additionally checks that replaying an older turn cannot undo a later conversation-profile selection. The combined API/DAO/runtime tests pass. -- 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]
